feat: upgrade vllm from 0.29.0 to 0.30.0 for CUDA and ROCm backends

README.md

@@ -56,8 +56,8 @@ The following table lists the supported accelerated backends and their correspon
 
 | CUDA Version <br/> (Variant) | vLLM                                                                                                                                                                                                              | SGLang                                                                                                            | VoxBox   |
 |------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------|----------|
-| 13.0                         | **`0.29.0`**, **`0.27.1`**,<br/> **`0.25.1`**, **`0.24.0`**,<br/> `0.22.1`, `0.21.0`,<br/> `0.20.2`, `0.19.1`,<br/> `0.18.1`                                                                                      | `0.5.18`, `0.5.15.post1`, <br/>`0.5.14`, `0.5.12.post1`                                                           |          |
-| 12.9                         | **`0.29.0`**, **`0.27.1`**,<br/> **`0.25.1`**, **`0.24.0`**,<br/> `0.22.1`, `0.21.0`,<br/> `0.20.2`, `0.19.1`,<br/> `0.18.1`, `0.17.1`,<br/> `0.16.0`, `0.15.1`,<br/> `0.14.1`, `0.13.0`,<br/> `0.12.0`, `0.11.2` | `0.5.18`, `0.5.15.post1`,<br/> `0.5.14`, `0.5.12.post1`, <br/>`0.5.9`, `0.5.8.post1`, <br/>`0.5.7`, `0.5.6.post2` |          |
+| 13.0                         | **`0.30.0`**, **`0.29.0`**, **`0.27.1`**,<br/> **`0.25.1`**, **`0.24.0`**,<br/> `0.22.1`, `0.21.0`,<br/> `0.20.2`, `0.19.1`,<br/> `0.18.1`                                                                                      | `0.5.18`, `0.5.15.post1`, <br/>`0.5.14`, `0.5.12.post1`                                                           |          |
+| 12.9                         | **`0.30.0`**, **`0.29.0`**, **`0.27.1`**,<br/> **`0.25.1`**, **`0.24.0`**,<br/> `0.22.1`, `0.21.0`,<br/> `0.20.2`, `0.19.1`,<br/> `0.18.1`, `0.17.1`,<br/> `0.16.0`, `0.15.1`,<br/> `0.14.1`, `0.13.0`,<br/> `0.12.0`, `0.11.2` | `0.5.18`, `0.5.15.post1`,<br/> `0.5.14`, `0.5.12.post1`, <br/>`0.5.9`, `0.5.8.post1`, <br/>`0.5.7`, `0.5.6.post2` |          |
 | 12.8                         | `0.17.1`, `0.16.0`, <br/>`0.15.1`, `0.14.1`, <br/>`0.13.0`, `0.12.0`, <br/>`0.11.2`, `0.10.2`                                                                                                                     | `0.5.9`, `0.5.8.post1`, <br/>`0.5.7`, `0.5.6.post2`, <br/>`0.5.5.post3`                                           | `0.0.21` |
 | 12.6                         | `0.15.1`, `0.14.1`, <br/>`0.13.0`, `0.12.0`, <br/>`0.11.2`, `0.10.2`                                                                                                                                              |                                                                                                                   | `0.0.21` |
 
@@ -96,7 +96,7 @@ The following table lists the supported accelerated backends and their correspon
 
 | ROCm Version <br/> (Variant) | vLLM                                                                                                          | SGLang                                                    |
 |------------------------------|---------------------------------------------------------------------------------------------------------------|-----------------------------------------------------------|
-| 7.2                          | **`0.29.0`**, **`0.27.1`**,<br/> **`0.25.1`**, **`0.24.0`**,<br/> `0.22.1`, `0.21.0`,<br/> `0.20.2`, `0.19.1` | `0.5.18`, `0.5.15.post1`, <br/>`0.5.14`, `0.5.12.post1`   |
+| 7.2                          | **`0.30.0`**, **`0.29.0`**, **`0.27.1`**,<br/> **`0.25.1`**, **`0.24.0`**,<br/> `0.22.1`, `0.21.0`,<br/> `0.20.2`, `0.19.1` | `0.5.18`, `0.5.15.post1`, <br/>`0.5.14`, `0.5.12.post1`   |
 | 7.1                          | `0.17.1`                                                                                                      |                                                           |
 | 7.0                          | `0.18.1`, `0.16.0`,<br/> `0.15.1`, `0.14.1`,<br/> `0.13.0`, `0.12.0`,<br/> `0.11.2`                           | `0.5.9`, `0.5.8.post1`, <br/>`0.5.7`, `0.5.6.post2`       |
 | 6.4                          | `0.16.0`, `0.15.1`,<br/> `0.14.1`, `0.13.0`,<br/> `0.12.0`, `0.11.2`,<br/> `0.10.2`                           | `0.5.8.post1`, `0.5.7`, <br/>`0.5.6.post2`, `0.5.5.post3` |

