feat: bump sglang to `0.5.6.post2` and swap to upstream runtime container (#4762)

Signed-off-by: Dillon Cullinan <dcullinan@nvidia.com>
Signed-off-by: Dmitry Tokarev <dtokarev@nvidia.com>
Co-authored-by: Dillon Cullinan <dcullinan@nvidia.com>
Co-authored-by: Dmitry Tokarev <dtokarev@nvidia.com>

.github/actions/docker-build/action.yml

@@ -164,8 +164,10 @@ runs:
 
         # Run the sanity check script inside the container
         # The script is located in /workspace/deploy/sanity_check.py in runtime containers
+        export WORKSPACE=/workspace
+
         set +e
-        docker run --rm "$IMAGE_TAG" python /workspace/deploy/sanity_check.py --runtime-check --no-gpu-check
+        docker run --rm "$IMAGE_TAG" python ${WORKSPACE}/deploy/sanity_check.py --runtime-check --no-gpu-check
         SANITY_CHECK_EXIT_CODE=$?
         set -e
         if [ ${SANITY_CHECK_EXIT_CODE} -ne 0 ]; then

benchmarks/pyproject.toml

@@ -46,7 +46,7 @@ dependencies = [
     "pydantic>=2",
     "tabulate",
     "types-tabulate",
-    # Satisfies vLLM 0.11.0 (>=4.55.2), vLLM 0.11.2 (>=4.56.0,<5), TRT-LLM 1.2.0rc5 (==4.56.0), SGLang 0.5.6 (==4.57.1)
+    # Satisfies vLLM 0.11.0 (>=4.55.2), vLLM 0.11.2 (>=4.56.0,<5), TRT-LLM 1.2.0rc5 (==4.56.0), SGLang 0.5.6.post2 (==4.57.1)
     "transformers>=4.56.0,<=4.57.1",
     "pytest-mypy",
 ]

container/Dockerfile.sglang-wideep

-# SPDX-FileCopyrightText: Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
-# SPDX-License-Identifier: Apache-2.0
-
-ARG SGLANG_IMAGE_TAG="v0.5.3.post2"
-ARG BRANCH_TYPE
-ARG CARGO_BUILD_JOBS
-
-FROM scratch AS local_src
-COPY . /src
-
-FROM lmsysorg/sglang:${SGLANG_IMAGE_TAG}
-
-WORKDIR /sgl-workspace
-
-ARG DYNAMO_COMMIT_SHA
-ENV DYNAMO_COMMIT_SHA=$DYNAMO_COMMIT_SHA
-
-# Install jq for JSON processing
-RUN apt-get update -y \
-    && DEBIAN_FRONTEND=noninteractive apt-get install -y --no-install-recommends \
-        jq \
-    && apt-get clean \
-    && rm -rf /var/lib/apt/lists/*
-
-# Install dynamo
-# Providing --build-arg BRANCH_TYPE=local will editable install the local dynamo repo
-# Providing --build-arg BRANCH_TYPE=remote will editable install the remote dynamo repo
-# Default is to install the latest published dynamo version
-ARG BRANCH_TYPE
-ARG CARGO_BUILD_JOBS
-
-COPY --from=local_src /src /tmp/local_src
-RUN if [ "$BRANCH_TYPE" = "local" ]; then \
-        cp -r /tmp/local_src /sgl-workspace/dynamo; \
-    elif [ "$BRANCH_TYPE" = "remote" ]; then \
-        git clone https://github.com/ai-dynamo/dynamo.git /sgl-workspace/dynamo; \
-    fi
-
-ENV CARGO_BUILD_JOBS=${CARGO_BUILD_JOBS:-16}
-
-# SGLang does not use a venv in their container
-RUN if [ "$BRANCH_TYPE" = "local" ]; then \
-    cd dynamo/lib/bindings/python && \
-    pip install --break-system-packages maturin && \
-    maturin build --release && \
-    pip install --break-system-packages target/wheels/*.whl && \
-    cd /sgl-workspace/dynamo && \
-    pip install --break-system-packages -e . && \
-    pip install --break-system-packages --requirement /tmp/local_src/container/deps/requirements.txt ; \
-  elif [ "$BRANCH_TYPE" = "remote" ]; then \
-    cd dynamo/lib/bindings/python && \
-    pip install --break-system-packages maturin && \
-    maturin build --release && \
-    pip install --break-system-packages target/wheels/*.whl && \
-    cd /sgl-workspace/dynamo && \
-    pip install --break-system-packages -e . && \
-    pip install --break-system-packages --requirement /sgl-workspace/dynamo/container/deps/requirements.txt ; \
-  else \
-    pip install --break-system-packages ai-dynamo ; \
-  fi \
-&& rm -rf /tmp/local_src
-
-# Install NATS and ETCD
-RUN case "$(uname -m)" in \
-      x86_64) ARCH=amd64 ;; \
-      aarch64) ARCH=arm64 ;; \
-      *) echo "Unsupported architecture: $(uname -m)" && exit 1 ;; \
-    esac && \
-    wget --tries=3 --waitretry=5 \
-      https://github.com/nats-io/nats-server/releases/download/v2.10.28/\
-nats-server-v2.10.28-${ARCH}.deb && \
-    dpkg -i nats-server-v2.10.28-${ARCH}.deb && rm nats-server-v2.10.28-${ARCH}.deb
-
-ENV ETCD_VERSION="v3.5.21"
-RUN case "$(uname -m)" in \
-      x86_64) ARCH=amd64 ;; \
-      aarch64) ARCH=arm64 ;; \
-      *) echo "Unsupported architecture: $(uname -m)" && exit 1 ;; \
-    esac && \
-    wget --tries=3 --waitretry=5 \
-      https://github.com/etcd-io/etcd/releases/download/${ETCD_VERSION}/\
-etcd-${ETCD_VERSION}-linux-${ARCH}.tar.gz -O /tmp/etcd.tar.gz && \
-    mkdir -p /usr/local/bin/etcd && \
-    tar -xzf /tmp/etcd.tar.gz \
-        -C /usr/local/bin/etcd --strip-components=1 && \
-    rm /tmp/etcd.tar.gz
-
-ENV PATH=/usr/local/bin/etcd:$PATH
-
-# Enable forceful shutdown of inflight requests
-ENV SGL_FORCE_SHUTDOWN=1
-
-WORKDIR /sgl-workspace/dynamo/examples/backends/sglang

