gpustack-runner 0.1.22.post2__tar.gz → 0.1.22.post4__tar.gz

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/Makefile

@@ -117,7 +117,7 @@ package:
 				JOB_EXTRA_ARGS+=("--cache-from=type=registry,ref=gpustack/runner-build-cache:$${TAG_CACHE}"); \
 			done; \
 		fi; \
-		if [[ "$(PACKAGE_PUSH)" == "true" ]] || [[ "$(PACKAGE_CACHE_PUSH)" == "true" ]]; then \
+		if [[ "$(PACKAGE_PUSH)" == "true" || "$(PACKAGE_CACHE_PUSH)" == "true" ]] && [[ -z "$(PACKAGE_POST_OPERATION)" ]]; then \
 			for TAG_CACHE in $${JOB_PLATFORM_CACHE}; do \
 				JOB_EXTRA_ARGS+=("--cache-to=type=registry,ignore-error=true,mode=max,compression=gzip,ref=$(PACKAGE_NAMESPACE)/$(PACKAGE_CACHE_REPOSITORY):$${TAG_CACHE}"); \
 			done; \

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gpustack-runner
-Version: 0.1.22.post2
+Version: 0.1.22.post4
 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
@@ -86,12 +86,12 @@ The following table lists the supported accelerated backends and their correspon
 > - Applied [Qwen2.5 VL patched](https://github.com/gpustack/gpustack/issues/3606) to vLLM 0.11.2.
 > - Applied [vLLM[audio] packages](https://github.com/vllm-project/vllm/blob/275de34170654274616082721348b7edd9741d32/setup.py#L720-L724) to vLLM 0.11.2.
 
-| CUDA Version <br/> (Variant) | vLLM                                                                       | SGLang                                                    | VoxBox   |
-|------------------------------|----------------------------------------------------------------------------|-----------------------------------------------------------|----------|
-| 12.9                         | `0.12.0`, **`0.11.2`**                                                     | `0.5.6.post2`                                             |          |
-| 12.8                         | `0.12.0`, **`0.11.2`**, <br/>`0.11.0`, `0.10.2`, <br/>`0.10.1.1`, `0.10.0` | `0.5.6.post2`, `0.5.5.post3`, <br/>`0.5.5`, `0.5.4.post3` | `0.0.20` |
-| 12.6                         | `0.12.0`, **`0.11.2`**, <br/>`0.11.0`, `0.10.2`, <br/>`0.10.1.1`, `0.10.0` | `0.5.6.post2`                                             | `0.0.20` |
-| 12.4                         | `0.11.0`, `0.10.2`, <br/>`0.10.1.1`, `0.10.0`                              |                                                           | `0.0.20` |
+| CUDA Version <br/> (Variant) | vLLM                                                                                      | SGLang                                                    | VoxBox             |
+|------------------------------|-------------------------------------------------------------------------------------------|-----------------------------------------------------------|--------------------|
+| 12.9                         | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**                                                     | `0.5.6.post2`                                             |                    |
+| 12.8                         | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**, `0.11.0`, <br/>`0.10.2`, `0.10.1.1`, <br/>`0.10.0` | `0.5.6.post2`, `0.5.5.post3`, <br/>`0.5.5`, `0.5.4.post3` | `0.0.21`, `0.0.20` |
+| 12.6                         | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**, `0.11.0`, <br/>`0.10.2`, `0.10.1.1`, <br/>`0.10.0` | `0.5.6.post2`                                             | `0.0.21`, `0.0.20` |
+| 12.4                         | `0.11.0`, `0.10.2`, <br/>`0.10.1.1`, `0.10.0`                                             |                                                           | `0.0.20`           |
 
 ### Hygon DTK
 
@@ -118,16 +118,17 @@ The following table lists the supported accelerated backends and their correspon
 > - ROCm 7.0 vLLM `0.11.2/0.11.0` are reusing the official ROCm 6.4 PyTorch 2.9 wheel package rather than a ROCm
     7.0 specific PyTorch build. Although supports ROCm 7.0 in vLLM `0.11.2/0.11.0`, `gfx1150/gfx1151` are not supported yet.
 > - SGLang supports `gfx942` only.
+> - ROCm 6.4 vLLM `0.13.0` supports `gfx903 gfx90a gfx942` only.
 
