[10/n] Migrate cuda_view and silu_and_mul_per_block_quant kernels to torch stale ABI. (#44334)

This commit is contained in:
Chris Leonard
2026-06-04 20:14:43 -07:00
committed by GitHub
parent 063ce98fb7
commit 56aff0dd15
25 changed files with 186 additions and 163 deletions
+7 -11
View File
@@ -307,10 +307,7 @@ endif()
#
set(VLLM_EXT_SRC
"csrc/cuda_view.cu"
"csrc/quantization/fused_kernels/fused_silu_mul_block_quant.cu"
"csrc/quantization/activation_kernels.cu"
"csrc/cuda_utils_kernels.cu"
"csrc/torch_bindings.cpp")
if(VLLM_GPU_LANG STREQUAL "CUDA")
@@ -346,9 +343,6 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
endif()
FetchContent_MakeAvailable(cutlass)
list(APPEND VLLM_EXT_SRC
"csrc/cutlass_extensions/common.cpp")
set_gencode_flags_for_srcs(
SRCS "${VLLM_EXT_SRC}"
CUDA_ARCHS "${CUDA_ARCHS}")
@@ -627,6 +621,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
#
set(VLLM_STABLE_EXT_SRC
"csrc/libtorch_stable/torch_bindings.cpp"
"csrc/libtorch_stable/cuda_view.cu"
"csrc/libtorch_stable/activation_kernels.cu"
"csrc/libtorch_stable/quantization/w8a8/int8/scaled_quant.cu"
"csrc/libtorch_stable/quantization/w8a8/fp8/common.cu"
@@ -639,6 +634,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
"csrc/libtorch_stable/layernorm_kernels.cu"
"csrc/libtorch_stable/layernorm_quant_kernels.cu"
"csrc/libtorch_stable/quantization/fused_kernels/fused_layernorm_dynamic_per_token_quant.cu"
"csrc/libtorch_stable/quantization/fused_kernels/fused_silu_mul_block_quant.cu"
"csrc/libtorch_stable/attention/merge_attn_states.cu"
"csrc/libtorch_stable/sampler.cu"
"csrc/libtorch_stable/topk.cu"
@@ -653,8 +649,8 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
if(VLLM_GPU_LANG STREQUAL "CUDA")
list(APPEND VLLM_STABLE_EXT_SRC
"csrc/cuda_utils_kernels.cu"
"csrc/cutlass_extensions/common.cpp"
"csrc/libtorch_stable/cuda_utils_kernels.cu"
"csrc/libtorch_stable/cutlass_extensions/common.cpp"
"csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_entry.cu"
"csrc/libtorch_stable/quantization/fp4/nvfp4_quant_entry.cu"
"csrc/libtorch_stable/quantization/fp4/nvfp4_scaled_mm_entry.cu"
@@ -1076,11 +1072,11 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
WITH_SOABI)
# Set TORCH_TARGET_VERSION for stable ABI compatibility.
# This ensures we only use C-shim APIs available in PyTorch 2.10.
# This ensures we only use C-shim APIs available in PyTorch 2.11.
# _C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
# which is currently set to 2.10.
# which is currently set to 2.11.
target_compile_definitions(_C_stable_libtorch PRIVATE
TORCH_TARGET_VERSION=0x020A000000000000ULL)
TORCH_TARGET_VERSION=0x020B000000000000ULL)
# Needed to use cuda/hip APIs from C-shim
if(VLLM_GPU_LANG STREQUAL "CUDA")