biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[BugFix][Performance] Restore flashinfer autotuning for all scenarios (#27904)

Browse Source
This commit is contained in:
Varun Sundar Rabindranath
2025-11-04 02:56:21 -05:00
committed by GitHub
parent 53f6e81dfd
commit 4022a9d279
4 changed files with 14 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
vllm/model_executor/warmup/kernel_warmup.py
View File

@@ -11,7 +11,6 @@ from typing import TYPE_CHECKING
import torch
import vllm.envs as envs
from vllm.config import CUDAGraphMode, VllmConfig
from vllm.logger import init_logger
from vllm.model_executor.warmup.deep_gemm_warmup import deep_gemm_warmup
from vllm.platforms import current_platform
@@ -25,26 +24,6 @@ if TYPE_CHECKING:
logger = init_logger(__name__)
def flashinfer_autotune_supported(vllm_config: VllmConfig) -> bool:
"""
Record known issues with vllm + flashinfer autotune here. Return True if
and only if flashinfer autotune will run through without issues.
"""
is_tp_or_dp = (vllm_config.parallel_config.data_parallel_size > 1) or (
vllm_config.parallel_config.tensor_parallel_size > 1
)
is_fi_mxfp4_backend = (
envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16
or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS
) or (
current_platform.is_cuda() and current_platform.is_device_capability(100)
) # on >=sm100, default mxfp4 backend is flashinfer
is_eager = vllm_config.compilation_config.cudagraph_mode == CUDAGraphMode.NONE
return not (is_tp_or_dp and is_fi_mxfp4_backend and is_eager)
def kernel_warmup(worker: "Worker"):
# Deep GEMM warmup
do_deep_gemm_warmup = (
@@ -58,11 +37,7 @@ def kernel_warmup(worker: "Worker"):
deep_gemm_warmup(model, max_tokens)
# FlashInfer autotune for Hopper (SM 9.0) and Blackwell (SM 10.0) GPUs
if (
has_flashinfer()
and current_platform.has_device_capability(90)
and flashinfer_autotune_supported(worker.vllm_config)
):
if has_flashinfer() and current_platform.has_device_capability(90):
flashinfer_autotune(worker.model_runner)
# FlashInfer attention warmup