 > [!IMPORTANT]
 > - Applied [vLLM[audio] packages](https://github.com/vllm-project/vllm/blob/275de34170654274616082721348b7edd9741d32/setup.py#L720-L724) to vLLM 0.11.2.
 > - Applied [petit-kernel package](https://github.com/vllm-project/vllm/blob/275de34170654274616082721348b7edd9741d32/setup.py#L728) to vLLM 0.11.2 and SGLang 0.5.5.post3.
 
-| ROCm Version <br/> (Variant) | vLLM                                   | SGLang                           |
-|------------------------------|----------------------------------------|----------------------------------|
-| 7.0                          | `0.12.0`, **`0.11.2`**, <br/> `0.11.0` | `0.5.6.post2`                    |
-| 6.4                          | `0.12.0`, **`0.11.2`**, <br/> `0.10.2` | `0.5.6.post2`, **`0.5.5.post3`** |
-| 6.3                          | `0.10.1.1`, `0.10.0`                   |                                  |
+| ROCm Version <br/> (Variant) | vLLM                                            | SGLang                           |
+|------------------------------|-------------------------------------------------|----------------------------------|
+| 7.0                          | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**, `0.11.0` | `0.5.6.post2`                    |
+| 6.4                          | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**, `0.10.2` | `0.5.6.post2`, **`0.5.5.post3`** |
+| 6.3                          | `0.10.1.1`, `0.10.0`                            |                                  |
 
 ## Directory Structure
 

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/README.md

@@ -66,12 +66,12 @@ The following table lists the supported accelerated backends and their correspon
 > - Applied [Qwen2.5 VL patched](https://github.com/gpustack/gpustack/issues/3606) to vLLM 0.11.2.
 > - Applied [vLLM[audio] packages](https://github.com/vllm-project/vllm/blob/275de34170654274616082721348b7edd9741d32/setup.py#L720-L724) to vLLM 0.11.2.
 
-| CUDA Version <br/> (Variant) | vLLM                                                                       | SGLang                                                    | VoxBox   |
-|------------------------------|----------------------------------------------------------------------------|-----------------------------------------------------------|----------|
-| 12.9                         | `0.12.0`, **`0.11.2`**                                                     | `0.5.6.post2`                                             |          |
-| 12.8                         | `0.12.0`, **`0.11.2`**, <br/>`0.11.0`, `0.10.2`, <br/>`0.10.1.1`, `0.10.0` | `0.5.6.post2`, `0.5.5.post3`, <br/>`0.5.5`, `0.5.4.post3` | `0.0.20` |
-| 12.6                         | `0.12.0`, **`0.11.2`**, <br/>`0.11.0`, `0.10.2`, <br/>`0.10.1.1`, `0.10.0` | `0.5.6.post2`                                             | `0.0.20` |
-| 12.4                         | `0.11.0`, `0.10.2`, <br/>`0.10.1.1`, `0.10.0`                              |                                                           | `0.0.20` |
+| CUDA Version <br/> (Variant) | vLLM                                                                                      | SGLang                                                    | VoxBox             |
+|------------------------------|-------------------------------------------------------------------------------------------|-----------------------------------------------------------|--------------------|
+| 12.9                         | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**                                                     | `0.5.6.post2`                                             |                    |
+| 12.8                         | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**, `0.11.0`, <br/>`0.10.2`, `0.10.1.1`, <br/>`0.10.0` | `0.5.6.post2`, `0.5.5.post3`, <br/>`0.5.5`, `0.5.4.post3` | `0.0.21`, `0.0.20` |
+| 12.6                         | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**, `0.11.0`, <br/>`0.10.2`, `0.10.1.1`, <br/>`0.10.0` | `0.5.6.post2`                                             | `0.0.21`, `0.0.20` |
+| 12.4                         | `0.11.0`, `0.10.2`, <br/>`0.10.1.1`, `0.10.0`                                             |                                                           | `0.0.20`           |
 
 ### Hygon DTK
 
@@ -98,16 +98,17 @@ The following table lists the supported accelerated backends and their correspon
 > - ROCm 7.0 vLLM `0.11.2/0.11.0` are reusing the official ROCm 6.4 PyTorch 2.9 wheel package rather than a ROCm
     7.0 specific PyTorch build. Although supports ROCm 7.0 in vLLM `0.11.2/0.11.0`, `gfx1150/gfx1151` are not supported yet.
 > - SGLang supports `gfx942` only.
+> - ROCm 6.4 vLLM `0.13.0` supports `gfx903 gfx90a gfx942` only.
 
