From 3ef9fd0f989d7dfc5e3633c0dcfc64acbaf2f8c5 Mon Sep 17 00:00:00 2001
From: Michael Goin <mgoin64@gmail.com>
Date: Mon, 23 Feb 2026 20:11:27 -0500
Subject: [PATCH] [Bugfix] Fix DSV3 kernels breaking _C and _moe_C on
 unsupported arches (#35123)

Signed-off-by: mgoin <mgoin64@gmail.com>
---
 CMakeLists.txt                     | 1 -
 csrc/dsv3_fused_a_gemm.cu          | 4 ++++
 csrc/moe/dsv3_router_gemm_entry.cu | 6 ++++++
 csrc/moe/torch_bindings.cpp        | 2 +-
 csrc/torch_bindings.cpp            | 2 +-
 5 files changed, 12 insertions(+), 3 deletions(-)

diff --git a/CMakeLists.txt b/CMakeLists.txt
index a6f7f69468d..55127a514f1 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -783,7 +783,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       SRCS "${DSV3_FUSED_A_GEMM_SRC}"
       CUDA_ARCHS "${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")
     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 "
diff --git a/csrc/dsv3_fused_a_gemm.cu b/csrc/dsv3_fused_a_gemm.cu
index 5b8374303ad..65dff9c84ba 100644
--- a/csrc/dsv3_fused_a_gemm.cu
+++ b/csrc/dsv3_fused_a_gemm.cu
@@ -745,3 +745,7 @@ void dsv3_fused_a_gemm(torch::Tensor& output, torch::Tensor const& mat_a,
         stream);
   }
 }
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("dsv3_fused_a_gemm", &dsv3_fused_a_gemm);
+}
diff --git a/csrc/moe/dsv3_router_gemm_entry.cu b/csrc/moe/dsv3_router_gemm_entry.cu
index 1ba97bd7640..38fb681c223 100644
--- a/csrc/moe/dsv3_router_gemm_entry.cu
+++ b/csrc/moe/dsv3_router_gemm_entry.cu
@@ -20,10 +20,12 @@
 
 #include <ATen/ATen.h>
 #include <ATen/cuda/CUDAContext.h>
+#include <torch/all.h>
 
 #include <cuda_bf16.h>
 #include <cuda_runtime.h>
 
+#include "core/registration.h"
 #include "dsv3_router_gemm_utils.h"
 
 static constexpr int DEFAULT_NUM_EXPERTS = 256;
@@ -161,3 +163,7 @@ void dsv3_router_gemm(at::Tensor& output,       // [num_tokens, num_experts]
     }
   }
 }
+
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
+  m.impl("dsv3_router_gemm", &dsv3_router_gemm);
+}
diff --git a/csrc/moe/torch_bindings.cpp b/csrc/moe/torch_bindings.cpp
index 22b00f20ad5..43859945145 100644
--- a/csrc/moe/torch_bindings.cpp
+++ b/csrc/moe/torch_bindings.cpp
@@ -127,7 +127,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, m) {
 
   // DeepSeek V3 optimized router GEMM for SM90+
   m.def("dsv3_router_gemm(Tensor! output, Tensor mat_a, Tensor mat_b) -> ()");
-  m.impl("dsv3_router_gemm", torch::kCUDA, &dsv3_router_gemm);
+  // conditionally compiled so impl registration is in source file
 #endif
 }
 
diff --git a/csrc/torch_bindings.cpp b/csrc/torch_bindings.cpp
index c16b9c223f6..39b6bc98a84 100644
--- a/csrc/torch_bindings.cpp
+++ b/csrc/torch_bindings.cpp
@@ -242,7 +242,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
   // DeepSeek V3 fused A GEMM (SM 9.0+, bf16 only, 1-16 tokens).
   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);
+  // conditionally compiled so impl registration is in source file
 
   // Quantized GEMM for AWQ.
   ops.def(