From e45b3abbdaf583bb4b12287251ccfa544db13ee6 Mon Sep 17 00:00:00 2001 From: Zachary Aristei Date: Sun, 20 Jul 2025 16:48:54 -0700 Subject: [PATCH 1/4] get correct nightlies via seeing what prints on dry run pip install on b200 --- container/deps/vllm/install_vllm.sh | 7 ++++--- 1 file changed, 4 insertions(+), 3 deletions(-) diff --git a/container/deps/vllm/install_vllm.sh b/container/deps/vllm/install_vllm.sh index ee48487a69..1a585ca772 100755 --- a/container/deps/vllm/install_vllm.sh +++ b/container/deps/vllm/install_vllm.sh @@ -119,9 +119,10 @@ if [ "$ARCH" = "arm64" ]; then # Try to install specific PyTorch version first, fallback to latest nightly echo "Attempting to install pinned PyTorch nightly versions..." - if ! uv pip install torch==2.9.0.dev20250712+cu128 torchvision==0.24.0.dev20250712+cu128 torchaudio==2.8.0.dev20250712+cu128 --index-url https://download.pytorch.org/whl/nightly/cu128; then - echo "Pinned versions failed, falling back to latest stable..." - uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128 + if ! uv pip install torch==2.8.0.dev20250613+cu128 torchaudio==2.8.0.dev20250616 torchvision==0.23.0.dev20250616 --index-url https://download.pytorch.org/whl/nightly/cu128; then + echo "Pinned versions failed" + exit 1 + # uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128 fi python use_existing_torch.py From a431bb7f725533e081d85702e3bd683cf7b52cc8 Mon Sep 17 00:00:00 2001 From: Zachary Aristei Date: Tue, 22 Jul 2025 14:24:34 -0700 Subject: [PATCH 2/4] fix nixl build args again --- container/Dockerfile.vllm | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/container/Dockerfile.vllm b/container/Dockerfile.vllm index f42124b6ee..1a2469a76d 100644 --- a/container/Dockerfile.vllm +++ b/container/Dockerfile.vllm @@ -171,7 +171,8 @@ ENV PATH="${VIRTUAL_ENV}/bin:${PATH}" # TEMP: disable gds backend for arm64 RUN if [ "$ARCH" = "arm64" ]; then \ cd ${NIXL_SRC_DIR} && uv build . --out-dir /workspace/wheels/nixl \ - --config-settings=setup-args="-Ddisable_gds_backend=true -Dgds_path=/usr/local/cuda/targets/sbsa-linux"; \ + --config-settings=setup-args="-Ddisable_gds_backend=true" \ + --config-settings=setup-args="-Dgds_path=/usr/local/cuda/targets/sbsa-linux"; \ else \ cd ${NIXL_SRC_DIR} && uv build . --out-dir /workspace/wheels/nixl; \ fi && \ From 2b9aec5b58069f2e9a905872b8161906caefdf9c Mon Sep 17 00:00:00 2001 From: Zachary Aristei Date: Tue, 22 Jul 2025 15:57:39 -0700 Subject: [PATCH 3/4] rm flashinfer for now --- container/deps/vllm/install_vllm.sh | 10 +++++----- 1 file changed, 5 insertions(+), 5 deletions(-) diff --git a/container/deps/vllm/install_vllm.sh b/container/deps/vllm/install_vllm.sh index 1a585ca772..a99e4d5365 100755 --- a/container/deps/vllm/install_vllm.sh +++ b/container/deps/vllm/install_vllm.sh @@ -164,10 +164,10 @@ python setup.py install # Install Flash Infer -cd $INSTALLATION_DIR -git clone https://github.com/flashinfer-ai/flashinfer.git --recursive -cd flashinfer -git checkout $FLASHINF_REF -python -m pip install -v . +#cd $INSTALLATION_DIR +#git clone https://github.com/flashinfer-ai/flashinfer.git --recursive +#cd flashinfer +#git checkout $FLASHINF_REF +#python -m pip install -v . echo "vllm installation completed successfully" From 3e0f3815352da8f588f6fb572f84b860c7c464fa Mon Sep 17 00:00:00 2001 From: Zachary Aristei Date: Tue, 22 Jul 2025 17:04:39 -0700 Subject: [PATCH 4/4] install stable flashinfer for now --- container/deps/vllm/install_vllm.sh | 14 +++++++++----- 1 file changed, 9 insertions(+), 5 deletions(-) diff --git a/container/deps/vllm/install_vllm.sh b/container/deps/vllm/install_vllm.sh index a99e4d5365..54384e18d1 100755 --- a/container/deps/vllm/install_vllm.sh +++ b/container/deps/vllm/install_vllm.sh @@ -164,10 +164,14 @@ python setup.py install # Install Flash Infer -#cd $INSTALLATION_DIR -#git clone https://github.com/flashinfer-ai/flashinfer.git --recursive -#cd flashinfer -#git checkout $FLASHINF_REF -#python -m pip install -v . +if [ "$ARCH" = "arm64" ]; then + uv pip install flashinfer-python +else + cd $INSTALLATION_DIR + git clone https://github.com/flashinfer-ai/flashinfer.git --recursive + cd flashinfer + git checkout $FLASHINF_REF + python -m pip install -v . +fi echo "vllm installation completed successfully"