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 6, 03:15 UTC
[Bugfix] Suppress spurious CPU KV cache warning in `launch render` (#37911)
Signed-off-by: Sage Ahrac <sagiahrak@gmail.com>
Sage committed
Mar 24, 2026 at 14:36 UTC
1c0aabdeb0cf77019a1f89b5bed5b8eebdd5c211
1 file changed
+6
-2
vllm/entrypoints/cli/launch.py
+6
-2
@@ -5,6 +5,8 @@ import argparse
5
6
import uvloop
7
8
+from vllm import envs
9
+from vllm.config import VllmConfig
10
from vllm.engine.arg_utils import AsyncEngineArgs
11
from vllm.entrypoints.cli.types import CLISubcommand
12
from vllm.entrypoints.openai.api_server import (
@@ -108,8 +110,6 @@ def cmd_init() -> list[CLISubcommand]:
110
111
async def run_launch_fastapi(args: argparse.Namespace) -> None:
112
"""Run the online serving layer with FastAPI (no GPU inference)."""
111
- from vllm.config import VllmConfig
112
-
113
# 1. Socket binding
114
listen_address, sock = setup_server(args)
115
@@ -121,6 +121,10 @@ async def run_launch_fastapi(args: argparse.Namespace) -> None:
121
# Clear quantization so VllmConfig skips quant dtype/capability validation.
122
model_config.quantization = None
123
124
+ # Render servers never allocate KV cache; suppress the spurious CPU KV
125
+ # cache space warning from CpuPlatform.check_and_update_config.
126
+ envs.VLLM_CPU_KVCACHE_SPACE = 0
127
+
128
vllm_config = VllmConfig(model_config=model_config)
129
shutdown_task = await build_and_serve_renderer(
130
vllm_config, listen_address, sock, args