 > [!IMPORTANT]
 > - Applied [vLLM[audio] packages](https://github.com/vllm-project/vllm/blob/275de34170654274616082721348b7edd9741d32/setup.py#L720-L724) to vLLM 0.11.2.
 > - Applied [petit-kernel package](https://github.com/vllm-project/vllm/blob/275de34170654274616082721348b7edd9741d32/setup.py#L728) to vLLM 0.11.2 and SGLang 0.5.5.post3.
 
-| ROCm Version <br/> (Variant) | vLLM                                   | SGLang                           |
-|------------------------------|----------------------------------------|----------------------------------|
-| 7.0                          | `0.12.0`, **`0.11.2`**, <br/> `0.11.0` | `0.5.6.post2`                    |
-| 6.4                          | `0.12.0`, **`0.11.2`**, <br/> `0.10.2` | `0.5.6.post2`, **`0.5.5.post3`** |
-| 6.3                          | `0.10.1.1`, `0.10.0`                   |                                  |
+| ROCm Version <br/> (Variant) | vLLM                                            | SGLang                           |
+|------------------------------|-------------------------------------------------|----------------------------------|
+| 7.0                          | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**, `0.11.0` | `0.5.6.post2`                    |
+| 6.4                          | `0.13.0`, `0.12.0`, <br/>**`0.11.2`**, `0.10.2` | `0.5.6.post2`, **`0.5.5.post3`** |
+| 6.3                          | `0.10.1.1`, `0.10.0`                            |                                  |
 
 ## Directory Structure
 

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/gpustack_runner/_version.py

@@ -27,8 +27,8 @@ version_tuple: VERSION_TUPLE
 __commit_id__: COMMIT_ID
 commit_id: COMMIT_ID
 
-__version__ = version = '0.1.22.post2'
-__version_tuple__ = version_tuple = (0, 1, 22, 'post2')
+__version__ = version = '0.1.22.post4'
+__version_tuple__ = version_tuple = (0, 1, 22, 'post4')
 try:
     from ._version_appendix import git_commit
     __commit_id__ = commit_id = git_commit

gpustack_runner-0.1.22.post4/gpustack_runner/_version_appendix.py

@@ -0,0 +1 @@
+git_commit = "f3f4d02"

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/gpustack_runner/runner.py.json

