Upgrade to cu130

builder/manywheel/README.md

@@ -3,7 +3,8 @@
 ## Building lmdeploy builder images
 
 To build all lmdeploy builder images, such as
-`lmdeploy-builder:manylinux_2_28-cuda12.8`, execute:
+`lmdeploy-builder:manylinux_2_28-cuda12.8` and
+`lmdeploy-builder:manylinux_2_28-cuda13.0`, execute:
 
 ```bash
 ./build_all_lmdeploy_builders.sh
@@ -15,7 +16,7 @@ WITH_PUSH=true ./build_all_lmdeploy_builders.sh
 For custom builds with specific versions:
 
 ```bash
-MANY_LINUX_VERSION=2_28 GPU_ARCH_VERSION=12.8 ./build_lmdeploy_builder.sh
+MANY_LINUX_VERSION=2_28 GPU_ARCH_VERSION=13.0 ./build_lmdeploy_builder.sh
 ```
 
 ## Build lmdeploy wheels

docker/Dockerfile

@@ -1,15 +1,20 @@
 # Base images
 ARG IMAGE_TYPE=final
-ARG CUDA_VERSION=cu12
+ARG CUDA_VERSION=cu128
 
-FROM nvidia/cuda:13.0.2-devel-ubuntu22.04 AS cu13
+FROM nvidia/cuda:13.0.3-devel-ubuntu24.04 AS cu130
 ENV CUDA_VERSION_SHORT=cu130
+ENV PYTORCH_CUDA_VERSION=cu130
 
-FROM nvidia/cuda:12.8.1-devel-ubuntu22.04 AS cu12.8
+FROM nvidia/cuda:12.8.1-devel-ubuntu24.04 AS cu128
 ENV CUDA_VERSION_SHORT=cu128
+# PyTorch 2.12 no longer publishes cu128 wheels; cu126 is its CUDA 12 build.
+ENV PYTORCH_CUDA_VERSION=cu126
 
-FROM nvidia/cuda:12.6.3-devel-ubuntu22.04 AS cu12
-ENV CUDA_VERSION_SHORT=cu126
+# Backward-compatible stage aliases.
+FROM cu130 AS cu13
+FROM cu128 AS cu12.8
+FROM cu128 AS cu12
 
 # Builder image
 FROM ${CUDA_VERSION} AS dev
@@ -20,7 +25,13 @@ ENV DEBIAN_FRONTEND=noninteractive
 ENV TZ=Etc/UTC
 
 RUN --mount=type=cache,target=/root/.cache \
-    sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' /etc/apt/sources.list && \
+    if [ -f /etc/apt/sources.list ]; then \
+        sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' /etc/apt/sources.list; \
+    fi && \
+    if [ -d /etc/apt/sources.list.d ]; then \
+        find /etc/apt/sources.list.d -type f -name '*.sources' -exec \
+            sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' {} +; \
+    fi && \
     apt-get update -y && \
     apt-get install -y --no-install-recommends \
         tzdata wget curl openssh-server ssh sudo git-core \
@@ -40,7 +51,9 @@ ENV PYTHON_VERSION=${PYTHON_VERSION}
 COPY requirements/common.txt requirements/runtime_cuda.txt /tmp/requirements/
 
 RUN --mount=type=cache,target=/root/.cache \
-    pip install -r /tmp/requirements/runtime_cuda.txt --extra-index-url https://download.pytorch.org/whl/${CUDA_VERSION_SHORT}
+    grep -E '^torch(vision)?([<>=]|$)' /tmp/requirements/runtime_cuda.txt > /tmp/requirements/torch.txt && \
+    pip install -r /tmp/requirements/torch.txt --index-url https://download.pytorch.org/whl/${PYTORCH_CUDA_VERSION} && \
+    pip install -r /tmp/requirements/runtime_cuda.txt
 
 RUN mkdir -p /wheels
 
@@ -67,24 +80,30 @@ COPY --from=thirdparty-wheels /wheels /wheels
 COPY --from=lmdeploy-wheel /wheels /wheels
 
 # Runtime image
-FROM nvidia/cuda:13.0.2-base-ubuntu22.04 AS cu13-base
+FROM nvidia/cuda:13.0.3-base-ubuntu24.04 AS cu130-base
 ENV CUDA_VERSION_SHORT=cu130
+ENV PYTORCH_CUDA_VERSION=cu130
 
-FROM nvidia/cuda:12.8.1-base-ubuntu22.04 AS cu12.8-base
+FROM nvidia/cuda:12.8.1-base-ubuntu24.04 AS cu128-base
 ENV CUDA_VERSION_SHORT=cu128
+ENV PYTORCH_CUDA_VERSION=cu126
 
-FROM nvidia/cuda:12.6.3-base-ubuntu22.04 AS cu12-base
-ENV CUDA_VERSION_SHORT=cu126
+# Backward-compatible stage aliases.
