Implement custom all reduce kernels (#2192)

This commit is contained in:
Hanzhi Zhou
2024-01-27 12:46:35 -08:00
committed by GitHub
parent 220a47627b
commit 380170038e
18 changed files with 1453 additions and 65 deletions
+4 -2
View File
@@ -51,8 +51,8 @@ if _is_hip():
"Cannot find ROCM_HOME. ROCm must be available to build the package."
)
NVCC_FLAGS += ["-DUSE_ROCM"]
NVCC_FLAGS += [f"-U__HIP_NO_HALF_CONVERSIONS__"]
NVCC_FLAGS += [f"-U__HIP_NO_HALF_OPERATORS__"]
NVCC_FLAGS += ["-U__HIP_NO_HALF_CONVERSIONS__"]
NVCC_FLAGS += ["-U__HIP_NO_HALF_OPERATORS__"]
if _is_cuda() and CUDA_HOME is None:
raise RuntimeError(
@@ -307,6 +307,7 @@ vllm_extension_sources = [
if _is_cuda():
vllm_extension_sources.append("csrc/quantization/awq/gemm_kernels.cu")
vllm_extension_sources.append("csrc/custom_all_reduce.cu")
if not _is_neuron():
vllm_extension = CUDAExtension(
@@ -316,6 +317,7 @@ if not _is_neuron():
"cxx": CXX_FLAGS,
"nvcc": NVCC_FLAGS,
},
libraries=["cuda"] if _is_cuda() else [],
)
ext_modules.append(vllm_extension)