From 1f807347eefa4d52a6e60dc7a2d10b7a94463ad2 Mon Sep 17 00:00:00 2001 From: SZCJW <792430652@qq.com> Date: Wed, 1 Apr 2026 01:08:27 +0800 Subject: [PATCH] x --- app/model_catalog.py | 6 ++++++ config.json | 13 ++++++++----- docker-compose.yml | 3 +++ 3 files changed, 17 insertions(+), 5 deletions(-) diff --git a/app/model_catalog.py b/app/model_catalog.py index eb76872..85f294a 100644 --- a/app/model_catalog.py +++ b/app/model_catalog.py @@ -125,4 +125,10 @@ def resolve_model_profile( env_vars = {str(k): str(v) for k, v in dict(profile.get("env", {})).items()} env_vars["HF_HUB_OFFLINE"] = "1" env_vars["TRANSFORMERS_OFFLINE"] = "1" + # 添加 vLLM RPC 超时环境变量,防止共享内存通信超时 + if "VLLM_RPC_TIMEOUT" not in env_vars: + env_vars["VLLM_RPC_TIMEOUT"] = "300" + # 添加 vLLM 工作进程超时环境变量 + if "VLLM_WORKER_MULTIPROC_METHOD" not in env_vars: + env_vars["VLLM_WORKER_MULTIPROC_METHOD"] = "spawn" return model_key, updates, env_vars diff --git a/config.json b/config.json index cb594d0..315509e 100644 --- a/config.json +++ b/config.json @@ -73,18 +73,21 @@ "local_path": "Qwen3.5-35B-A3B-GPTQ-Int4", "dtype": "float16", "quantization": "gptq", - "ctx": "131072", + "ctx": "65536", "trust_remote": true, "valid_tp": [1, 2], - "max_num_seqs": "8", - "max_tokens": "8192", - "gpu_util": "0.90", + "max_num_seqs": "4", + "max_tokens": "4096", + "gpu_util": "0.85", "enforce_eager": false, "tool_call_parser": "qwen3_coder", "reasoning_parser": "qwen3", "enable_auto_tool_choice": true, "served_model_name": "Qwen3.5-35B-A3B-GPTQ-Int4", - "hf_model_id": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4" + "hf_model_id": "Qwen/Qwen3.5-35B-A3B-GPTQ-Int4", + "env": { + "VLLM_RPC_TIMEOUT": "300" + } }, "Qwen3.5-35B-A3B-FP8": { "local_path": "Qwen3.5-35B-A3B-FP8", diff --git a/docker-compose.yml b/docker-compose.yml index 700aac8..892c4c1 100644 --- a/docker-compose.yml +++ b/docker-compose.yml @@ -14,6 +14,9 @@ services: - ./config.json:/workspace/config.json:ro environment: HF_HOME: /opt/model + VLLM_RPC_TIMEOUT: "300" + VLLM_WORKER_MULTIPROC_METHOD: "spawn" + PYTHONUNBUFFERED: "1" devices: - /dev/kfd - /dev/dri