+FROM cu130-base AS cu13-base
+FROM cu128-base AS cu12.8-base
+FROM cu128-base AS cu12-base
 
 FROM ${CUDA_VERSION}-base AS final
 ARG PYTHON_VERSION=3.12
 ENV PYTHON_VERSION=${PYTHON_VERSION}
 
+# Install the CUDA runtime requirements explicitly so torch is selected from
+# the matching PyTorch CUDA index before local wheels resolve dependencies.
 # Some dependencies such as timm (required by InternVL models) are listed in requirements/serve.txt.
-COPY requirements/serve.txt /tmp/requirements/serve.txt
+COPY requirements/common.txt requirements/runtime_cuda.txt requirements/serve.txt /tmp/requirements/
 COPY docker/install.sh /tmp/install.sh
 RUN --mount=type=cache,target=/root/.cache \
-    --mount=type=cache,target=/wheels,from=wheels,source=/wheels \
+    --mount=type=bind,target=/wheels,from=wheels,source=/wheels \
     /tmp/install.sh
 
 # explicitly set ptxas path for triton

docker/Dockerfile_dev

@@ -1,16 +1,23 @@
-FROM nvidia/cuda:12.8.1-devel-ubuntu22.04 AS cu12.8
+FROM nvidia/cuda:12.8.1-devel-ubuntu24.04 AS cu128
 
 # environment variables
 ENV DEBIAN_FRONTEND=noninteractive \
     TZ=Etc/UTC \
     PATH=/opt/py3/bin:/root/.local/bin:${PATH} \
     TRITON_PTXAS_PATH=/usr/local/cuda/bin/ptxas \
-    CUDA_VERSION_SHORT=cu128
+    CUDA_VERSION_SHORT=cu128 \
+    PYTORCH_CUDA_VERSION=cu126
 
 # Install dependencies and create python virtual environment
 RUN --mount=type=cache,target=/var/cache/apt \
     --mount=type=cache,target=/root/.cache \
-    sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' /etc/apt/sources.list && \
+    if [ -f /etc/apt/sources.list ]; then \
+        sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' /etc/apt/sources.list; \
+    fi && \
+    if [ -d /etc/apt/sources.list.d ]; then \
+        find /etc/apt/sources.list.d -type f -name '*.sources' -exec \
+            sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' {} +; \
+    fi && \
     apt-get update -y && \
     apt-get install -y --no-install-recommends \
         tzdata wget curl openssh-server ssh sudo git-core \
@@ -33,7 +40,9 @@ RUN mkdir -p /wheels
 
 # install lmdeploy and its dependencies
 RUN --mount=type=cache,target=/root/.cache/uv \
-    uv pip install -r requirements_cuda.txt --extra-index-url https://download.pytorch.org/whl/cu128 && \
+    grep -E '^torch(vision)?([<>=]|$)' requirements/runtime_cuda.txt > /tmp/requirements_torch.txt && \
+    uv pip install -r /tmp/requirements_torch.txt --index-url https://download.pytorch.org/whl/${PYTORCH_CUDA_VERSION} && \
+    uv pip install -r requirements_cuda.txt && \
     uv pip install -e .