pack/cuda/Dockerfile.sglang

@@ -9,7 +9,7 @@ ARG SGLANG_TORCH_CUDA_VERSION=${CUDA_VERSION}
 ## Keep >= 0.4.6 and in sync with the ROCm and vLLM pins: SGLang's lmc_radix_cache.py imports
 ## `lmcache.integration.sglang.multi_process_adapter`, added in 0.4.6, and the MP wire protocol
 ## has no version handshake. The SGLang base ships no lmcache, so it is always built from source.
-ARG SGLANG_LMCACHE_VERSION=0.5.4
+ARG SGLANG_LMCACHE_VERSION=0.5.5
 ARG SGLANG_NVIDIA_HPCX_VERSION=2.24.1_cuda13
 ARG SGLANG_AWS_EFA_VERSION=1.46.0
 
@@ -192,6 +192,10 @@ RUN <<EOF
     ## Note: don't pin cuda-bindings —— cu129 base has 12.9.4, cu130 base has 13.2.0,
     ## both compatible with their respective torch. Hardcoding 12.9.4 on cu130 downgrades
     ## and breaks `torch -> cuda-bindings<14,>=13.0.3` requirement.
+    ## 0.5.5 generates its gRPC stubs at build time (setup.py _BuildPyWithGrpcStubs);
+    ## provide the pinned codegen tools its build-system requirements name, since
+    ## `--no-isolation` skips installing them.
+    uv pip install grpcio==1.84.0 grpcio-tools==1.84.0
     pushd /tmp/lmcache \
         && sed -i "s/\"torch==.*\"/\"torch\"/g" /tmp/lmcache/pyproject.toml \
         && cat /tmp/lmcache/pyproject.toml \
@@ -373,7 +377,7 @@ RUN <<EOF
     # both share the same cupy/ tree, so uninstalling one breaks the other. Purge both, then
     # reinstall only the variant matching the CUDA major we build for, at the version already
     # present in the image. Since 0.4.6 lmcache splits its requirements per CUDA major, so our
-    # 0.5.4 pin agrees with this reinstall (see LMCACHE_CUDA_MAJOR in the build stage above),
+    # 0.5.5 pin agrees with this reinstall (see LMCACHE_CUDA_MAJOR in the build stage above),
     # leaving MSCCL++ as the only source that has to be corrected here.
 
     IFS="." read -r CUDA_MAJOR CUDA_MINOR CUDA_PATCH <<< "${CUDA_VERSION}"

pack/cuda/Dockerfile.vllm

@@ -2,17 +2,22 @@ ARG PYTHON_VERSION=3.12
 ARG CMAKE_MAX_JOBS
 ARG CUDA_VERSION=13.0.1
 ARG CUDA_ARCHS
-ARG VLLM_BASE_IMAGE=vllm/vllm-openai:v0.29.0-ubuntu2404
-ARG VLLM_VERSION=0.29.0
-ARG VLLM_TORCH_VERSION=2.13.0
+ARG VLLM_BASE_IMAGE=vllm/vllm-openai:v0.30.0-ubuntu2404
+ARG VLLM_VERSION=0.30.0
+ARG VLLM_TORCH_VERSION=2.14.0
+## torchvision released in lockstep with torch (0.29.0 pairs with 2.14.0); torchaudio
+## has no 2.12.0 release, so the base's 2.11.0 stays on both cu129 and cu130.
+ARG VLLM_TORCHVISION_VERSION=0.29.0
 ARG VLLM_TORCH_CUDA_VERSION=${CUDA_VERSION}
 ## Built from source, replacing the base's own lmcache: that one is a PyPI wheel compiled
 ## against lmcache's build-time torch pin, which need not match the torch this base ships, and
 ## on a mismatch lmcache silently degrades to its torch baseline (vllm-project/vllm#53424).
 ## Keep in sync with the ROCm and SGLang pins —— the MP wire protocol has no version handshake.
-ARG VLLM_LMCACHE_VERSION=0.5.4
+ARG VLLM_LMCACHE_VERSION=0.5.5
 ARG VLLM_ROUTER_VERSION=0.1.15
-ARG VLLM_OMNI_COMMIT=aff7d64948f6
+## A tag, not a commit: v0.30.0rc1 is the omni release rebased onto vLLM 0.30.0
+## (the previous commit pinned a 0.29.0-era tree).
+ARG VLLM_OMNI_COMMIT=v0.30.0rc1
 