@@ -604,6 +604,28 @@
     "docker_image": "gpustack/runner:cuda12.9-sglang0.5.6.post2",
     "deprecated": false
   },
+  {
+    "backend": "cuda",
+    "backend_version": "12.9",
+    "original_backend_version": "12.9.1",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.13.0",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:cuda12.9-vllm0.13.0",
+    "deprecated": false
+  },
+  {
+    "backend": "cuda",
+    "backend_version": "12.9",
+    "original_backend_version": "12.9.1",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.13.0",
+    "platform": "linux/arm64",
+    "docker_image": "gpustack/runner:cuda12.9-vllm0.13.0",
+    "deprecated": false
+  },
   {
     "backend": "cuda",
     "backend_version": "12.9",
@@ -725,6 +747,28 @@
     "docker_image": "gpustack/runner:cuda12.8-sglang0.5.4.post3",
     "deprecated": false
   },
+  {
+    "backend": "cuda",
+    "backend_version": "12.8",
+    "original_backend_version": "12.8.1",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.13.0",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:cuda12.8-vllm0.13.0",
+    "deprecated": false
+  },
+  {
+    "backend": "cuda",
+    "backend_version": "12.8",
+    "original_backend_version": "12.8.1",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.13.0",
+    "platform": "linux/arm64",
+    "docker_image": "gpustack/runner:cuda12.8-vllm0.13.0",
+    "deprecated": false
+  },
   {
     "backend": "cuda",
     "backend_version": "12.8",
@@ -857,6 +901,28 @@
     "docker_image": "gpustack/runner:cuda12.8-vllm0.10.0",
     "deprecated": false
   },
+  {
+    "backend": "cuda",
+    "backend_version": "12.8",
+    "original_backend_version": "12.8.1",
+    "backend_variant": "",
+    "service": "voxbox",
+    "service_version": "0.0.21",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:cuda12.8-voxbox0.0.21",
+    "deprecated": false
+  },
+  {
+    "backend": "cuda",
+    "backend_version": "12.8",
+    "original_backend_version": "12.8.1",
+    "backend_variant": "",
+    "service": "voxbox",
+    "service_version": "0.0.21",
+    "platform": "linux/arm64",
+    "docker_image": "gpustack/runner:cuda12.8-voxbox0.0.21",
+    "deprecated": false
+  },
   {
     "backend": "cuda",
     "backend_version": "12.8",
@@ -879,6 +945,28 @@
     "docker_image": "gpustack/runner:cuda12.8-voxbox0.0.20",
     "deprecated": false
   },
+  {
+    "backend": "cuda",
+    "backend_version": "12.6",
+    "original_backend_version": "12.6.3",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.13.0",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:cuda12.6-vllm0.13.0",
+    "deprecated": false
+  },
+  {
+    "backend": "cuda",
+    "backend_version": "12.6",
+    "original_backend_version": "12.6.3",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.13.0",
+    "platform": "linux/arm64",
+    "docker_image": "gpustack/runner:cuda12.6-vllm0.13.0",
+    "deprecated": false
+  },
   {
     "backend": "cuda",
     "backend_version": "12.6",
@@ -1011,6 +1099,28 @@
     "docker_image": "gpustack/runner:cuda12.6-vllm0.10.0",
     "deprecated": false
   },
+  {
+    "backend": "cuda",
+    "backend_version": "12.6",
+    "original_backend_version": "12.6.3",
+    "backend_variant": "",
+    "service": "voxbox",
+    "service_version": "0.0.21",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:cuda12.6-voxbox0.0.21",
+    "deprecated": false
+  },
+  {
+    "backend": "cuda",
+    "backend_version": "12.6",
+    "original_backend_version": "12.6.3",
+    "backend_variant": "",
+    "service": "voxbox",
+    "service_version": "0.0.21",
+    "platform": "linux/arm64",
+    "docker_image": "gpustack/runner:cuda12.6-voxbox0.0.21",
+    "deprecated": false
+  },
   {
     "backend": "cuda",
     "backend_version": "12.6",
@@ -1198,6 +1308,17 @@
     "docker_image": "gpustack/runner:rocm7.0-sglang0.5.6.post2",
     "deprecated": false
   },
+  {
+    "backend": "rocm",
+    "backend_version": "7.0",
+    "original_backend_version": "7.0.2",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.13.0",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:rocm7.0-vllm0.13.0",
+    "deprecated": false
+  },
   {
     "backend": "rocm",
     "backend_version": "7.0",
@@ -1253,6 +1374,17 @@
     "docker_image": "gpustack/runner:rocm6.4-sglang0.5.5.post3",
     "deprecated": false
   },
+  {
+    "backend": "rocm",
+    "backend_version": "6.4",
+    "original_backend_version": "6.4.4",
+    "backend_variant": "",
+    "service": "vllm",
+    "service_version": "0.13.0",
+    "platform": "linux/amd64",
+    "docker_image": "gpustack/runner:rocm6.4-vllm0.13.0",
+    "deprecated": false
+  },
   {
     "backend": "rocm",
     "backend_version": "6.4",

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/pack/cuda/Dockerfile

@@ -95,11 +95,11 @@ ARG CMAKE_MAX_JOBS
 ARG CUDA_VERSION=12.9.1
 ARG CUDA_ARCHS
 ARG VOXBOX_BASE_IMAGE=gpustack/runner:cuda${CUDA_VERSION}-python${PYTHON_VERSION}
-ARG VOXBOX_VERSION=0.0.20
+ARG VOXBOX_VERSION=0.0.21
 ARG VOXBOX_TORCH_VERSION=2.7.1
 ARG VOXBOX_TORCH_CUDA_VERSION=${CUDA_VERSION}
 ARG VLLM_BASE_IMAGE=gpustack/runner:cuda${CUDA_VERSION}-python${PYTHON_VERSION}
-ARG VLLM_VERSION=0.12.0
+ARG VLLM_VERSION=0.13.0
 ARG VLLM_TORCH_VERSION=2.9.0
 ARG VLLM_TORCH_CUDA_VERSION=${CUDA_VERSION}
 ARG VLLM_BUILD_BASE_IMAGE=gpustack/runner:cuda${VLLM_TORCH_CUDA_VERSION}-python${PYTHON_VERSION}
@@ -112,7 +112,7 @@ ARG VLLM_DEEPEP_COMMIT=b57e5e21
 ARG VLLM_DEEPGEMM_COMMIT=9b680f42
 ARG VLLM_FLASHINFER_VERSION=0.5.3
 ARG VLLM_FLASHATTENTION_VERSION=2.8.3
-ARG VLLM_LMCACHE_VERSION=0.3.10.post1
+ARG VLLM_LMCACHE_VERSION=0.3.11
 ARG VLLM_MOONCAKE_VERSION=0.3.7.post2
 ARG SGLANG_BASE_IMAGE=vllm
 ARG SGLANG_VERSION=0.5.6.post2
@@ -492,6 +492,7 @@ einops
 cuda-python==${CUDA_MAJOR}.${CUDA_MINOR}
 pynvml==${CUDA_MAJOR}
 nvidia-nvshmem-cu${CUDA_MAJOR}
+nvshmem4py-cu${CUDA_MAJOR}
 EOT
     uv pip install \
         -r /tmp/requirements.txt
@@ -575,6 +576,20 @@ RUN <<EOF
 
