gpustack-runner 0.1.23.post4__tar.gz → 0.1.23.post5__tar.gz

{gpustack_runner-0.1.23.post4 → gpustack_runner-0.1.23.post5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpustack-runner
-Version: 0.1.23.post4
+Version: 0.1.23.post5
 Summary: GPUStack Runner is library for registering runnable accelerated backends and services in GPUStack.
 Project-URL: Homepage, https://github.com/gpustack/runner
 Project-URL: Bug Tracker, https://github.com/gpustack/gpustack/issues
@@ -97,9 +97,9 @@ The following table lists the supported accelerated backends and their correspon
 
 ### Hygon DTK
 
-| DTK Version <br/> (Variant) | vLLM             |
-|-----------------------------|------------------|
-| 25.04                       | `0.9.2`, `0.8.5` |
+| DTK Version <br/> (Variant) | vLLM                       |
+|-----------------------------|----------------------------|
+| 25.04                       | `0.11.0`, `0.9.2`, `0.8.5` |
 
 ### MetaX MACA
 
@@ -108,6 +108,13 @@ The following table lists the supported accelerated backends and their correspon
 | 3.2                          | `0.10.2` |
 | 3.0                          | `0.9.1`  |
 
+### MThreads MUSA
+
+| MUSA Version <br/> (Variant) | vLLM    | SGLang  |
+|------------------------------|---------|---------|
+| 4.3.2                        |         | `0.5.2` |
+| 4.1.0                        | `0.9.2` |         |
+
 ### AMD ROCm
 
 > [!CAUTION]
@@ -171,6 +178,7 @@ ARG PYTHON_VERSION=...                                 # REQUIRED
 ARG CMAKE_MAX_JOBS=...                                 # REQUIRED
 ARG {OTHERS}                                           # OPTIONAL
 ARG {BACKEND}_VERSION=...                              # REQUIRED
+ARG {BACKEND}_VERSION_EXTRA=...                        # OPTIONAL
 ARG {BACKEND}_ARCHS=...                                # REQUIRED
 ARG {BACKEND}_{OTHERS}=...                             # OPTIONAL
 ARG {SERVICE}_BASE_IMAGE=...                           # REQUIRED

{gpustack_runner-0.1.23.post4 → gpustack_runner-0.1.23.post5}/README.md

@@ -77,9 +77,9 @@ The following table lists the supported accelerated backends and their correspon
 
 ### Hygon DTK
 
-| DTK Version <br/> (Variant) | vLLM             |
-|-----------------------------|------------------|
-| 25.04                       | `0.9.2`, `0.8.5` |
+| DTK Version <br/> (Variant) | vLLM                       |
+|-----------------------------|----------------------------|
+| 25.04                       | `0.11.0`, `0.9.2`, `0.8.5` |
 
 ### MetaX MACA
 
@@ -88,6 +88,13 @@ The following table lists the supported accelerated backends and their correspon
 | 3.2                          | `0.10.2` |
 | 3.0                          | `0.9.1`  |
 
+### MThreads MUSA
+
+| MUSA Version <br/> (Variant) | vLLM    | SGLang  |
+|------------------------------|---------|---------|
+| 4.3.2                        |         | `0.5.2` |
+| 4.1.0                        | `0.9.2` |         |
+
 ### AMD ROCm
 
 > [!CAUTION]
@@ -151,6 +158,7 @@ ARG PYTHON_VERSION=...                                 # REQUIRED
 ARG CMAKE_MAX_JOBS=...                                 # REQUIRED
 ARG {OTHERS}                                           # OPTIONAL
 ARG {BACKEND}_VERSION=...                              # REQUIRED
+ARG {BACKEND}_VERSION_EXTRA=...                        # OPTIONAL
 ARG {BACKEND}_ARCHS=...                                # REQUIRED
 ARG {BACKEND}_{OTHERS}=...                             # OPTIONAL
 ARG {SERVICE}_BASE_IMAGE=...                           # REQUIRED