 
 RUN --mount=type=cache,target=/root/.cache/uv \

docker/build.sh

@@ -2,7 +2,7 @@
 
 mkdir -p /wheels
 
-if [[ "${CUDA_VERSION_SHORT}" = "cu130" ]]; then
+if [[ "${CUDA_VERSION_SHORT}" == cu13* ]]; then
     pip install nvidia-nccl-cu13
 else
     pip install nvidia-nccl-cu12

docker/install.sh

@@ -4,20 +4,40 @@
 if [ ! -f "/opt/py3/bin/python" ]; then
     # install system packages
     export DEBIAN_FRONTEND=noninteractive TZ=Etc/UTC
-    sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' /etc/apt/sources.list
+    if [ -f /etc/apt/sources.list ]; then
+        sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' /etc/apt/sources.list
+    fi
+    if [ -d /etc/apt/sources.list.d ]; then
+        find /etc/apt/sources.list.d -type f -name '*.sources' -exec \
+            sed -i 's|http://archive.ubuntu.com|http://azure.archive.ubuntu.com|g' {} +
+    fi
     apt-get update -y
     apt-get install -y --no-install-recommends \
         tzdata wget curl ssh sudo git-core vim libibverbs1 ibverbs-providers ibverbs-utils librdmacm1 libibverbs-dev rdma-core libmlx5-1
 
-    if [[ ${PYTHON_VERSION} != "3.10" ]]; then
+    # Ubuntu may expose the interpreter package without the matching -dev and
+    # -venv packages. Check the complete set before deciding whether the PPA
+    # is needed.
+    python_packages=(
+        "python${PYTHON_VERSION}"
+        "python${PYTHON_VERSION}-dev"
+        "python${PYTHON_VERSION}-venv"
+    )
+    need_python_ppa=0
+    for package in "${python_packages[@]}"; do
+        if ! apt-cache show "${package}" >/dev/null 2>&1; then
+            need_python_ppa=1
+            break
+        fi
+    done
+    if ((need_python_ppa)); then
         apt-get install -y --no-install-recommends software-properties-common
         add-apt-repository -y ppa:deadsnakes/ppa
         apt-get update -y
     fi
 
     # install python, create virtual env
-    apt-get install -y --no-install-recommends \
-        python${PYTHON_VERSION} python${PYTHON_VERSION}-dev python${PYTHON_VERSION}-venv
+    apt-get install -y --no-install-recommends "${python_packages[@]}"
 
     pushd /opt >/dev/null
         python${PYTHON_VERSION} -m venv py3
@@ -46,13 +66,35 @@ export PATH=/opt/py3/bin:$PATH
 
 pip install -U pip wheel setuptools
 
-if [[ "${CUDA_VERSION_SHORT}" = "cu130" ]]; then
+# In the production image, install torch/torchvision from the matching CUDA
+# index first. Installing all local wheels directly would otherwise let pip
+# resolve torch from the default PyPI index.
+if [ -f /tmp/requirements/runtime_cuda.txt ]; then
+    grep -E '^torch(vision)?([<>=]|$)' \
+        /tmp/requirements/runtime_cuda.txt > /tmp/requirements/torch.txt
+    pip install -r /tmp/requirements/torch.txt \
+        --index-url "https://download.pytorch.org/whl/${PYTORCH_CUDA_VERSION}"
+    pip install -r /tmp/requirements/runtime_cuda.txt
+fi
+
+if [[ "${CUDA_VERSION_SHORT}" == cu13* ]]; then
     pip install nvidia-nvshmem-cu13==3.4.5
 else
     pip install nvidia-nvshmem-cu12==3.4.5
 fi
 
+for package in deep_ep deep_gemm flash_attn_3; do
+    if ! compgen -G "/wheels/${package}-*.whl" >/dev/null; then
+        echo "${package} wheel is missing from /wheels" >&2
+        exit 1
+    fi
+done
 pip install /wheels/*.whl
+python - <<'PY'
+import deep_ep
+import deep_gemm
+import lmdeploy.pytorch.third_party.flash_attn_interface
+PY
 pip install dlslime==0.0.2.post1
 
 pip install ninja einops packaging
@@ -67,4 +109,33 @@ if [[ "${CUDA_VERSION_SHORT}" = "cu128" ]]; then
     # window registration may cause memory leaks in NCCL 2.27, NCCL 2.28+ resolves the issue,
