[Distributed] Add opt-in FlashInfer PCIe IPC all-reduce backend (#53576)

Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com>
Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Gabriel Wu
2026-09-01 03:34:56 +00:00
committed by GitHub
co-authored by Yongye Zhu mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent e16b5e518d
commit 446c769482
7 changed files with 450 additions and 4 deletions
+7
View File
@@ -264,6 +264,7 @@ if TYPE_CHECKING:
VLLM_ROCM_FP8_MFMA_PAGE_ATTN: bool = False
VLLM_ALLREDUCE_USE_SYMM_MEM: bool = True
VLLM_ALLREDUCE_USE_FLASHINFER: bool = True
VLLM_ALLREDUCE_USE_FLASHINFER_PCIE_IPC: bool = False
VLLM_TUNED_CONFIG_FOLDER: str | None = None
VLLM_ENABLE_STARTUP_PLAN: bool = False
VLLM_GPT_OSS_SYSTEM_TOOL_MCP_LABELS: set[str] = set()
@@ -1864,6 +1865,12 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_ALLREDUCE_USE_FLASHINFER": lambda: bool(
int(os.getenv("VLLM_ALLREDUCE_USE_FLASHINFER", "1"))
),
# Whether to use FlashInfer's single-node PCIe CUDA-IPC all-reduce.
# This backend has a strict single-stream contract and is opt-in while its
# integration is being qualified.
"VLLM_ALLREDUCE_USE_FLASHINFER_PCIE_IPC": lambda: bool(
int(os.getenv("VLLM_ALLREDUCE_USE_FLASHINFER_PCIE_IPC", "0"))
),
# Experimental: use this to enable MCP tool calling for non harmony models
"VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT": lambda: bool(
int(os.getenv("VLLM_USE_EXPERIMENTAL_PARSER_CONTEXT", "0"))