     IFS="." read -r TORCH_MAJOR TORCH_MINOR TORCH_PATCH <<< "${VLLM_TORCH_VERSION}"
     IFS="." read -r CUDA_MAJOR CUDA_MINOR CUDA_PATCH <<< "${VLLM_TORCH_CUDA_VERSION}"
+    IFS="." read -r PYTHON_MAJOR PYTHON_MINOR <<< "${PYTHON_VERSION}"
+
+    PYTHON_MAJOR_MINOR="${PYTHON_MAJOR}${PYTHON_MINOR}"
+    for ABI in FALSE TRUE; do
+        PREBUILD_URL="https://github.com/Dao-AILab/flash-attention/releases/download/v${VLLM_FLASHATTENTION_VERSION}/flash_attn-${VLLM_FLASHATTENTION_VERSION}+cu${CUDA_MAJOR}torch${TORCH_MAJOR}.${TORCH_MINOR}cxx11abi${ABI}-cp${PYTHON_MAJOR_MINOR}-cp${PYTHON_MAJOR_MINOR}-linux_$(uname -m).whl"
+        if curl --retry 3 --retry-connrefused -fsSIL "${PREBUILD_URL}" >/dev/null 2>&1; then
+            echo "Downloading prebuilt FlashAttention wheel from ${PREBUILD_URL}..."
+            curl --retry 3 --retry-connrefused -fL "${PREBUILD_URL}" -o "/tmp/flash_attn-${VLLM_FLASHATTENTION_VERSION}+cu${CUDA_MAJOR}torch${TORCH_MAJOR}.${TORCH_MINOR}cxx11abi${ABI}-cp${PYTHON_MAJOR_MINOR}-cp${PYTHON_MAJOR_MINOR}-linux_$(uname -m).whl"
+            mkdir -p /workspace \
+                && mv /tmp/*.whl /workspace \
+                && tree -hs /workspace
+            exit 0
+        fi
+    done
 
     # Support ARM64 only
     if [[ "${TARGETARCH}" != "amd64" ]]; then
@@ -582,16 +597,6 @@ RUN <<EOF
         exit 0
     fi
 
-    PREBUILD_URL="https://github.com/Dao-AILab/flash-attention/releases/download/v${VLLM_FLASHATTENTION_VERSION}/flash_attn-${VLLM_FLASHATTENTION_VERSION}+cu${CUDA_MAJOR}torch${TORCH_MAJOR}.${TORCH_MINOR}cxx11abiFALSE-cp310-cp310-linux_$(uname -m).whl"
-    if curl --retry 3 --retry-connrefused -fsSIL "${PREBUILD_URL}" >/dev/null 2>&1; then
-        echo "Downloading prebuilt FlashAttention wheel from ${PREBUILD_URL}..."
-        curl --retry 3 --retry-connrefused -fL "${PREBUILD_URL}" -o "/tmp/flash_attn-${VLLM_FLASHATTENTION_VERSION}+cu${CUDA_MAJOR}torch${TORCH_MAJOR}.${TORCH_MINOR}cxx11abiFALSE-cp310-cp310-linux_$(uname -m).whl"
-        mkdir -p /workspace \
-            && mv /tmp/*.whl /workspace \
-            && tree -hs /workspace
-        exit 0
-    fi
-
     # Download
     git -C /tmp clone --recursive --shallow-submodules \
         --depth 1 --branch v${VLLM_FLASHATTENTION_VERSION} --single-branch \
@@ -962,7 +967,7 @@ ARG VLLM_VERSION
 
 ENV VLLM_VERSION=${VLLM_VERSION}
 
-RUN <<EOF
+RUN --mount=type=bind,from=vllm-build-flashattention,source=/,target=/flashattention,rw <<EOF
     # vLLM
 
