mirror of
https://git.datalinker.icu/vllm-project/vllm.git
synced 2026-04-16 22:37:09 +08:00
Merge 86323206761d09de969a6dc1f0e90b8b07e9dd97 into 254f6b986720c92ddf97fbb1a6a6465da8e87e29
This commit is contained in:
Andrey Talman
GitHub
commit
767510811d
@ -196,3 +196,4 @@ steps:
|
||||
env:
|
||||
DOCKER_BUILDKIT: "1"
|
||||
DOCKERHUB_USERNAME: "vllmbot"
|
||||
|
||||
|
||||
@ -54,7 +54,11 @@ ARG PIP_EXTRA_INDEX_URL
|
||||
ARG UV_INDEX_URL=${PIP_INDEX_URL}
|
||||
ARG UV_EXTRA_INDEX_URL=${PIP_EXTRA_INDEX_URL}
|
||||
|
||||
# Flag to install PyTorch nightly instead of stable
|
||||
ARG USE_TORCH_NIGHTLY=false
|
||||
|
||||
# PyTorch provides its own indexes for standard and nightly builds
|
||||
# This will be set conditionally based on USE_TORCH_NIGHTLY in the base stage
|
||||
ARG PYTORCH_CUDA_INDEX_BASE_URL=https://download.pytorch.org/whl
|
||||
|
||||
# PIP supports multiple authentication schemes, including keyring
|
||||
@ -128,15 +132,40 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
|
||||
# Install PyTorch and core CUDA dependencies
|
||||
# This is ~2GB and rarely changes
|
||||
ARG PYTORCH_CUDA_INDEX_BASE_URL
|
||||
ARG USE_TORCH_NIGHTLY
|
||||
|
||||
# Set PyTorch index URL based on USE_TORCH_NIGHTLY
|
||||
# We compute the index URL once and reuse it across all stages
|
||||
RUN if [ "$USE_TORCH_NIGHTLY" = "true" ]; then \
|
||||
PYTORCH_SUFFIX="/nightly"; \
|
||||
else \
|
||||
PYTORCH_SUFFIX=""; \
|
||||
fi && \
|
||||
echo "PYTORCH_INDEX=${PYTORCH_CUDA_INDEX_BASE_URL}${PYTORCH_SUFFIX}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')" >> /etc/environment && \
|
||||
if [ "$USE_TORCH_NIGHTLY" = "true" ]; then \
|
||||
echo "PRERELEASE_FLAG=--prerelease=allow" >> /etc/environment; \
|
||||
else \
|
||||
echo "PRERELEASE_FLAG=" >> /etc/environment; \
|
||||
fi
|
||||
|
||||
ENV PYTORCH_INDEX=""
|
||||
ENV PRERELEASE_FLAG=""
|
||||
RUN . /etc/environment && echo "PYTORCH_INDEX=${PYTORCH_INDEX}" && echo "PRERELEASE_FLAG=${PRERELEASE_FLAG}"
|
||||
|
||||
WORKDIR /workspace
|
||||
|
||||
# install build and runtime dependencies
|
||||
COPY requirements/common.txt requirements/common.txt
|
||||
COPY requirements/cuda.txt requirements/cuda.txt
|
||||
COPY use_existing_torch.py use_existing_torch.py
|
||||
RUN --mount=type=cache,target=/root/.cache/uv \
|
||||
. /etc/environment && \
|
||||
if [ "$USE_TORCH_NIGHTLY" = "true" ]; then \
|
||||
echo ">>> Running use_existing_torch.py to reset torch dependencies for nightly build" && \
|
||||
python3 use_existing_torch.py; \
|
||||
fi && \
|
||||
uv pip install --python /opt/venv/bin/python3 -r requirements/cuda.txt \
|
||||
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
|
||||
--extra-index-url ${PYTORCH_INDEX} ${PRERELEASE_FLAG}
|
||||
|
||||
# CUDA arch list used by torch
|
||||
# Explicitly set the list to avoid issues with torch 2.2
|
||||
@ -151,10 +180,10 @@ ARG TARGETPLATFORM
|
||||
|
||||
ARG PIP_INDEX_URL UV_INDEX_URL
|
||||
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
|
||||
ARG PYTORCH_CUDA_INDEX_BASE_URL
|
||||
|
||||
# install build dependencies
|
||||
COPY requirements/build.txt requirements/build.txt
|
||||
COPY use_existing_torch.py use_existing_torch.py
|
||||
|
||||
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
|
||||
# Reference: https://github.com/astral-sh/uv/pull/1694
|
||||
@ -164,8 +193,13 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
|
||||
ENV UV_LINK_MODE=copy
|
||||
|
||||
RUN --mount=type=cache,target=/root/.cache/uv \
|
||||
. /etc/environment && \
|
||||
if [ "$USE_TORCH_NIGHTLY" = "true" ]; then \
|
||||
echo ">>> Running use_existing_torch.py to reset torch dependencies for nightly build" && \
|