 FROM ${VLLM_BASE_IMAGE} AS vllm-build
 SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
@@ -121,12 +126,56 @@ ENV UV_SYSTEM_PYTHON=1 \
 
 ARG VLLM_VERSION
 ARG VLLM_TORCH_VERSION
+ARG VLLM_TORCHVISION_VERSION
 ARG VLLM_TORCH_CUDA_VERSION
 
 ENV VLLM_VERSION=${VLLM_VERSION} \
     VLLM_TORCH_VERSION=${VLLM_TORCH_VERSION} \
     VLLM_TORCH_CUDA_VERSION=${VLLM_TORCH_CUDA_VERSION}
 
+## Upgrade Torch
+#
+# The base image ships the torch it was built against (2.13.0 for the v0.30.0
+# bases). When a rule targets a newer torch, install it from the matching
+# PyTorch wheel index; the base's prebuilt wheels (vllm, flashinfer, fa3-fwd)
+# were compiled against the older torch and rely on torch's C++ ABI staying
+# stable within the major version. A rule whose base already ships the target
+# torch (cu129: 2.14.0 is published for cu130 only) passes VLLM_TORCH_VERSION
+# equal to the base's, and the step is a no-op.
+
+ARG VLLM_TORCH_VERSION
+ARG VLLM_TORCHVISION_VERSION
+ARG VLLM_TORCH_CUDA_VERSION
+
+RUN <<EOF
+    # Torch
+
+    INSTALLED_TORCH="$(python -c 'import torch; print(torch.__version__)')"
+    if [[ "${INSTALLED_TORCH}" == "${VLLM_TORCH_VERSION}"* ]]; then
+        echo "Torch ${INSTALLED_TORCH} already matches ${VLLM_TORCH_VERSION}; skipping the upgrade."
+        exit 0
+    fi
+
+    IFS="." read -r CUDA_MAJOR CUDA_MINOR CUDA_PATCH <<< "${VLLM_TORCH_CUDA_VERSION}"
+
+    ## `--extra-index-url`, not `--index-url`: the PyTorch index carries torch/torchvision
+    ## only, while the transitive deps (filelock, sympy, ...) and the nvidia-* wheels
+    ## resolve from PyPI. The `+cu${CUDA_MAJOR}${CUDA_MINOR}` local versions pin the
+    ## index wheels exactly, so the PyPI copy of the same version can never win.
+    uv pip install \
+        "torch==${VLLM_TORCH_VERSION}+cu${CUDA_MAJOR}${CUDA_MINOR}" \
+        "torchvision==${VLLM_TORCHVISION_VERSION}+cu${CUDA_MAJOR}${CUDA_MINOR}" \
+        --extra-index-url "https://download.pytorch.org/whl/cu${CUDA_MAJOR}${CUDA_MINOR}"
+
+    # Review
+    uv pip show torch torchvision
+    python -c 'import torch; print("torch", torch.__version__, "cuda", torch.version.cuda)'
+
+    # Cleanup
+    rm -rf /var/tmp/* \
+        && rm -rf /tmp/*
+EOF
+
 