     IFS="." read -r CUDA_MAJOR CUDA_MINOR CUDA_PATCH <<< "${VLLM_TORCH_CUDA_VERSION}"
@@ -986,6 +991,8 @@ RUN <<EOF
     export TORCH_CUDA_ARCH_LIST="${VL_CUDA_ARCHS}"
     export COMPILE_CUSTOM_KERNELS=1
     export NVCC_THREADS=1
+    echo "Building vLLM with the following environment variables:"
+    env
 
     # Install
     git -C /tmp clone --recursive --shallow-submodules \
@@ -1047,6 +1054,9 @@ RUN --mount=type=bind,from=vllm-build-vllm,source=/,target=/vllm,rw <<EOF
     export MAX_JOBS="${CMAKE_MAX_JOBS}"
     export TORCH_CUDA_ARCH_LIST="${LC_CUDA_ARCHS}"
     export NVCC_THREADS=1
+    echo "Building LMCache with the following environment variables:"
+    env
+
     git -C /tmp clone --recursive --shallow-submodules \
         --depth 1 --branch v${VLLM_LMCACHE_VERSION} --single-branch \
         https://github.com/LMCache/LMCache.git lmcache

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/pack/matrix.yaml

@@ -102,7 +102,6 @@ rules:
   ##
   - backend: "cuda"
     services:
-      - "voxbox"
       - "vllm"
       - "sglang"
     args:

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/pack/rocm/Dockerfile

@@ -69,10 +69,10 @@
 #   which is used to build the SGLang from source.
 ARG PYTHON_VERSION=3.12
 ARG CMAKE_MAX_JOBS
-ARG ROCM_VERSION=7.1.1
+ARG ROCM_VERSION=7.0.2
 ARG ROCM_ARCHS
 ARG VLLM_BASE_IMAGE=gpustack/runner:rocm${ROCM_VERSION}-python${PYTHON_VERSION}
-ARG VLLM_VERSION=0.12.0
+ARG VLLM_VERSION=0.13.0
 ARG VLLM_TORCH_VERSION=2.9.1
 ARG VLLM_TORCH_ROCM_VERSION=${ROCM_VERSION}
 ARG VLLM_TORCH_SOURCE=pytorch
@@ -80,7 +80,7 @@ ARG VLLM_BUILD_BASE_IMAGE=gpustack/runner:rocm${VLLM_TORCH_ROCM_VERSION}-python$
 ARG VLLM_TRITON_COMMIT=57c693b6
 ARG VLLM_FLASHATTENTION_VERSION=2.8.3
 ARG VLLM_AITER_VERSION=0.1.7.post5
-ARG VLLM_LMCACHE_VERSION=0.3.10.post1
+ARG VLLM_LMCACHE_VERSION=0.3.11
 ARG VLLM_MOONCAKE_VERSION=0.3.7.post2
 ARG SGLANG_BASE_IMAGE=vllm
 ARG SGLANG_VERSION=0.5.6.post2
@@ -679,12 +679,12 @@ ARG VLLM_VERSION
 
 ENV VLLM_VERSION=${VLLM_VERSION}
 
-RUN --mount=type=bind,from=vllm-build-triton,source=/,target=/triton,rw \
-    --mount=type=bind,from=vllm-build-flashattention,source=/,target=/flashattention,rw \
+RUN --mount=type=bind,from=vllm-build-flashattention,source=/,target=/flashattention,rw \
     --mount=type=bind,from=vllm-build-aiter,source=/,target=/aiter,rw <<EOF
     # vLLM
 
     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
@@ -697,6 +697,14 @@ RUN --mount=type=bind,from=vllm-build-triton,source=/,target=/triton,rw \
     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}.${VL_MINOR} == 0.13" | bc -l) )); then
+                # TODO(thxCode): Temporarily remove gfx1030 for vLLM ROCm build due to build error in ROCm 6.4.4.
+                # #15 134.9 /tmp/vllm/build/temp.linux-x86_64-cpython-312/csrc/sampler.hip:564:63: error: local memory (66032) exceeds limit (65536) in 'void vllm::topKPerRowDecode<1024, true, false, true>(float const*, int const*, int*, int, int, int, int, float*, int, int const*)'
+                # ##15 134.9   564 | static __global__ __launch_bounds__(kNumThreadsPerBlock) void topKPerRowDecode(
+                # ##15 134.9       |                                                               ^
+                # ##15 134.9 16 warnings and 1 error generated when compiling for gfx1030.
+                VL_ROCM_ARCHS="gfx908;gfx90a;gfx942"
+            fi
         else
             VL_ROCM_ARCHS="gfx908;gfx90a;gfx942;gfx950;gfx1030;gfx1100;gfx1101;gfx1200;gfx1201;gfx1150;gfx1151"
         fi
@@ -704,6 +712,8 @@ RUN --mount=type=bind,from=vllm-build-triton,source=/,target=/triton,rw \
     export MAX_JOBS="${CMAKE_MAX_JOBS}"
     export COMPILE_CUSTOM_KERNELS=1
     export PYTORCH_ROCM_ARCH="${VL_ROCM_ARCHS}"
+    echo "Building vLLM with the following environment variables:"
+    env
 
