A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 00:30 UTC
[BugFix] Stopgap - Flashinfer Autotuner + GPT-OSS + DP/TP (#27762)
Signed-off-by: Varun Sundar Rabindranath <vsundarr@redhat.com> Co-authored-by: Varun Sundar Rabindranath <vsundarr@redhat.com>
Varun Sundar Rabindranath committed
Oct 30, 2025 at 11:24 UTC
e5e076cad7c1c922fa6d48049c45bead505f52a6
1 file changed
+13
-7
vllm/model_executor/warmup/kernel_warmup.py
+13
-7
@@ -11,7 +11,7 @@ from typing import TYPE_CHECKING
11
import torch
12
13
import vllm.envs as envs
14
-from vllm.config import VllmConfig
14
+from vllm.config import CUDAGraphMode, VllmConfig
15
from vllm.logger import init_logger
16
from vllm.model_executor.warmup.deep_gemm_warmup import deep_gemm_warmup
17
from vllm.platforms import current_platform
@@ -30,13 +30,19 @@ def flashinfer_autotune_supported(vllm_config: VllmConfig) -> bool:
30
Record known issues with vllm + flashinfer autotune here. Return True if
31
and only if flashinfer autotune will run through without issues.
32
"""
33
- return not (
34
- vllm_config.parallel_config.data_parallel_size > 1
35
- and (
36
- envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16
37
- or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
38
- )
33
+ is_tp_or_dp = (vllm_config.parallel_config.data_parallel_size > 1) or (
34
+ vllm_config.parallel_config.tensor_parallel_size > 1
35
)
36
+ is_fi_mxfp4_backend = (
37
+ envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8
38
+ or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_BF16
39
+ or envs.VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8_CUTLASS
40
+ ) or (
41
+ current_platform.is_cuda() and current_platform.is_device_capability(100)
42
+ ) # on >=sm100, default mxfp4 backend is flashinfer
43
+ is_eager = vllm_config.compilation_config.cudagraph_mode == CUDAGraphMode.NONE
44
+
45
+ return not (is_tp_or_dp and is_fi_mxfp4_backend and is_eager)
46
47
48
def kernel_warmup(worker: "Worker"):