[Build] Fix DSV3_FUSED_A_GEMM_ARCHS to only include SM 9.0 (Hopper)

CMakeLists.txt

@@ -771,11 +771,12 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
     endif()
   endif()
 
-  # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper and later)
+  # DeepSeek V3 fused A GEMM kernel (requires SM 9.0+, Hopper only).
+  # SM 10.0/11.0/12.0 are excluded as this kernel is only intended for datacenter GPUs.
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 13.0)
-    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0f;11.0f" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a" "${CUDA_ARCHS}")
   else()
-    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a;10.0a;10.1a;10.3a" "${CUDA_ARCHS}")
+    cuda_archs_loose_intersection(DSV3_FUSED_A_GEMM_ARCHS "9.0a" "${CUDA_ARCHS}")
   endif()
   if(${CMAKE_CUDA_COMPILER_VERSION} VERSION_GREATER_EQUAL 12.0 AND DSV3_FUSED_A_GEMM_ARCHS)
     set(DSV3_FUSED_A_GEMM_SRC "csrc/dsv3_fused_a_gemm.cu")
@@ -784,6 +785,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       CUDA_ARCHS "${DSV3_FUSED_A_GEMM_ARCHS}")
     list(APPEND VLLM_EXT_SRC ${DSV3_FUSED_A_GEMM_SRC})
     list(APPEND VLLM_GPU_FLAGS "-DENABLE_DSV3_FUSED_A_GEMM=1")
+    add_compile_definitions(ENABLE_DSV3_FUSED_A_GEMM=1)
     message(STATUS "Building dsv3_fused_a_gemm for archs: ${DSV3_FUSED_A_GEMM_ARCHS}")
   else()
     message(STATUS "Not building dsv3_fused_a_gemm as no compatible archs found "

csrc/ops.h

@@ -412,6 +412,8 @@ int64_t qr_max_size();
 #endif
 
 #ifndef USE_ROCM
+  #ifdef ENABLE_DSV3_FUSED_A_GEMM
 void dsv3_fused_a_gemm(torch::Tensor& output, torch::Tensor const& mat_a,
                        torch::Tensor const& mat_b);
+  #endif
 #endif

csrc/torch_bindings.cpp

@@ -240,9 +240,12 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // Quantization ops
 #ifndef USE_ROCM
   // DeepSeek V3 fused A GEMM (SM 9.0+, bf16 only, 1-16 tokens).
+  // Only available when the kernel is built for compatible architectures.
+  #ifdef ENABLE_DSV3_FUSED_A_GEMM
   ops.def(
       "dsv3_fused_a_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
   ops.impl("dsv3_fused_a_gemm", torch::kCUDA, &dsv3_fused_a_gemm);
+  #endif
 
   // Quantized GEMM for AWQ.
   ops.def(

vllm/_custom_ops.py

@@ -2789,22 +2789,39 @@ def sm100_cutlass_mla_get_workspace_size(
     )
 
 
-def dsv3_fused_a_gemm(
-    output: torch.Tensor,
-    mat_a: torch.Tensor,
-    mat_b: torch.Tensor,
-) -> None:
-    """DeepSeek V3 fused A GEMM (SM 9.0+, bf16 only, 1-16 tokens).
+if hasattr(torch.ops._C, "dsv3_fused_a_gemm"):
 
-    Computes output = mat_a @ mat_b.T where:
-      mat_a: [num_tokens, 7168] row-major bf16 (hidden states)
-      mat_b: [7168, 2112] column-major bf16 (weight transposed)
-      output: [num_tokens, 2112] row-major bf16
+    def dsv3_fused_a_gemm(
+        output: torch.Tensor,
+        mat_a: torch.Tensor,
+        mat_b: torch.Tensor,
+    ) -> None:
+        """DeepSeek V3 fused A GEMM (SM 9.0+, bf16 only, 1-16 tokens).
 
-    Optimized for the DeepSeek V2/V3 QKV A-projection at small batch sizes.
-    Requires SM 9.0+ (Hopper).
-    """
-    torch.ops._C.dsv3_fused_a_gemm(output, mat_a, mat_b)
+        Computes output = mat_a @ mat_b.T where:
+          mat_a: [num_tokens, 7168] row-major bf16 (hidden states)
+          mat_b: [7168, 2112] column-major bf16 (weight transposed)
+          output: [num_tokens, 2112] row-major bf16
+
+        Optimized for the DeepSeek V2/V3 QKV A-projection at small batch sizes.
+        Requires SM 9.0+ (Hopper).
+        """
+        torch.ops._C.dsv3_fused_a_gemm(output, mat_a, mat_b)
+
+else:
+
+    def dsv3_fused_a_gemm(
+        output: torch.Tensor,
+        mat_a: torch.Tensor,
+        mat_b: torch.Tensor,
+    ) -> None:
+        """DeepSeek V3 fused A GEMM (SM 9.0+, bf16 only, 1-16 tokens).
+
+        This kernel is only available for SM 9.0 (Hopper) architectures.
+        """
+        raise RuntimeError(
+            "dsv3_fused_a_gemm is only supported on SM 9.0 (Hopper) architectures"
+        )
 
 
 if hasattr(torch.ops._C, "weight_packed_linear"):

vllm/model_executor/models/deepseek_v2.py

@@ -730,17 +730,15 @@ def __init__(
         )
 
         # Check if the DeepSeek V3 fused A GEMM kernel can be used.
-        # This kernel supports PDL and is optimized for low batch size.
+        # This kernel is only supported on Hopper (SM 9.0) and above.
+        # Note: SM 10.0/11.0/12.0 (GeForce Blackwell) are not supported.
         self._use_min_latency_gemm = (
             hasattr(self, "weight")
             and self.weight.dtype == torch.bfloat16
             and self.weight.shape[0] == 2112
             and self.weight.shape[1] == 7168
             and current_platform.is_cuda()
-            and (
-                current_platform.is_device_capability(90)
-                or current_platform.is_device_capability_family(100)
-            )
+            and current_platform.is_device_capability(90)
         )
 
     def forward(