{gpustack_runner-0.1.23.post4 → gpustack_runner-0.1.23.post5}/gpustack_runner/_version.py

@@ -27,8 +27,8 @@ version_tuple: VERSION_TUPLE
 __commit_id__: COMMIT_ID
 commit_id: COMMIT_ID
 
-__version__ = version = '0.1.23.post4'
-__version_tuple__ = version_tuple = (0, 1, 23, 'post4')
+__version__ = version = '0.1.23.post5'
+__version_tuple__ = version_tuple = (0, 1, 23, 'post5')
 try:
     from ._version_appendix import git_commit
     __commit_id__ = commit_id = git_commit

gpustack_runner-0.1.23.post5/gpustack_runner/_version_appendix.py

@@ -0,0 +1 @@
+git_commit = "d297d69"

{gpustack_runner-0.1.23.post4 → gpustack_runner-0.1.23.post5}/gpustack_runner/cmds/images.py

@@ -444,14 +444,14 @@ class SaveImagesSubCommand(SubCommand):
 
                 command = [
                     "skopeo",
-                    "--override-os",
-                    override_os,
-                    "--override-arch",
-                    override_arch,
                     "copy",
                     "--src-tls-verify=false",
                     "--retry-times",
                     str(self.max_retries),
+                    "--override-os",
+                    override_os,
+                    "--override-arch",
+                    override_arch,
                 ]
                 if self.source_username and self.source_password:
                     command.extend(
@@ -771,6 +771,10 @@ class CopyImagesSubCommand(SubCommand):
                 print(f"❌ Error syncing image '{img_name}'")
                 failures.append((img_name, img_err))
 
+            override_os, override_arch = None, None
+            if self.platform:
+                override_os, override_arch = self.platform.split("/", maxsplit=1)
+
             # Submit tasks
             for img in images:
                 command = [
@@ -778,10 +782,20 @@ class CopyImagesSubCommand(SubCommand):
                     "copy",
                     "--src-tls-verify=false",
                     "--dest-tls-verify=false",
-                    "--all",
                     "--retry-times",
                     str(self.max_retries),
                 ]
+                if override_os and override_arch:
+                    command.extend(
+                        [
+                            "--override-os",
+                            override_os,
+                            "--override-arch",
+                            override_arch,
+                        ],
+                    )
+                else:
+                    command.append("--all")
                 if self.source_username and self.source_password:
                     command.extend(
                         [

{gpustack_runner-0.1.23.post4 → gpustack_runner-0.1.23.post5}/gpustack_runner/runner.py

@@ -13,7 +13,7 @@ from dataclasses_json import dataclass_json
 from . import envs
 
 _RE_DOCKER_IMAGE = re.compile(
-    r"(?:(?P<prefix>[\w\\.\-]+(?:/[\w\\.\-]+)*)/)?runner:(?P<backend>(Host|cann|corex|cuda|dtk|maca|rocm))(?P<backend_version>[XY\d\\.]+)(?:-(?P<backend_variant>\w+))?-(?P<service>(vllm|voxbox|mindie|sglang))(?P<service_version>[\w\\.]+)(?:-(?P<suffix>\w+))?",
+    r"(?:(?P<prefix>[\w\\.\-]+(?:/[\w\\.\-]+)*)/)?runner:(?P<backend>(Host|cann|corex|cuda|dtk|maca|musa|rocm))(?P<backend_version>[XY\d\\.]+)(?:-(?P<backend_variant>\w+))?-(?P<service>(vllm|voxbox|mindie|sglang))(?P<service_version>[\w\\.]+)(?:-(?P<suffix>\w+))?",
 )
 """
 Regex for Docker image parsing,

{gpustack_runner-0.1.23.post4 → gpustack_runner-0.1.23.post5}/gpustack_runner/runner.py.json