     # but turbomind engine will use nccl GIN for EP in future, which is brought in since 2.29
     pip install "nvidia-nccl-cu12>2.29"
+elif [[ "${CUDA_VERSION_SHORT}" == cu13* ]]; then
+    pip install "nvidia-nccl-cu13>2.29"
 fi
+
+python - <<'PY'
+import os
+from importlib.metadata import distributions
+
+import torch
+
+expected = os.environ['PYTORCH_CUDA_VERSION']
+actual = torch.version.cuda
+installed = {
+    dist.metadata['Name'].lower().replace('_', '-')
+    for dist in distributions()
+    if dist.metadata['Name']
+}
+if expected == 'cu126':
+    assert actual is not None and actual.startswith('12.6'), actual
+    unexpected = sorted(name for name in installed if name.endswith('-cu13'))
+elif expected == 'cu130':
+    assert actual is not None and actual.startswith('13.0'), actual
+    unexpected = sorted(name for name in installed if name.endswith('-cu12'))
+else:
+    raise AssertionError(f'unsupported PYTORCH_CUDA_VERSION: {expected}')
+assert not unexpected, f'unexpected CUDA packages: {unexpected}'
+print(f'Validated torch {torch.__version__} with CUDA {actual}')
+PY
+
+pip check

docker/prepare_3rdparty_wheel.sh

@@ -18,8 +18,8 @@ DEEP_GEMM_VERSION=88965b0
 FLASH_MLA_VERSION=1408756  # no release, pick the latest commit
 
 # DeepEP
-if [[ "${CUDA_VERSION_SHORT}" = "cu130" ]]; then
-    export CPLUS_INCLUDE_PATH="/usr/local/cuda/include/cccl":${CPLUS_INCLUDE_PATH}
+if [[ "${CUDA_VERSION_SHORT}" == cu13* ]]; then
+    export CPLUS_INCLUDE_PATH="/usr/local/cuda/include/cccl${CPLUS_INCLUDE_PATH:+:${CPLUS_INCLUDE_PATH}}"
     pip install nvidia-nvshmem-cu13==3.4.5
 else
     pip install nvidia-nvshmem-cu12==3.4.5
@@ -33,10 +33,16 @@ pip wheel -v --no-build-isolation --no-deps -w /wheels "git+https://github.com/d
 # sm100 compilation for Flash MLA requires NVCC 12.9 or higher
 FLASH_MLA_DISABLE_SM100=1 pip wheel -v --no-build-isolation --no-deps -w /wheels "git+https://github.com/deepseek-ai/FlashMLA.git@${FLASH_MLA_VERSION}"
 
-# flash_attn_3 (prebuilt wheels; CUDA + torch must match this image)
-TORCH_VER=$(python3 -c "import torch; print(''.join(torch.__version__.split('+')[0].split('.')))")
-FA3_WHEELS_URL="https://windreamer.github.io/flash-attention3-wheels/${CUDA_VERSION_SHORT}_torch${TORCH_VER}"
-pip download --no-deps -d /wheels "flash_attn_3" --find-links "${FA3_WHEELS_URL}"
+# flash_attn_3 (official, LibTorch ABI-stable wheels for torch >= 2.9).
+# Select the FA3 channel matching the PyTorch binary. CUDA 13 FA3 wheels are
+# published in the cu130 channel.
+FA3_CUDA_VERSION="${PYTORCH_CUDA_VERSION:-${CUDA_VERSION_SHORT}}"
+if [[ "${FA3_CUDA_VERSION}" == cu13* ]]; then
+    FA3_CUDA_VERSION=cu130
+fi
+pip download --no-deps --only-binary=:all: -d /wheels \
+    "flash-attn-3==3.0.0" \
+    --index-url "https://download.pytorch.org/whl/${FA3_CUDA_VERSION}"
 
 # GDRCopy debs
 apt-get update -y \

lmdeploy/pytorch/check_env/triton.py

@@ -3,7 +3,7 @@
 
 from .base import BaseChecker
 
-MAX_TRITON_VERSION = '3.6.0'
+MAX_TRITON_VERSION = '3.7.1'
 MIN_TRITON_VERSION = '3.0.0'
 
 

requirements/runtime_cuda.txt

@@ -8,6 +8,6 @@ opencv-python-headless
 peft<=0.14.0
 prometheus_client
 tilelang==0.1.11; sys_platform == "linux" and "aarch64" not in platform_machine and "arm" not in platform_machine
-torch<=2.10.0,>=2.0.0
-torchvision<=0.25.0,>=0.15.0
-triton<=3.6.0,>=3.0.0; sys_platform == "linux" and "aarch64" not in platform_machine and "arm" not in platform_machine
+torch<=2.12.1,>=2.0.0
+torchvision<=0.27.1,>=0.15.0
+triton<=3.7.1,>=3.0.0; sys_platform == "linux" and "aarch64" not in platform_machine and "arm" not in platform_machine