||||
python3 use_existing_torch.py; \
|
||||
fi && \
|
||||
uv pip install --python /opt/venv/bin/python3 -r requirements/build.txt \
|
||||
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
|
||||
--extra-index-url ${PYTORCH_INDEX} ${PRERELEASE_FLAG}
|
||||
|
||||
WORKDIR /workspace
|
||||
|
||||
@ -290,10 +324,10 @@ ARG TARGETPLATFORM
|
||||
|
||||
ARG PIP_INDEX_URL UV_INDEX_URL
|
||||
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
|
||||
ARG PYTORCH_CUDA_INDEX_BASE_URL
|
||||
|
||||
# install build dependencies
|
||||
COPY requirements/build.txt requirements/build.txt
|
||||
COPY use_existing_torch.py use_existing_torch.py
|
||||
|
||||
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
|
||||
# Reference: https://github.com/astral-sh/uv/pull/1694
|
||||
@ -303,8 +337,13 @@ ENV UV_INDEX_STRATEGY="unsafe-best-match"
|
||||
ENV UV_LINK_MODE=copy
|
||||
|
||||
RUN --mount=type=cache,target=/root/.cache/uv \
|
||||
. /etc/environment && \
|
||||
if [ "$USE_TORCH_NIGHTLY" = "true" ]; then \
|
||||
echo ">>> Running use_existing_torch.py to reset torch dependencies for nightly build" && \
|
||||
python3 use_existing_torch.py; \
|
||||
fi && \
|
||||
uv pip install --python /opt/venv/bin/python3 -r requirements/build.txt \
|
||||
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
|
||||
--extra-index-url ${PYTORCH_INDEX} ${PRERELEASE_FLAG}
|
||||
|
||||
WORKDIR /workspace
|
||||
|
||||
@ -352,7 +391,6 @@ FROM base AS dev
|
||||
|
||||
ARG PIP_INDEX_URL UV_INDEX_URL
|
||||
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
|
||||
ARG PYTORCH_CUDA_INDEX_BASE_URL
|
||||
|
||||
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
|
||||
# Reference: https://github.com/astral-sh/uv/pull/1694
|
||||
@ -367,8 +405,9 @@ COPY requirements/lint.txt requirements/lint.txt
|
||||
COPY requirements/test.txt requirements/test.txt
|
||||
COPY requirements/dev.txt requirements/dev.txt
|
||||
RUN --mount=type=cache,target=/root/.cache/uv \
|
||||
. /etc/environment && \
|
||||
uv pip install --python /opt/venv/bin/python3 -r requirements/dev.txt \
|
||||
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
|
||||
--extra-index-url ${PYTORCH_INDEX} ${PRERELEASE_FLAG}
|
||||
#################### DEV IMAGE ####################
|
||||
#################### vLLM installation IMAGE ####################
|
||||
# image with vLLM installed
|
||||
@ -462,12 +501,37 @@ RUN ldconfig /usr/local/cuda-$(echo $CUDA_VERSION | cut -d. -f1,2)/compat/
|
||||
# Install PyTorch and core CUDA dependencies
|
||||
# This is ~2GB and rarely changes
|
||||
ARG PYTORCH_CUDA_INDEX_BASE_URL
|
||||
ARG USE_TORCH_NIGHTLY
|
||||
|
||||
# Set PyTorch index URL based on USE_TORCH_NIGHTLY
|
||||
# We compute the index URL once and reuse it across all stages
|
||||
RUN if [ "$USE_TORCH_NIGHTLY" = "true" ]; then \
|
||||
PYTORCH_SUFFIX="/nightly"; \
|
||||
else \
|
||||
PYTORCH_SUFFIX=""; \
|
||||
fi && \
|
||||
echo "PYTORCH_INDEX=${PYTORCH_CUDA_INDEX_BASE_URL}${PYTORCH_SUFFIX}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')" >> /etc/environment && \
|
||||
if [ "$USE_TORCH_NIGHTLY" = "true" ]; then \
|
||||
echo "PRERELEASE_FLAG=--prerelease=allow" >> /etc/environment; \
|
||||
else \
|
||||
echo "PRERELEASE_FLAG=" >> /etc/environment; \
|
||||
fi
|
||||
|
||||
ENV PYTORCH_INDEX=""
|
||||
ENV PRERELEASE_FLAG=""
|
||||
|
||||
COPY requirements/common.txt /tmp/common.txt
|
||||
COPY requirements/cuda.txt /tmp/requirements-cuda.txt
|
||||
COPY use_existing_torch.py /tmp/use_existing_torch.py
|
||||
RUN --mount=type=cache,target=/root/.cache/uv \
|
||||
. /etc/environment && \
|
||||
if [ "$USE_TORCH_NIGHTLY" = "true" ]; then \
|
||||