@@ -1363,6 +1363,17 @@
     "docker_image": "gpustack/runner:cuda12.4-voxbox0.0.20",
     "deprecated": true
   },
+  {
+    "backend": "dtk",
+    "backend_version": "25.04",
+    "original_backend_version": "25.04.2",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.11.0",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:dtk25.04-vllm0.11.0",
+    "deprecated": false
+  },
   {
     "backend": "dtk",
     "backend_version": "25.04",
@@ -1407,6 +1418,28 @@
     "docker_image": "gpustack/runner:maca3.0-vllm0.9.1",
     "deprecated": false
   },
+  {
+    "backend": "musa",
+    "backend_version": "4.3",
+    "original_backend_version": "4.3.2",
+    "backend_variant": "",
+    "service": "sglang",
+    "service_version": "0.5.7",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:musa4.3-sglang0.5.7",
+    "deprecated": false
+  },
+  {
+    "backend": "musa",
+    "backend_version": "4.1",
+    "original_backend_version": "4.1.0",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.9.2",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:musa4.1-vllm0.9.2",
+    "deprecated": false
+  },
   {
     "backend": "rocm",
     "backend_version": "7.0",

gpustack_runner-0.1.23.post5/pack/.post_operation/20260105_vllm_install_omni/cann/Dockerfile

@@ -0,0 +1,81 @@
+ARG CMAKE_MAX_JOBS
+ARG CANN_VERSION=8.3
+ARG CANN_ARCHS=910b
+ARG VLLM_VERSION=0.12.0
+ARG VLLM_OMNI_COMMIT=75cdf1c
+
+FROM gpustack/runner:cann${CANN_VERSION}-${CANN_ARCHS}-vllm${VLLM_VERSION} AS vllm-build-omni
+SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
+
+ARG TARGETPLATFORM
+ARG TARGETOS
+ARG TARGETARCH
+
+## Build Omni
+
+ARG CMAKE_MAX_JOBS
+ARG VLLM_OMNI_COMMIT
+
+ENV VLLM_OMNI_COMMIT=${VLLM_OMNI_COMMIT}
+
+RUN <<EOF
+    # Omni
+
+    CMAKE_MAX_JOBS="${CMAKE_MAX_JOBS}"
+    if [[ -z "${CMAKE_MAX_JOBS}" ]]; then
+        CMAKE_MAX_JOBS="$(( $(nproc) / 2 ))"
+    fi
+    if (( $(echo "${CMAKE_MAX_JOBS} > 4" | bc -l) )); then
+        CMAKE_MAX_JOBS="4"
+    fi
+    export MAX_JOBS="${CMAKE_MAX_JOBS}"
+    export COMPILE_CUSTOM_KERNELS=1
+    export LD_LIBRARY_PATH="${LD_LIBRARY_PATH}:${CANN_HOME}/ascend-toolkit/latest/$(uname -i)-linux/devlib"
+    export VLLM_TARGET_DEVICE="empty"
+    echo "Building vLLM Omni with the following environment variables:"
+    env
+
+    # Build
+    git -C /tmp clone --recursive --shallow-submodules \
+        https://github.com/vllm-project/vllm-omni vllm_omni \
+        && pushd /tmp/vllm_omni \
+        && git checkout ${VLLM_OMNI_COMMIT} \
+        && git submodule update --init --recursive
+    pushd /tmp/vllm_omni \
+        && python -v -m build --no-isolation --wheel \
+        && tree -hs /tmp/vllm_omni/dist \
+        && mv /tmp/vllm_omni/dist /workspace
+
+    # Cleanup
+    rm -rf /var/tmp/* \
+        && rm -rf /tmp/*
+EOF
+
+FROM gpustack/runner:cann${CANN_VERSION}-${CANN_ARCHS}-vllm${VLLM_VERSION} AS vllm
+SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
+
+ARG TARGETPLATFORM
+ARG TARGETOS
+ARG TARGETARCH
+
+## Install Omni
+
+RUN --mount=type=bind,from=vllm-build-omni,source=/,target=/omni,rw <<EOF
+    # Omni
+
+    # Install
+    uv pip install --no-build-isolation \
+        /omni/workspace/*.whl
+
+    # Review
+    uv pip tree
+
+    # Cleanup
+    rm -rf /var/tmp/* \
+        && rm -rf /tmp/*
+EOF
+
+## Entrypoint
+
+WORKDIR /
+ENTRYPOINT [ "tini", "--" ]

gpustack_runner-0.1.23.post5/pack/.post_operation/20260105_vllm_install_omni/cuda/Dockerfile

