[Perf] Extend Qwen Triton warmup to avoid first-request latency spikes (#54797)

Signed-off-by: Juqi Li <2223621784@qq.com>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
vhagor
2026-09-07 11:46:00 +08:00
committed by GitHub
co-authored by Isotr0py
parent 6748217fb9
commit 3dc7a68ce4
7 changed files with 368 additions and 19 deletions
@@ -34,10 +34,12 @@ from vllm.model_executor.warmup.flashinfer_sparse_mla_warmup import (
from vllm.model_executor.warmup.kimi_k3_triton_warmup import (
kimi_k3_triton_warmup,
)
from vllm.model_executor.warmup.mamba_triton_warmup import mamba_triton_warmup
from vllm.model_executor.warmup.qwen4_exp_qsa_warmup import (
qwen4_exp_qsa_triton_warmup,
)
from vllm.model_executor.warmup.qwen_triton_warmup import qwen_triton_warmup
from vllm.model_executor.warmup.qwen_vl_triton_warmup import qwen_vl_triton_warmup
from vllm.model_executor.warmup.replayssm_warmup import (
replayssm_autotune_warmup,
)
@@ -149,6 +151,8 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
)
qwen_triton_warmup(worker.model_runner, worker.vllm_config.model_config)
qwen_vl_triton_warmup(worker.model_runner)
mamba_triton_warmup(worker.model_runner)
compilation_config = worker.vllm_config.compilation_config
cudagraph_capture_sizes = list(compilation_config.cudagraph_capture_sizes or [])