[Feature] Integrate SM100 DeepGEMM support (#20087)

2025-07-10 23:18:05 -04:00
parent 5b032352cc
commit e2de455c34
16 changed files with 397 additions and 114 deletions
--- a/benchmarks/kernels/benchmark_moe.py
+++ b/benchmarks/kernels/benchmark_moe.py
@@ -86,6 +86,9 @@ def benchmark_config(
            (num_experts, 2 * shard_intermediate_size), dtype=torch.float32
        )
        w2_scale = torch.randn((hidden_size, num_experts), dtype=torch.float32)
+    if use_deep_gemm:
+        # we use the default block shape for deepgemm
+        block_quant_shape = [128, 128]
    if use_fp8_w8a8:
        if block_quant_shape:
            block_n, block_k = block_quant_shape[0], block_quant_shape[1]