@@ -0,0 +1,93 @@
+ARG CMAKE_MAX_JOBS
+ARG CUDA_VERSION=12.8
+ARG VLLM_VERSION=0.12.0
+ARG VLLM_OMNI_COMMIT=75cdf1c
+
+FROM gpustack/runner:cuda${CUDA_VERSION}-vllm${VLLM_VERSION} AS vllm-build-omni
+SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
+
+ARG TARGETPLATFORM
+ARG TARGETOS
+ARG TARGETARCH
+
+ARG TARGETPLATFORM
+ARG TARGETOS
+ARG TARGETARCH
+
+## Build Omni
+
+ARG CMAKE_MAX_JOBS
+ARG VLLM_OMNI_COMMIT
+
+ENV VLLM_OMNI_COMMIT=${VLLM_OMNI_COMMIT}
+
+RUN <<EOF
+    # Omni
+
+    IFS="." read -r CUDA_MAJOR CUDA_MINOR CUDA_PATCH <<< "${VLLM_TORCH_CUDA_VERSION}"
+
+    CMAKE_MAX_JOBS="${CMAKE_MAX_JOBS}"
+    if [[ -z "${CMAKE_MAX_JOBS}" ]]; then
+        CMAKE_MAX_JOBS="$(( $(nproc) / 2 ))"
+    fi
+    if (( $(echo "${CMAKE_MAX_JOBS} > 4" | bc -l) )); then
+        CMAKE_MAX_JOBS="4"
+    fi
+    VL_CUDA_ARCHS="${CUDA_ARCHS}"
+    if [[ -z "${VL_CUDA_ARCHS}" ]]; then
+        if (( $(echo "${CUDA_MAJOR}.${CUDA_MINOR} < 12.9" | bc -l) )); then
+            VL_CUDA_ARCHS="7.5 8.0+PTX 8.9 9.0 10.0+PTX 12.0+PTX"
+        else
+            VL_CUDA_ARCHS="7.5 8.0+PTX 8.9 9.0 10.0 10.3 12.0 12.1+PTX"
+        fi
+    fi
+    export MAX_JOBS="${CMAKE_MAX_JOBS}"
+    export TORCH_CUDA_ARCH_LIST="${VL_CUDA_ARCHS}"
+    export NVCC_THREADS=1
+    echo "Building vLLM Omni with the following environment variables:"
+    env
+
+    # Build
+    git -C /tmp clone --recursive --shallow-submodules \
+        https://github.com/vllm-project/vllm-omni vllm_omni \
+        && pushd /tmp/vllm_omni \
+        && git checkout ${VLLM_OMNI_COMMIT} \
+        && git submodule update --init --recursive
+    pushd /tmp/vllm_omni \
+        && python -v -m build --no-isolation --wheel \
+        && tree -hs /tmp/vllm_omni/dist \
+        && mv /tmp/vllm_omni/dist /workspace
+
+    # Cleanup
+    rm -rf /var/tmp/* \
+        && rm -rf /tmp/*
+EOF
+
+FROM gpustack/runner:cuda${CUDA_VERSION}-vllm${VLLM_VERSION} AS vllm
+SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
+
+ARG TARGETPLATFORM
+ARG TARGETOS
+ARG TARGETARCH
+
+## Install Omni
+
+RUN --mount=type=bind,from=vllm-build-omni,source=/,target=/omni,rw <<EOF
+    # Omni
+
+    # Install
+    uv pip install --no-build-isolation \
+        /omni/workspace/*.whl
+
+    # Review
+    uv pip tree
+
+    # Cleanup
+    rm -rf /var/tmp/* \
+        && rm -rf /tmp/*
+EOF
+
+## Entrypoint
+
+WORKDIR /
+ENTRYPOINT [ "tini", "--" ]

gpustack_runner-0.1.23.post5/pack/.post_operation/20260105_vllm_install_omni/matrix.yaml

