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 13, 08:15 UTC
[CI/Build] Install uv for AMD MI300: Language Models Tests (Hybrid) %N (#28142)
Signed-off-by: amdfaa <107946068+amdfaa@users.noreply.github.com> Signed-off-by: zhewenli <zhewenli@meta.com> Co-authored-by: zhewenli <zhewenli@meta.com>
amdfaa committed
Nov 13, 2025 at 09:34 UTC
a7791eac9d29a4a26b007db42130a9e28b3e77ee
1 file changed
+27
-8
docker/Dockerfile.rocm
+27
-8
@@ -15,6 +15,20 @@ RUN apt-get update -q -y && apt-get install -q -y \
15
# Remove sccache
16
RUN python3 -m pip install --upgrade pip
17
RUN apt-get purge -y sccache; python3 -m pip uninstall -y sccache; rm -f "$(which sccache)"
18
+
19
+# Install UV
20
+RUN curl -LsSf https://astral.sh/uv/install.sh | sh
21
+
22
+# Activate virtual environment and add uv to PATH
23
+ENV PATH="/root/.local/bin:$PATH"
24
+
25
+# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
26
+# Reference: https://github.com/astral-sh/uv/pull/1694
27
+ENV UV_HTTP_TIMEOUT=500
28
+ENV UV_INDEX_STRATEGY="unsafe-best-match"
29
+# Use copy mode to avoid hardlink failures with Docker cache mounts
30
+ENV UV_LINK_MODE=copy
31
+
32
ARG COMMON_WORKDIR
33
WORKDIR ${COMMON_WORKDIR}
34
@@ -59,13 +73,15 @@ FROM base AS test
73
74
RUN python3 -m pip install --upgrade pip && rm -rf /var/lib/apt/lists/*
75
62
-# Install vLLM
76
+# Install vLLM using uv (inherited from base stage)
77
+# Note: No -U flag to avoid upgrading PyTorch ROCm to CUDA version
78
RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
79
+ --mount=type=cache,target=/root/.cache/uv \
80
cd /install \
65
- && pip install -U -r requirements/rocm.txt \
66
- && pip install -U -r requirements/rocm-test.txt \
81
+ && uv pip install --system -r requirements/rocm.txt \
82
+ && uv pip install --system -r requirements/rocm-test.txt \
83
&& pip uninstall -y vllm \
68
- && pip install *.whl
84
+ && uv pip install --system *.whl
85
86
WORKDIR /vllm-workspace
87
ARG COMMON_WORKDIR
@@ -89,14 +105,17 @@ RUN case "$(which python3)" in \
105
rm -rf /opt/conda/envs/py_3.9/lib/python3.9/site-packages/numpy-1.20.3.dist-info/;; \
106
*) ;; esac
107
92
-RUN python3 -m pip install --upgrade huggingface-hub[cli]
108
+RUN --mount=type=cache,target=/root/.cache/uv \
109
+ uv pip install --system --upgrade huggingface-hub[cli]
110
94
-# Install vLLM
111
+# Install vLLM using uv (inherited from base stage)
112
+# Note: No -U flag to avoid upgrading PyTorch ROCm to CUDA version
113
RUN --mount=type=bind,from=export_vllm,src=/,target=/install \
114
+ --mount=type=cache,target=/root/.cache/uv \
115
cd /install \
97
- && pip install -U -r requirements/rocm.txt \
116
+ && uv pip install --system -r requirements/rocm.txt \
117
&& pip uninstall -y vllm \
99
- && pip install *.whl
118
+ && uv pip install --system *.whl
119
120
ARG COMMON_WORKDIR
121