[Fast Start] Preload the FlashInfer autotune table on the weight cache daemon (#60085)

Signed-off-by: liusy58 <mg21330037@smail.nju.edu.cn>
Signed-off-by: Isotr0py <Isotr0py@outlook.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
Co-authored-by: Isotr0py <Isotr0py@outlook.com>
This commit is contained in:
siyu
2026-10-10 15:48:06 +00:00
committed by GitHub
co-authored by Claude Opus 5 Isotr0py
parent 3709632ff2
commit f2376cdc4d
10 changed files with 253 additions and 17 deletions
+8 -1
View File
@@ -39,6 +39,7 @@ from vllm.logger import init_logger
from vllm.model_executor.model_loader.weight_cache.protocol import (
check_ipc_platform_support,
)
from vllm.usage.usage_lib import UsageContext
from vllm.utils.network_utils import get_distributed_init_method, get_open_port
if typing.TYPE_CHECKING:
@@ -157,7 +158,13 @@ class PreloadSubcommand(CLISubcommand):
)
engine_args = EngineArgs.from_cli_args(args)
vllm_config = engine_args.create_engine_config()
# Warmup context for flashinfer autotune.
usage_context = (
UsageContext.OPENAI_API_SERVER
if engine_args.kernel_config.enable_flashinfer_autotune is not False
else None
)
vllm_config = engine_args.create_engine_config(usage_context=usage_context)
if vllm_config.load_config.load_format == "ipc_cache":
raise ValueError(
"The weight cache daemon itself must load from disk; use the "