@@ -0,0 +1,78 @@
+rules:
+
+  #
+  # NVIDIA CUDA
+  #
+
+  ## Packed NVIDIA CUDA 12.9.
+  ##
+  - backend: "cuda"
+    services:
+      - "vllm"
+    args:
+      - "CUDA_VERSION=12.9"
+      - "VLLM_VERSION=0.12.0"
+  ## Packed NVIDIA CUDA 12.8.
+  ##
+  - backend: "cuda"
+    services:
+      - "vllm"
+    args:
+      - "CUDA_VERSION=12.8"
+      - "VLLM_VERSION=0.12.0"
+  ## Packed NVIDIA CUDA 12.6.
+  ##
+  - backend: "cuda"
+    services:
+      - "vllm"
+    args:
+      - "CUDA_VERSION=12.6"
+      - "VLLM_VERSION=0.12.0"
+
+  #
+  # AMD ROCm
+  #
+
+  ## Packed AMD ROCm 7.0.
+  ##
+  - backend: "rocm"
+    services:
+      - "vllm"
+    platforms:
+      - "linux/amd64"
+    args:
+      - "ROCM_VERSION=7.0"
+      - "VLLM_VERSION=0.12.0"
+  ## Packed AMD ROCm 6.4.
+  ##
+  - backend: "rocm"
+    services:
+      - "vllm"
+    platforms:
+      - "linux/amd64"
+    args:
+      - "ROCM_VERSION=6.4"
+      - "VLLM_VERSION=0.12.0"
+
+  #
+  # Ascend CANN
+  #
+
+  ## Packed Ascend CANN 8.3, using CANN Kernel for A3.
+  ##
+  - backend: "cann"
+    services:
+      - "vllm"
+    args:
+      - "CANN_VERSION=8.3"
+      - "CANN_ARCHS=a3"
+      - "VLLM_VERSION=0.12.0"
+  ## Packed Ascend CANN 8.3, using CANN Kernel for 910B.
+  ##
+  - backend: "cann"
+    services:
+      - "vllm"
+    args:
+      - "CANN_VERSION=8.3"
+      - "CANN_ARCHS=910b"
+      - "VLLM_VERSION=0.12.0"

gpustack_runner-0.1.23.post5/pack/.post_operation/20260105_vllm_install_omni/rocm/Dockerfile