     # Build
     git -C /tmp clone --recursive --shallow-submodules \
@@ -712,7 +722,9 @@ RUN --mount=type=bind,from=vllm-build-triton,source=/,target=/triton,rw \
     pushd /tmp/vllm \
         && sed -i "s/\"torch ==.*\"/\"torch\"/g" /tmp/vllm/pyproject.toml \
         && sed -i "s/\"torch==.*\"/\"torch\"/g" /tmp/vllm/requirements/rocm-build.txt \
+        && sed -i "s/\"torchvision==.*\"/\"torchvision\"/g" /tmp/vllm/requirements/rocm-build.txt \
         && sed -i "s/\"torchaudio==.*\"/\"torchaudio\"/g" /tmp/vllm/requirements/rocm-build.txt \
+        && sed -i "s/\"triton==.*\"/\"triton\"/g" /tmp/vllm/requirements/rocm-build.txt \
         && VLLM_TARGET_DEVICE="rocm" python -v -m build --no-isolation --wheel \
         && tree -hs /tmp/vllm/dist \
         && mv /tmp/vllm/dist /workspace
@@ -769,6 +781,8 @@ RUN --mount=type=bind,from=vllm-build-vllm,source=/,target=/vllm,rw <<EOF
     export TORCH_DONT_CHECK_COMPILER_ABI=1
     export CXX=hipcc
     export BUILD_WITH_HIP=1
+    echo "Building LMCache with the following environment variables:"
+    env
 
     # Install LMCache
     git -C /tmp clone --recursive --shallow-submodules \
@@ -1403,7 +1417,7 @@ RUN --mount=type=bind,target=/workspace,rw <<EOF
 
     tree -hs /workspace/patches
     pushd $(pip show sglang | grep Location: | cut -d" " -f 2) \
-        && patch -p1 < /workspace/patches/*.patch
+        && patch -p1 < /workspace/patches/sglang_*.patch
 EOF
 
 ## Entrypoint

{gpustack_runner-0.1.22.post2 → gpustack_runner-0.1.22.post4}/tests/gpustack_runner/fixtures/test_list_runners_by_backend.json

