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 15, 03:30 UTC
[BugFix] Fix: ImportError when building on hopper systems (#20513)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Lucas Wilkinson committed
Jul 6, 2025 at 00:17 UTC
40b86aa05e4458bf28f038666942d620d89c8c3d
4 files changed
+10
-9
.github/CODEOWNERS
+1
-1
@@ -16,7 +16,7 @@
16
/vllm/lora @jeejeelee
17
/vllm/reasoning @aarnphm
18
/vllm/entrypoints @aarnphm
19
-CMakeLists.txt @tlrmchlsmth
19
+CMakeLists.txt @tlrmchlsmth @LucasWilkinson
20
21
# Any change to the VllmConfig changes can have a large user-facing impact,
22
# so spam a lot of people
csrc/ops.h
-5
@@ -239,11 +239,6 @@ void cutlass_moe_mm(
239
torch::Tensor const& b_strides, torch::Tensor const& c_strides,
240
bool per_act_token, bool per_out_ch);
241
242
-void cutlass_blockwise_scaled_grouped_mm(
243
- torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
244
- const torch::Tensor& scales_a, const torch::Tensor& scales_b,
245
- const torch::Tensor& problem_sizes, const torch::Tensor& expert_offsets);
246
-
242
void cutlass_fp4_group_mm(
243
torch::Tensor& output, const torch::Tensor& a, const torch::Tensor& b,
244
const torch::Tensor& a_blockscale, const torch::Tensor& b_blockscales,
csrc/quantization/cutlass_w8a8/moe/blockwise_scaled_group_mm_sm100.cu
+8
-1
@@ -1,3 +1,5 @@
1
+#include "core/registration.h"
2
+
3
#include <torch/all.h>
4
#include <cutlass/arch/arch.h>
5
@@ -364,4 +366,9 @@ void cutlass_blockwise_scaled_grouped_mm(
366
TORCH_CHECK(false, "Unsupported output tensor type");
367
}
368
#endif
367
-}
\ No newline at end of file
369
+}
370
+
371
+TORCH_LIBRARY_IMPL_EXPAND(TORCH_EXTENSION_NAME, CUDA, m) {
372
+ m.impl("cutlass_blockwise_scaled_grouped_mm",
373
+ &cutlass_blockwise_scaled_grouped_mm);
374
+}
csrc/torch_bindings.cpp
+1
-2
@@ -399,8 +399,7 @@ TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
399
"Tensor scales_a, Tensor scales_b, "
400
"Tensor problem_sizes, Tensor expert_offsets) -> ()",
401
{stride_tag});
402
- ops.impl("cutlass_blockwise_scaled_grouped_mm", torch::kCUDA,
403
- &cutlass_blockwise_scaled_grouped_mm);
402
+ // conditionally compiled so impl registration is in source file
403
404
// cutlass nvfp4 block scaled group GEMM
405
ops.def(