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 2, 04:45 UTC
Migrate header files to torch stable abi (#44013)
Chris Leonard committed
Jun 2, 2026 at 11:09 UTC
4d93bc35c9f4a1b470ced63392f640feddf7789b
18 files changed
+13
-16
.pre-commit-config.yaml
+1
-1
@@ -21,7 +21,7 @@ repos:
21
rev: v21.1.2
22
hooks:
23
- id: clang-format
24
- exclude: 'csrc/(moe/topk_softmax_kernels.cu|quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
24
+ exclude: 'csrc/(moe/topk_softmax_kernels.cu|libtorch_stable/quantization/gguf/(ggml-common.h|dequantize.cuh|vecdotq.cuh|mmq.cuh|mmvq.cuh))|vllm/third_party/.*'
25
types_or: [c++, cuda]
26
args: [--style=file, --verbose]
27
- repo: https://github.com/DavidAnson/markdownlint-cli2
csrc/libtorch_stable/async_util.cuh
renamed
csrc/libtorch_stable/cutlass_extensions/epilogue/broadcast_load_epilogue_c2x.hpp
renamed
csrc/libtorch_stable/cutlass_extensions/epilogue/scaled_mm_epilogues_c2x.hpp
+1
-1
@@ -2,7 +2,7 @@
2
3
#include <torch/csrc/stable/tensor.h>
4
5
-#include "cutlass_extensions/epilogue/broadcast_load_epilogue_c2x.hpp"
5
+#include "broadcast_load_epilogue_c2x.hpp"
6
7
/*
8
This file defines custom epilogues for fusing channel scales, token scales,
csrc/libtorch_stable/fused_qknorm_rope_kernel.cu
+1
-1
@@ -20,7 +20,7 @@
20
21
#include "torch_utils.h"
22
23
-#include "../async_util.cuh"
23
+#include "async_util.cuh"
24
#include "../cuda_compat.h"
25
#include "../type_convert.cuh"
26
#include "dispatch_utils.h"
csrc/libtorch_stable/launch_bounds_utils.h
renamed
csrc/libtorch_stable/persistent_topk.cuh
renamed
csrc/libtorch_stable/quantization/fp4/activation_nvfp4_quant_fusion_kernels.cu
+1
-1
@@ -25,7 +25,7 @@
25
#include <cuda_fp8.h>
26
27
#include "cuda_utils.h"
28
-#include "launch_bounds_utils.h"
28
+#include "libtorch_stable/launch_bounds_utils.h"
29
30
// Define before including nvfp4_utils.cuh so the header
31
// can use this macro during compilation.
csrc/libtorch_stable/quantization/fp4/mxfp4_experts_quant.cu
+1
-1
@@ -34,7 +34,7 @@
34
static_assert(CVT_FP4_ELTS_PER_THREAD == 16,
35
"MXFP4 experts quant requires PACK16 mode (CUDA >= 12.9)");
36
37
-#include "launch_bounds_utils.h"
37
+#include "libtorch_stable/launch_bounds_utils.h"
38
39
namespace vllm {
40
csrc/libtorch_stable/quantization/fp4/nvfp4_experts_quant.cu
+1
-1
@@ -26,7 +26,7 @@
26
27
#include "cuda_utils.h"
28
#include "nvfp4_utils.cuh"
29
-#include "launch_bounds_utils.h"
29
+#include "libtorch_stable/launch_bounds_utils.h"
30
31
namespace vllm {
32
csrc/libtorch_stable/quantization/fp4/nvfp4_quant_kernels.cu
+1
-1
@@ -26,7 +26,7 @@
26
#include "../../cuda_vec_utils.cuh"
27
28
#include "cuda_utils.h"
29
-#include "launch_bounds_utils.h"
29
+#include "libtorch_stable/launch_bounds_utils.h"
30
31
// Define before including nvfp4_utils.cuh so the header
32
// can use this macro during compilation.
csrc/libtorch_stable/quantization/gguf/dequantize.cuh
renamed
csrc/libtorch_stable/quantization/gguf/ggml-common.h
renamed
csrc/libtorch_stable/quantization/gguf/gguf_kernel.cu
+5
-8
@@ -7,14 +7,11 @@
7
8
#include <torch/csrc/stable/ops.h>
9
10
-// NOTE: These headers are intentionally kept in csrc/quantization/gguf/ (not
11
-// moved to libtorch_stable) to avoid unnecessary reformatting that would break
12
-// git rename detection and pollute blame history.
13
-#include "../../../quantization/gguf/ggml-common.h"
14
-#include "../../../quantization/gguf/vecdotq.cuh"
15
-#include "../../../quantization/gguf/dequantize.cuh"
16
-#include "../../../quantization/gguf/mmvq.cuh"
17
-#include "../../../quantization/gguf/mmq.cuh"
10
+#include "ggml-common.h"
11
+#include "vecdotq.cuh"
12
+#include "dequantize.cuh"
13
+#include "mmvq.cuh"
14
+#include "mmq.cuh"
15
#include "moe.cuh"
16
#include "moe_vec.cuh"
17
csrc/libtorch_stable/quantization/gguf/mmq.cuh
renamed
csrc/libtorch_stable/quantization/gguf/mmvq.cuh
renamed
csrc/libtorch_stable/quantization/gguf/vecdotq.cuh
renamed
csrc/libtorch_stable/topk.cu
+1
-1
@@ -7,7 +7,7 @@
7
#include "torch_utils.h"
8
9
#ifndef USE_ROCM
10
- #include "../persistent_topk.cuh"
10
+ #include "persistent_topk.cuh"
11
#endif
12
13
namespace {