 FROM vllm-build AS vllm-build-omni
 SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
@@ -260,10 +309,20 @@ RUN <<EOF
     ## and breaks `torch -> cuda-bindings<14,>=13.0.3` requirement.
     ## Note: separate statements, not an `&&` chain —— `bash -e` ignores a failing non-final
     ## command of a list, which would leave /workspace empty and fail later at install time.
+    ## 0.5.5 generates its gRPC stubs at build time (setup.py _BuildPyWithGrpcStubs);
+    ## provide the pinned codegen tools its build-system requirements name, since
+    ## `--no-isolation` skips installing them.
+    uv pip install grpcio==1.84.0 grpcio-tools==1.84.0
     cd /tmp/lmcache
     ## Drop the build-time torch pin: we compile against the base's own torch on purpose.
     sed -i "s/\"torch==.*\"/\"torch\"/g" /tmp/lmcache/pyproject.toml
     cat /tmp/lmcache/pyproject.toml
+    ## torch 2.14's headers require C++20 (#error in torch/all.h), but lmcache 0.5.5
+    ## hardcodes -std=c++17 for its g++ sources (pybind.cpp); nvcc already defaults
+    ## to C++20 under CUDA 13. Harmless on the cu129 base's torch 2.13.
+    sed -i 's/-std=c++17/-std=c++20/g' \
+        /tmp/lmcache/setup_extensions/common_cpp.py \
+        /tmp/lmcache/setup_extensions/build_profiles/cuda.py
     ## `--skip-dependency-check`: the base does not ship torch's full dependency closure.
     python -m build --no-isolation --skip-dependency-check --wheel
     tree -hs /tmp/lmcache/dist
@@ -473,7 +532,7 @@ soundfile
 mistral_common[audio]
 
 # tokenizer extras
-fastokens==0.2.0
+fastokens==0.3.2
 EOT
     uv pip install \
         -r /tmp/requirements.txt

pack/matrix.yaml

@@ -101,8 +101,8 @@ rules:
       - "linux/amd64"
     args:
       - "ROCM_VERSION=7.2.3"
-      - "VLLM_VERSION=0.29.0"
-      - "VLLM_BASE_IMAGE=vllm/vllm-openai-rocm:v0.29.0"
+      - "VLLM_VERSION=0.30.0"
+      - "VLLM_BASE_IMAGE=vllm/vllm-openai-rocm:v0.30.0"
   ## AMD ROCm 7.2.1 - SGLang 0.5.18 (rocm720-mi30x)
   ##
   - backend: "rocm"
@@ -119,15 +119,17 @@ rules:
   # NVIDIA CUDA
   #
 
-  ## NVIDIA CUDA 13.0.1
+  ## NVIDIA CUDA 13.0.3
+  ## (torch is upgraded from the base's 2.13.0 to 2.14.0 —— see the
+  ##  "Upgrade Torch" step in pack/cuda/Dockerfile.vllm)
   ##
   - backend: "cuda"
     services:
       - "vllm"
     args:
-      - "CUDA_VERSION=13.0.1"
-      - "VLLM_VERSION=0.29.0"
-      - "VLLM_BASE_IMAGE=vllm/vllm-openai:v0.29.0-ubuntu2404"
+      - "CUDA_VERSION=13.0.3"
+      - "VLLM_VERSION=0.30.0"
+      - "VLLM_BASE_IMAGE=vllm/vllm-openai:v0.30.0-ubuntu2404"
   ## NVIDIA CUDA 13.0.1 - SGLang 0.5.18
   ##
   - backend: "cuda"
@@ -138,14 +140,18 @@ rules:
       - "SGLANG_VERSION=0.5.18"
       - "SGLANG_BASE_IMAGE=lmsysorg/sglang:v0.5.18-cu130"
   ## NVIDIA CUDA 12.9.1
+  ## (stays on the base's torch 2.13.0: 2.14.0 is published for cu130 only,
+  ##  so VLLM_TORCH_VERSION is pinned to the base's version and the
+  ##  "Upgrade Torch" step in pack/cuda/Dockerfile.vllm is a no-op)
   ##
   - backend: "cuda"
     services:
       - "vllm"
     args:
       - "CUDA_VERSION=12.9.1"
-      - "VLLM_VERSION=0.29.0"
-      - "VLLM_BASE_IMAGE=vllm/vllm-openai:v0.29.0-cu129-ubuntu2404"
+      - "VLLM_VERSION=0.30.0"
+      - "VLLM_BASE_IMAGE=vllm/vllm-openai:v0.30.0-cu129-ubuntu2404"
+      - "VLLM_TORCH_VERSION=2.13.0"
   ## NVIDIA CUDA 12.9.1 - SGLang 0.5.18
   ##
   - backend: "cuda"

pack/rocm/Dockerfile.sglang

@@ -8,7 +8,7 @@ ARG SGLANG_TORCH_VERSION=2.9.1
 ARG SGLANG_TORCH_ROCM_VERSION=${ROCM_VERSION}
 ## Keep >= 0.4.6 and in sync with the CUDA pin —— see pack/cuda/Dockerfile.sglang.
 ## Always built from source: upstream's ROCm wheel targets torch 2.11.0+rocm7.2.
-ARG SGLANG_LMCACHE_VERSION=0.5.4
+ARG SGLANG_LMCACHE_VERSION=0.5.5
 