echo ">>> Running use_existing_torch.py to reset torch dependencies for nightly build" && \
|
||||
cd /tmp && python3 use_existing_torch.py; \
|
||||
fi && \
|
||||
uv pip install --system -r /tmp/requirements-cuda.txt \
|
||||
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.') && \
|
||||
rm /tmp/requirements-cuda.txt /tmp/common.txt
|
||||
--extra-index-url ${PYTORCH_INDEX} ${PRERELEASE_FLAG} && \
|
||||
rm /tmp/requirements-cuda.txt /tmp/common.txt /tmp/use_existing_torch.py
|
||||
|
||||
# Install FlashInfer pre-compiled kernel cache and binaries
|
||||
# This is ~1.1GB and only changes when FlashInfer version bumps
|
||||
@ -518,12 +582,14 @@ ARG PIP_INDEX_URL UV_INDEX_URL
|
||||
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
|
||||
ARG PYTORCH_CUDA_INDEX_BASE_URL
|
||||
ARG PIP_KEYRING_PROVIDER UV_KEYRING_PROVIDER
|
||||
ARG USE_TORCH_NIGHTLY
|
||||
|
||||
# Install vllm wheel first, so that torch etc will be installed.
|
||||
RUN --mount=type=bind,from=build,src=/workspace/dist,target=/vllm-workspace/dist \
|
||||
--mount=type=cache,target=/root/.cache/uv \
|
||||
. /etc/environment && \
|
||||
uv pip install --system dist/*.whl --verbose \
|
||||
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
|
||||
--extra-index-url ${PYTORCH_INDEX} ${PRERELEASE_FLAG}
|
||||
|
||||
RUN --mount=type=cache,target=/root/.cache/uv \
|
||||
. /etc/environment && \
|
||||
@ -544,8 +610,9 @@ ENV LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATH
|
||||
# Install EP kernels wheels (pplx-kernels and DeepEP) that have been built in the `build` stage
|
||||
RUN --mount=type=bind,from=build,src=/tmp/ep_kernels_workspace/dist,target=/vllm-workspace/ep_kernels/dist \
|
||||
--mount=type=cache,target=/root/.cache/uv \
|
||||
. /etc/environment && \
|
||||
uv pip install --system ep_kernels/dist/*.whl --verbose \
|
||||
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
|
||||
--extra-index-url ${PYTORCH_INDEX} ${PRERELEASE_FLAG}
|
||||
|
||||
# CUDA image changed from /usr/local/nvidia to /usr/local/cuda in 12.8 but will
|
||||
# return to /usr/local/nvidia in 13.0 to allow container providers to mount drivers
|
||||
@ -569,7 +636,6 @@ ARG PYTHON_VERSION
|
||||
|
||||
ARG PIP_INDEX_URL UV_INDEX_URL
|
||||
ARG PIP_EXTRA_INDEX_URL UV_EXTRA_INDEX_URL
|
||||
ARG PYTORCH_CUDA_INDEX_BASE_URL
|
||||
|
||||
# This timeout (in seconds) is necessary when installing some dependencies via uv since it's likely to time out
|
||||
# Reference: https://github.com/astral-sh/uv/pull/1694
|
||||
@ -587,8 +653,9 @@ RUN echo 'tzdata tzdata/Areas select America' | debconf-set-selections \
|
||||
RUN --mount=type=cache,target=/root/.cache/uv \
|
||||
CUDA_MAJOR="${CUDA_VERSION%%.*}"; \
|
||||
if [ "$CUDA_MAJOR" -ge 12 ]; then \
|
||||
. /etc/environment && \
|
||||
uv pip install --system -r requirements/dev.txt \
|
||||
--extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.'); \
|
||||
--extra-index-url ${PYTORCH_INDEX} ${PRERELEASE_FLAG}; \
|
||||
fi
|
||||
|
||||
# install development dependencies (for testing)
|
||||
|
||||
@ -2,17 +2,85 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import glob
|
||||
import os
|
||||
import re
|
||||
|
||||
# Collect all files to process
|
||||
files_to_process = list(glob.glob("requirements/*.txt"))
|
||||
|
||||
# Add pyproject.toml if it exists
|
||||
if os.path.exists("pyproject.toml"):
|
||||
files_to_process.append("pyproject.toml")
|
||||
|
||||
# Pattern to match torch package names we want to unpin
|
||||
TORCH_PACKAGES = ['torch', 'torchaudio', 'torchvision', 'triton']
|
||||
|
||||
def unpin_torch_dependency(line):
|
||||
"""Remove version pinning from torch-related packages, keep the package name."""
