[ROCm][AITER] Directly Implement AITER Custom All-reduce in CudaCommunicator (#46065)

Signed-off-by: BadrBasowid <badr.basowid@gmail.com>
This commit is contained in:
BadrBasowid
2026-07-06 12:16:32 +00:00
committed by GitHub
parent 40cc2e8327
commit 740f379fae
12 changed files with 499 additions and 148 deletions
+6
View File
@@ -119,6 +119,7 @@ if TYPE_CHECKING:
VLLM_USE_OINK_OPS: bool = False
VLLM_MXFP8_EMULATION_DEQUANT_AT_LOAD: bool = True
VLLM_ROCM_USE_AITER: bool = False
VLLM_ROCM_USE_AITER_CUSTOM_AR: bool = True
VLLM_ROCM_USE_AITER_PAGED_ATTN: bool = False
VLLM_ROCM_USE_AITER_LINEAR: bool = True
VLLM_ROCM_USE_AITER_LINEAR_HIPBMM: bool = False
@@ -1146,6 +1147,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_ROCM_USE_AITER": lambda: (
os.getenv("VLLM_ROCM_USE_AITER", "False").lower() in ("true", "1")
),
# Use AITER's CustomAllreduce as the custom-allreduce backend inside vLLM's
# CudaCommunicator on ROCm.
"VLLM_ROCM_USE_AITER_CUSTOM_AR": lambda: (
os.getenv("VLLM_ROCM_USE_AITER_CUSTOM_AR", "True").lower() in ("true", "1")
),
# Whether to use aiter paged attention.
# By default is disabled.
"VLLM_ROCM_USE_AITER_PAGED_ATTN": lambda: (