 FROM ${SGLANG_BASE_IMAGE} AS sglang-build
 SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
@@ -198,6 +198,10 @@ RUN <<EOF
     git -C /tmp clone --recursive --shallow-submodules \
         --depth 1 --branch v${SGLANG_LMCACHE_VERSION} --single-branch \
         https://github.com/LMCache/LMCache.git lmcache
+    ## 0.5.5 generates its gRPC stubs at build time (setup.py _BuildPyWithGrpcStubs);
+    ## provide the pinned codegen tools its build-system requirements name, since
+    ## `--no-isolation` skips installing them.
+    uv pip install grpcio==1.84.0 grpcio-tools==1.84.0
     pushd /tmp/lmcache \
         && sed -i "s/\"torch==.*\"/\"torch\"/g" /tmp/lmcache/pyproject.toml \
         && cat /tmp/lmcache/pyproject.toml \

pack/rocm/Dockerfile.vllm

@@ -2,17 +2,21 @@ ARG PYTHON_VERSION=3.12
 ARG CMAKE_MAX_JOBS
 ARG ROCM_VERSION=7.2.3
 ARG ROCM_ARCHS
-ARG VLLM_BASE_IMAGE=vllm/vllm-openai-rocm:v0.29.0
-ARG VLLM_VERSION=0.29.0
+ARG VLLM_BASE_IMAGE=vllm/vllm-openai-rocm:v0.30.0
+ARG VLLM_VERSION=0.30.0
+## The v0.30.0 ROCm base is built from the ROCm pytorch release/2.12 branch,
+## so it keeps the base's torch; there is no ROCm equivalent of the CUDA
+## "Upgrade Torch" step.
 ARG VLLM_TORCH_VERSION=2.12.0
 ARG VLLM_TORCH_ROCM_VERSION=${ROCM_VERSION}
 ## The ROCm base ships no lmcache of its own, so it is always built from source here.
 ## Keep in sync with the CUDA and SGLang pins —— see pack/cuda/Dockerfile.vllm.
-ARG VLLM_LMCACHE_VERSION=0.5.4
+ARG VLLM_LMCACHE_VERSION=0.5.5
 ## Upstream's own HIP build —— same source and CMake options we used to compile here,
 ## but produced in a clean rocm/dev-ubuntu-22.04 and smoke-tested before release.
 ARG VLLM_MOONCAKE_VERSION=0.3.13.post1
-ARG VLLM_OMNI_COMMIT=aff7d64948f6
+## A tag, not a commit: v0.30.0rc1 is the omni release rebased onto vLLM 0.30.0.
+ARG VLLM_OMNI_COMMIT=v0.30.0rc1
 
 FROM ${VLLM_BASE_IMAGE} AS vllm-build
 SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
@@ -185,6 +189,10 @@ RUN <<EOF
     git -C /tmp clone --recursive --shallow-submodules \
         --depth 1 --branch v${VLLM_LMCACHE_VERSION} --single-branch \
         https://github.com/LMCache/LMCache.git lmcache
+    ## 0.5.5 generates its gRPC stubs at build time (setup.py _BuildPyWithGrpcStubs);
+    ## provide the pinned codegen tools its build-system requirements name, since
+    ## `--no-isolation` skips installing them.
+    uv pip install grpcio==1.84.0 grpcio-tools==1.84.0
     pushd /tmp/lmcache \
         && sed -i "s/\"torch==.*\"/\"torch\"/g" /tmp/lmcache/pyproject.toml \
         && cat /tmp/lmcache/pyproject.toml \
@@ -410,7 +418,7 @@ mistral_common[audio]
 petit-kernel
 
 # tokenizer extras
-fastokens==0.2.0
+fastokens==0.3.2
 EOT
     uv pip install \
         -r /tmp/requirements.txt