1 line
No EOL
459 KiB
JSON
1 line
No EOL
459 KiB
JSON
{"GEMV": {}, "GEMV_REVSPLITK": {"(1, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 16, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 16, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 16, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 16, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 16, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 16, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 16, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 16, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 16, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 16, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 16, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4096, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 8192, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3840, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 2, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3840, 15360, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 8, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 1, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 30720, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 8, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}}, "GEMV_SPLITK": {"(1, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 1, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 1, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 4, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 2, "BLOCK_SIZE_K": 2048, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 4, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 1, "BLOCK_SIZE_K": 2048, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 1, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 1, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 1, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 4096, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 8192, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3840, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 3840, 15360, 15360, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 8, "BLOCK_SIZE_K": 1024, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1, 30720, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 1, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "dot_prod_mode": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}}, "GEMM_SPLITK": {"(64, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 2, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 2, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 2, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 2, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 2, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 2, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 2, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(48, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(24, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4096, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4096, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4096, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4096, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 4096, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 4096, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 4096, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 4096, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 8192, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 8192, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 8192, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 8192, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 8192, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 8192, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 8192, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 8192, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3840, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3840, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3840, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3840, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3840, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3840, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3840, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3840, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3840, 15360, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3840, 15360, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3840, 15360, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3840, 15360, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 16, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 3840, 15360, 15360, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 3840, 15360, 15360, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 3840, 15360, 15360, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 3840, 15360, 15360, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 30720, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 30720, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 30720, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 30720, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "SPLIT_K": 4, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(64, 30720, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(32, 30720, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 32, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(16, 30720, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(8, 30720, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 16, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "SPLIT_K": 1, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 2, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}}, "GEMM": {"(4096, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4096, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4096, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4096, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 6144, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 6144, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 6144, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4096, 14336, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4096, 14336, 14336, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4096, 14336, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 28672, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 3, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 28672, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 3, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 28672, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3584, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3584, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3584, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4608, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4608, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 4608, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3584, 18944, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3584, 18944, 18944, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3584, 18944, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 37888, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 37888, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 37888, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3072, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3072, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3072, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 1024, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 1024, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 1024, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 14336, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 14336, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 14336, 4096, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 512, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 512, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 512, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 18944, 3584, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 18944, 3584, 3584, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 18944, 3584, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3072, 8192, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3072, 8192, 8192, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 3072, 8192, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 16384, 3072, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 16384, 3072, 3072, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 16384, 3072, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2048, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2048, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2048, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2560, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2560, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2560, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2048, 11008, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2048, 11008, 11008, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 512, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2048, 11008, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 22016, 2048, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 3, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 22016, 2048, 2048, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 22016, 2048, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 7168, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 7168, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 7168, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 13824, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 13824, 13824, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 5120, 13824, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 27648, 5120, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 3, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 27648, 5120, 5120, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 3, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 27648, 5120, 128, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2048, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(4096, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(3072, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(2048, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1536, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(1024, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(768, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(384, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(256, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(192, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(96, 2048, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4096, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4096, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 4096, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 4096, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 8192, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 8192, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 8192, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 8192, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3840, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3840, 4096, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3840, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3840, 4096, 4096, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3840, 15360, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3840, 15360, 64, 8, 104)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 32, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3840, 15360, 15360, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 3840, 15360, 15360, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 256, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 4, "num_ctas": 1, "num_stages": 5, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 30720, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 256, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "num_warps": 4, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 30720, 3840, 64, 8, 104)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 128, "BLOCK_SIZE_K": 32, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 1, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 30720, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 128, "BLOCK_SIZE_N": 256, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(128, 30720, 3840, 3840, 1, 408)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 512, "BLOCK_SIZE_K": 128, "GROUP_SIZE_M": 8, "A_load_order": 2, "num_warps": 8, "num_ctas": 1, "num_stages": 3, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}, "(512, 3072, 3072, 128, 4, 102)": {"BLOCK_SIZE_M": 64, "BLOCK_SIZE_N": 64, "BLOCK_SIZE_K": 64, "GROUP_SIZE_M": 8, "A_load_order": 0, "NUM_STAGES": 4, "num_warps": 4, "num_ctas": 1, "num_stages": 4, "num_buffers_warp_spec": 0, "num_consumer_groups": 0, "reg_dec_producer": 0, "reg_inc_consumer": 0}}} |