[Bugfix] Fix DSV3 kernels breaking _C and _moe_C on unsupported arches (#35123)
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
@@ -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);
|
||||
}
|
||||
|
||||
@@ -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
|
||||
}
|
||||
|
||||
|
||||
Reference in New Issue
Block a user