@@ -626,6 +626,28 @@
         "docker_image": "gpustack/runner:cuda12.9-sglang0.5.6.post2",
         "deprecated": false
       },
+      {
+        "backend": "cuda",
+        "backend_version": "12.9",
+        "original_backend_version": "12.9.1",
+        "backend_variant": "",
+        "service": "vllm",
+        "service_version": "0.13.0",
+        "platform": "linux/amd64",
+        "docker_image": "gpustack/runner:cuda12.9-vllm0.13.0",
+        "deprecated": false
+      },
+      {
+        "backend": "cuda",
+        "backend_version": "12.9",
+        "original_backend_version": "12.9.1",
+        "backend_variant": "",
+        "service": "vllm",
+        "service_version": "0.13.0",
+        "platform": "linux/arm64",
+        "docker_image": "gpustack/runner:cuda12.9-vllm0.13.0",
+        "deprecated": false
+      },
       {
         "backend": "cuda",
         "backend_version": "12.9",
@@ -747,6 +769,28 @@
         "docker_image": "gpustack/runner:cuda12.8-sglang0.5.4.post3",
         "deprecated": false
       },
+      {
+        "backend": "cuda",
+        "backend_version": "12.8",
+        "original_backend_version": "12.8.1",
+        "backend_variant": "",
+        "service": "vllm",
+        "service_version": "0.13.0",
+        "platform": "linux/amd64",
+        "docker_image": "gpustack/runner:cuda12.8-vllm0.13.0",
+        "deprecated": false
+      },
+      {
+        "backend": "cuda",
+        "backend_version": "12.8",
+        "original_backend_version": "12.8.1",
+        "backend_variant": "",
+        "service": "vllm",
+        "service_version": "0.13.0",
+        "platform": "linux/arm64",
+        "docker_image": "gpustack/runner:cuda12.8-vllm0.13.0",
+        "deprecated": false
+      },
       {
         "backend": "cuda",
         "backend_version": "12.8",
@@ -879,6 +923,28 @@
         "docker_image": "gpustack/runner:cuda12.8-vllm0.10.0",
         "deprecated": false
       },
+      {
+        "backend": "cuda",
+        "backend_version": "12.8",
+        "original_backend_version": "12.8.1",
+        "backend_variant": "",
+        "service": "voxbox",
+        "service_version": "0.0.21",
+        "platform": "linux/amd64",
+        "docker_image": "gpustack/runner:cuda12.8-voxbox0.0.21",
+        "deprecated": false
+      },
+      {
+        "backend": "cuda",
+        "backend_version": "12.8",
+        "original_backend_version": "12.8.1",
+        "backend_variant": "",
+        "service": "voxbox",
+        "service_version": "0.0.21",
+        "platform": "linux/arm64",
+        "docker_image": "gpustack/runner:cuda12.8-voxbox0.0.21",
+        "deprecated": false
+      },
       {
         "backend": "cuda",
         "backend_version": "12.8",
@@ -901,6 +967,28 @@
         "docker_image": "gpustack/runner:cuda12.8-voxbox0.0.20",
         "deprecated": false
       },
+      {
+        "backend": "cuda",
+        "backend_version": "12.6",
+        "original_backend_version": "12.6.3",
+        "backend_variant": "",
+        "service": "vllm",
+        "service_version": "0.13.0",
+        "platform": "linux/amd64",
+        "docker_image": "gpustack/runner:cuda12.6-vllm0.13.0",
+        "deprecated": false
+      },
+      {
+        "backend": "cuda",
+        "backend_version": "12.6",
+        "original_backend_version": "12.6.3",
+        "backend_variant": "",
+        "service": "vllm",
+        "service_version": "0.13.0",
+        "platform": "linux/arm64",
+        "docker_image": "gpustack/runner:cuda12.6-vllm0.13.0",
+        "deprecated": false
+      },
       {
         "backend": "cuda",
         "backend_version": "12.6",
@@ -1033,6 +1121,28 @@
         "docker_image": "gpustack/runner:cuda12.6-vllm0.10.0",
         "deprecated": false
       },
+      {
+        "backend": "cuda",
+        "backend_version": "12.6",
+        "original_backend_version": "12.6.3",
+        "backend_variant": "",
+        "service": "voxbox",
+        "service_version": "0.0.21",
+        "platform": "linux/amd64",
+        "docker_image": "gpustack/runner:cuda12.6-voxbox0.0.21",
+        "deprecated": false
+      },
+      {
+        "backend": "cuda",
+        "backend_version": "12.6",
+        "original_backend_version": "12.6.3",
+        "backend_variant": "",
+        "service": "voxbox",
+        "service_version": "0.0.21",
+        "platform": "linux/arm64",
+        "docker_image": "gpustack/runner:cuda12.6-voxbox0.0.21",
+        "deprecated": false
+      },
       {
         "backend": "cuda",
         "backend_version": "12.6",
@@ -1244,6 +1354,17 @@
         "docker_image": "gpustack/runner:rocm7.0-sglang0.5.6.post2",
         "deprecated": false
       },
+      {
+        "backend": "rocm",
+        "backend_version": "7.0",
+        "original_backend_version": "7.0.2",
+        "backend_variant": "",
+        "service": "vllm",
+        "service_version": "0.13.0",
+        "platform": "linux/amd64",
+        "docker_image": "gpustack/runner:rocm7.0-vllm0.13.0",
+        "deprecated": false
+      },
       {
         "backend": "rocm",
         "backend_version": "7.0",
@@ -1299,6 +1420,17 @@
         "docker_image": "gpustack/runner:rocm6.4-sglang0.5.5.post3",
         "deprecated": false
       },
+      {
+        "backend": "rocm",
+        "backend_version": "6.4",
+        "original_backend_version": "6.4.4",
+        "backend_variant": "",
+        "service": "vllm",
+        "service_version": "0.13.0",
+        "platform": "linux/amd64",
+        "docker_image": "gpustack/runner:rocm6.4-vllm0.13.0",
+        "deprecated": false
+      },
       {
         "backend": "rocm",
         "backend_version": "6.4",

gpustack_runner-0.1.22.post2/gpustack_runner/_version_appendix.py

@@ -1 +0,0 @@
-git_commit = "457b969"