x

Dockerfile

@@ -3,7 +3,7 @@ FROM rocm/vllm:rocm7.12.0_gfx120X-all_ubuntu24.04_py3.12_pytorch_2.9.1_vllm_0.16
 WORKDIR /workspace
 
 COPY requirements.txt /workspace/requirements.txt
-RUN pip install --no-cache-dir -r /workspace/requirements.txt
+RUN pip install --no-cache-dir --retries 10 --timeout 180 -r /workspace/requirements.txt
 
 COPY app /workspace/app
 

README.md

@@ -155,4 +155,6 @@ docker compose up -d --force-recreate
 
 - 如果看到 `No services to build`，说明你执行的是 `docker compose up -d --force-recreate` 且未触发重建；必须先单独执行 `docker compose build --no-cache`。
 
+- 构建阶段若出现 `git clone https://github.com/huggingface/transformers.git` 失败，请改用 PyPI 版本依赖（项目已默认 `transformers>=4.57.0,<5`），避免源码拉取中断导致构建失败。
+
 - 若仍报相同错误，说明当前 `vLLM/Transformers` 组合不支持该模型架构，建议在 `config.json` 切到其他本地模型，或升级基础镜像到更新的 vLLM 版本。

config.json

@@ -17,7 +17,7 @@
   "revision": "",
   "models": {
     "default": "Qwen3.5-35B-A3B-GPTQ-Int4",
-    "selected": "Qwen3.5-35B-A3B-GPTQ-Int4",
+    "selected": "Qwen3.5-27B-FP8",
     "profiles": {
       "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit": {
         "local_path": "Qwen3-Next-80B-A3B-Instruct-AWQ-4bit",
@@ -51,7 +51,7 @@
       },
       "Qwen3.5-27B-FP8": {
         "local_path": "Qwen3.5-27B-FP8",
-        "ctx": "32768",
+        "ctx": "65536",
         "trust_remote": true,
         "valid_tp": [1, 2],
         "max_num_seqs": "64",
@@ -64,7 +64,7 @@
       },
       "Qwen3.5-35B-A3B-GPTQ-Int4": {
         "local_path": "Qwen3.5-35B-A3B-GPTQ-Int4",
-        "ctx": "32768",
+        "ctx": "65536",
         "trust_remote": true,
         "valid_tp": [1, 2],
         "max_num_seqs": "64",

requirements.txt

@@ -1,4 +1,4 @@
 fastapi==0.116.1
 uvicorn==0.35.0
 pydantic==2.11.7
-transformers @ git+https://github.com/huggingface/transformers.git
+transformers>=4.57.0,<5