|
||||
original_line = line
|
||||
line_stripped = line.strip()
|
||||
|
||||
# Skip empty lines
|
||||
if not line_stripped:
|
||||
return line
|
||||
|
||||
# Skip full comment lines
|
||||
if line_stripped.startswith('#'):
|
||||
return line
|
||||
|
||||
# Check if this line contains a torch package
|
||||
for pkg in TORCH_PACKAGES:
|
||||
# Check if line starts with the package name (case insensitive)
|
||||
if line_stripped.lower().startswith(pkg):
|
||||
# Extract inline comment if present
|
||||
comment = ''
|
||||
if '#' in line:
|
||||
pkg_and_version, comment = line.split('#', 1)
|
||||
comment = ' #' + comment.rstrip('\n')
|
||||
else:
|
||||
pkg_and_version = line
|
||||
|
||||
# Check if there's a version specifier
|
||||
# Matches any version constraint operators: ==, >=, <=, >, <, !=, ~=
|
||||
if re.search(r'[=<>!~]', pkg_and_version):
|
||||
# Get original capitalization of package name from the original line
|
||||
orig_pkg = line_stripped.split()[0] if line_stripped.split() else pkg
|
||||
# Extract just the package name without any version info
|
||||
orig_pkg = re.split(r'[=<>!~]', orig_pkg)[0]
|
||||
|
||||
result = f"{orig_pkg}{comment}\n" if comment else f"{orig_pkg}\n"
|
||||
print(f" unpinned: {line.strip()} -> {result.strip()}")
|
||||
return result
|
||||
|
||||
return line
|
||||
|
||||
for file in files_to_process:
|
||||
if not os.path.exists(file):
|
||||
print(f">>> skipping {file} (does not exist)")
|
||||
continue
|
||||
|
||||
for file in (*glob.glob("requirements/*.txt"), "pyproject.toml"):
|
||||
print(f">>> cleaning {file}")
|
||||
with open(file) as f:
|
||||
lines = f.readlines()
|
||||
if "torch" in "".join(lines).lower():
|
||||
print("removed:")
|
||||
with open(file, "w") as f:
|
||||
for line in lines:
|
||||
if "torch" not in line.lower():
|
||||
f.write(line)
|
||||
else:
|
||||
print(line.strip())
|
||||
try:
|
||||
with open(file) as f:
|
||||
lines = f.readlines()
|
||||
except Exception as e:
|
||||
print(f"!!! error reading {file}: {e}")
|
||||
continue
|
||||
|
||||
# Check if we need to process this file
|
||||
has_torch = any(any(pkg in line.lower() for pkg in TORCH_PACKAGES) for line in lines)
|
||||
|
||||
if has_torch:
|
||||
print("unpinning torch dependencies:")
|
||||
try:
|
||||
with open(file, "w") as f:
|
||||
for line in lines:
|
||||
new_line = unpin_torch_dependency(line)
|
||||
f.write(new_line)
|
||||
except Exception as e:
|
||||
print(f"!!! error writing {file}: {e}")
|
||||
continue
|
||||
else:
|
||||
print(" (no torch dependencies found)")
|
||||
|
||||
print(f"<<< done cleaning {file}\n")
|
||||
|
||||
Loading…
x
Reference in New Issue
Block a user