[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
@@ -13,6 +13,9 @@ from typing import TYPE_CHECKING
import torch
import vllm.envs as envs
from vllm.distributed.device_communicators.flashinfer_pcie_ipc_all_reduce import (
warmup_flashinfer_pcie_ipc_allreduce,
)
from vllm.logger import init_logger
from vllm.model_executor.warmup.b12x_warmup import b12x_warmup
from vllm.model_executor.warmup.cutedsl_warmup import cutedsl_warmup
@@ -182,6 +185,10 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
minimax_m3_msa_warmup(worker)
# Allocate the exact decode-sized workspace, autotune cache misses, and
# resolve every CUDA Graph bucket before capture begins.
warmup_flashinfer_pcie_ipc_allreduce(worker)
enable_flashinfer_autotune = (
worker.vllm_config.kernel_config.enable_flashinfer_autotune
)