@@ -0,0 +1,98 @@
+ARG CMAKE_MAX_JOBS
+ARG ROCM_VERSION=6.4
+ARG VLLM_VERSION=0.12.0
+ARG VLLM_OMNI_COMMIT=75cdf1c
+
+FROM gpustack/runner:rocm${ROCM_VERSION}-vllm${VLLM_VERSION} AS vllm-build-omni
+SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
+
+ARG TARGETPLATFORM
+ARG TARGETOS
+ARG TARGETARCH
+
+## Build Omni
+
+ARG CMAKE_MAX_JOBS
+ARG VLLM_OMNI_COMMIT
+
+ENV VLLM_OMNI_COMMIT=${VLLM_OMNI_COMMIT}
+
+RUN <<EOF
+    # Omni
+
+    IFS="." read -r ROCM_MAJOR ROCM_MINOR ROCM_PATCH <<< "${VLLM_TORCH_ROCM_VERSION}"
+    IFS="." read -r VL_MAJOR VL_MINOR VL_PATCH <<< "${VLLM_VERSION}"
+
+    CMAKE_MAX_JOBS="${CMAKE_MAX_JOBS}"
+    if [[ -z "${CMAKE_MAX_JOBS}" ]]; then
+        CMAKE_MAX_JOBS="$(( $(nproc) / 2 ))"
+    fi
+    if (( $(echo "${CMAKE_MAX_JOBS} > 4" | bc -l) )); then
+        CMAKE_MAX_JOBS="4"
+    fi
+    VL_ROCM_ARCHS="${ROCM_ARCHS}"
+    if [[ -z "${VL_ROCM_ARCHS}" ]]; then
+        if (( $(echo "${ROCM_MAJOR}.${ROCM_MINOR} < 7.0" | bc -l) )); then
+            VL_ROCM_ARCHS="gfx908;gfx90a;gfx942;gfx1030;gfx1100"
+            if (( $(echo "${VL_MAJOR}.${VL_MINOR} == 0.13" | bc -l) )); then
+                # TODO(thxCode): Temporarily remove gfx1030 for vLLM ROCm build due to build error in ROCm 6.4.4.
+                # #15 134.9 /tmp/vllm/build/temp.linux-x86_64-cpython-312/csrc/sampler.hip:564:63: error: local memory (66032) exceeds limit (65536) in 'void vllm::topKPerRowDecode<1024, true, false, true>(float const*, int const*, int*, int, int, int, int, float*, int, int const*)'
+                # ##15 134.9   564 | static __global__ __launch_bounds__(kNumThreadsPerBlock) void topKPerRowDecode(
+                # ##15 134.9       |                                                               ^
+                # ##15 134.9 16 warnings and 1 error generated when compiling for gfx1030.
+                VL_ROCM_ARCHS="gfx908;gfx90a;gfx942"
+            fi
+        else
+            VL_ROCM_ARCHS="gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151"
+        fi
+    fi
+    export MAX_JOBS="${CMAKE_MAX_JOBS}"
+    export COMPILE_CUSTOM_KERNELS=1
+    export PYTORCH_ROCM_ARCH="${VL_ROCM_ARCHS}"
+    echo "Building vLLM Omni with the following environment variables:"
+    env
+
+    # Build
+    git -C /tmp clone --recursive --shallow-submodules \
+        https://github.com/vllm-project/vllm-omni vllm_omni \
+        && pushd /tmp/vllm_omni \
+        && git checkout ${VLLM_OMNI_COMMIT} \
+        && git submodule update --init --recursive
+    pushd /tmp/vllm_omni \
+        && python -v -m build --no-isolation --wheel \
+        && tree -hs /tmp/vllm_omni/dist \
+        && mv /tmp/vllm_omni/dist /workspace
+
+    # Cleanup
+    rm -rf /var/tmp/* \
+        && rm -rf /tmp/*
+EOF
+
+FROM gpustack/runner:rocm${ROCM_VERSION}-vllm${VLLM_VERSION} AS vllm
+SHELL ["/bin/bash", "-eo", "pipefail", "-c"]
+
+ARG TARGETPLATFORM
+ARG TARGETOS
+ARG TARGETARCH
+
+## Install Omni
+
+RUN --mount=type=bind,from=vllm-build-omni,source=/,target=/omni,rw <<EOF
+    # Omni
+
+    # Install
+    uv pip install --no-build-isolation \
+        /omni/workspace/*.whl
+
+    # Review
+    uv pip tree
+
+    # Cleanup
+    rm -rf /var/tmp/* \
+        && rm -rf /tmp/*
+EOF
+
+## Entrypoint
+
+WORKDIR /
+ENTRYPOINT [ "tini", "--" ]

{gpustack_runner-0.1.23.post4 → gpustack_runner-0.1.23.post5}/pack/.post_operation/README.md

@@ -32,3 +32,4 @@ We leverage the matrix expansion feature of GPUStack Runner to achieve this, and
 - [x] 2025-12-19: Install `vLLM[audio]` packages for vLLM 0.12.0/0.11.2 of CUDA/ROCm released images.
 - [x] 2025-12-19: Install `petit-kernel` package for vLLM 0.12.0/0.11.2 and SGLang 0.5.6.post2/0.5.5.post3 of ROcm released images.
 - [x] 2025-12-24: Apply ATB config patches to MindIE 2.2.rc1 for CANN released images.
+- [ ] 2026-01-05: Install `vllm-omni` packages for vLLM 0.12.0 of CUDA/ROCm/CANN released images.