A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[CI/Build] Conditionally register cutlass_fp4_group_mm to fix building on Hopper (#26138)
Signed-off-by: mgoin <mgoin64@gmail.com>
Michael Goin committed
Oct 2, 2025 at 23:32 UTC
5d5146eee35a989f14de6f4203cabd77174fa1ba
2 files changed
+7
-1
csrc/quantization/fp4/nvfp4_blockwise_moe_kernel.cu
+6
@@ -14,6 +14,8 @@
14
* limitations under the License.
15
*/
16
17
+#include "core/registration.h"
18
+
19
#include <torch/all.h>
20
#include <cutlass/arch/arch.h>
21
@@ -418,3 +420,7 @@ void cutlass_fp4_group_mm(
420
"12.8 or above.");
421
#endif
422
}
423
+
424
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
425
+ m.impl("cutlass_fp4_group_mm", &cutlass_fp4_group_mm);
426
+}
csrc/torch_bindings.cpp
+1
-1
@@ -397,7 +397,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
397
" Tensor a_blockscale, Tensor b_blockscales, Tensor alphas,"
398
" Tensor problem_sizes, Tensor expert_offsets, Tensor sf_offsets) -> ()",
399
{stride_tag});
400
- ops.impl("cutlass_fp4_group_mm", torch::kCUDA, &cutlass_fp4_group_mm);
400
+ // conditionally compiled so impl registration is in source file
401
402
// CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
403
// quantization, as well as bias