.

Dockerfile

@@ -82,6 +82,12 @@ RUN pip install . && \
 WORKDIR /opt
 RUN pip download --only-binary=:all: --no-deps vllm==0.9.1 -d /tmp/vllm_wheels
 
+# 编译安装compressed_tensors
+RUN git clone https://github.com/vllm-project/compressed-tensors.git && \
+    cd compressed-tensors && \
+    pip install . && \
+    python3 -m pip wheel . --no-deps -w /tmp/compressed_tensors_wheels
+
 # ── 编译你本地 sglang 源码并打 wheel ───────────────────────────────────────
 COPY ./sglang /sgl/sglang
 WORKDIR /sgl/sglang/python
@@ -96,13 +102,14 @@ RUN mkdir -p /wheels && \
     cp /opt/flashinfer/dist/flashinfer_python-*.whl /wheels/ && \
     cp /tmp/vllm_wheels/vllm-*.whl /wheels/ && \
     cp /tmp/sg_wheels/sglang-*.whl /wheels/ && \
+    cp /tmp/compressed_tensors/compressed_tensors*.whl /wheels/ && \
     pip wheel filelock typing-extensions sympy fsspec jinja2 networkx -w /wheels
 
 # ── ✅ 再打包 runtime 阶段必需依赖 ────────────────────────────────────────────
 RUN pip wheel \
     pydantic orjson psutil pyzmq pynvml \
     transformers==4.52.0 uvicorn fastapi IPython aiohttp \
-    setproctitle uvloop sentencepiece triton pillow cachetools msgspec blake3 cloudpickle compressed-tensors \
+    setproctitle uvloop sentencepiece triton pillow cachetools msgspec blake3 cloudpickle \
     -w /wheels
 
 ###############################################################################