[Bugfix] [ROCm] [Critical] fallback to regular abi for ROCm (#44648)

Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
TJian
2026-06-05 15:51:17 +00:00
committed by GitHub
parent 6a894574bf
commit aa6fb8a329
6 changed files with 127 additions and 25 deletions
+16 -8
View File
@@ -590,8 +590,11 @@ endif()
if (VLLM_GPU_LANG STREQUAL "HIP")
# Add QuickReduce kernels (ROCm-only; not part of stable ABI migration).
# TODO: Remove the cuda_view when ROCm upgrade to torch 2.11.
list(APPEND VLLM_EXT_SRC
"csrc/custom_quickreduce.cu"
"csrc/cuda_view.cu"
"csrc/libtorch_stable/cuda_utils_kernels.cu"
)
# if ROCM endif
endif()
@@ -621,7 +624,6 @@ 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"
@@ -649,6 +651,7 @@ 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/libtorch_stable/cuda_view.cu"
"csrc/libtorch_stable/cuda_utils_kernels.cu"
"csrc/libtorch_stable/cutlass_extensions/common.cpp"
"csrc/libtorch_stable/quantization/w8a8/cutlass/scaled_mm_entry.cu"
@@ -1071,20 +1074,25 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" OR VLLM_GPU_LANG STREQUAL "HIP")
USE_SABI 3
WITH_SOABI)
# Set TORCH_TARGET_VERSION for stable ABI compatibility.
# 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 2.11.
target_compile_definitions(_C_stable_libtorch PRIVATE
TORCH_TARGET_VERSION=0x020B000000000000ULL)
# Needed to use cuda/hip APIs from C-shim
if(VLLM_GPU_LANG STREQUAL "CUDA")
# Set TORCH_TARGET_VERSION for stable ABI compatibility.
# 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 2.11.
target_compile_definitions(_C_stable_libtorch PRIVATE
TORCH_TARGET_VERSION=0x020B000000000000ULL)
target_compile_definitions(_C_stable_libtorch PRIVATE USE_CUDA)
# Needed by CUTLASS kernels
target_compile_definitions(_C_stable_libtorch PRIVATE
CUTLASS_ENABLE_DIRECT_CUDA_DRIVER_CALL=1)
elseif(VLLM_GPU_LANG STREQUAL "HIP")
# Set TORCH_TARGET_VERSION for stable ABI compatibility.
# This ensures we only use C-shim APIs available in PyTorch 2.10.
# _C_stable_libtorch is abi compatible with PyTorch >= TORCH_TARGET_VERSION
# which is currently set to 2.10.
target_compile_definitions(_C_stable_libtorch PRIVATE
TORCH_TARGET_VERSION=0x020A000000000000ULL)
target_compile_definitions(_C_stable_libtorch PRIVATE USE_ROCM)
endif()