container/build.sh

@@ -115,13 +115,13 @@ VLLM_RUNTIME_IMAGE_TAG_CU13="13.0.2-runtime-ubuntu24.04"
 NONE_BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
 NONE_BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
 
-SGLANG_CUDA_VERSION="12.9.1"
-# This is for Dockerfile
+
 SGLANG_BASE_IMAGE="nvcr.io/nvidia/cuda-dl-base"
-SGLANG_BASE_IMAGE_TAG="25.01-cuda12.8-devel-ubuntu24.04"
-# This is for Dockerfile.sglang. Unlike the other frameworks, it is using a different base image
-SGLANG_FRAMEWORK_IMAGE="nvcr.io/nvidia/cuda"
-SGLANG_FRAMEWORK_IMAGE_TAG="${SGLANG_CUDA_VERSION}-cudnn-devel-ubuntu24.04"
+SGLANG_BASE_IMAGE_TAG="25.06-cuda12.9-devel-ubuntu24.04"
+SGLANG_CUDA_VERSION="12.9.1"
+SGLANG_PYTHON_VERSION="3.10"
+
+PYTHON_VERSION="3.12"
 
 NIXL_REF=0.8.0
 NIXL_UCX_REF=v1.20.0-rc1
@@ -909,13 +909,10 @@ fi
 
 if [[ $FRAMEWORK == "SGLANG" ]]; then
     echo "Customizing Python, CUDA, and framework images for sglang images"
-    BUILD_ARGS+=" --build-arg PYTHON_VERSION=3.10"
+    BUILD_ARGS+=" --build-arg PYTHON_VERSION=${SGLANG_PYTHON_VERSION}"
     BUILD_ARGS+=" --build-arg CUDA_VERSION=${SGLANG_CUDA_VERSION}"
-    # Unlike the other two frameworks, SGLang's framework image is different from the base image, so we need to set it explicitly.
-    BUILD_ARGS+=" --build-arg FRAMEWORK_IMAGE=${SGLANG_FRAMEWORK_IMAGE}"
-    BUILD_ARGS+=" --build-arg FRAMEWORK_IMAGE_TAG=${SGLANG_FRAMEWORK_IMAGE_TAG}"
 else
-    BUILD_ARGS+=" --build-arg PYTHON_VERSION=3.12"
+    BUILD_ARGS+=" --build-arg PYTHON_VERSION=${PYTHON_VERSION}"
 fi
 # Add sccache build arguments
 if [ "$USE_SCCACHE" = true ]; then

container/deps/requirements.txt

@@ -19,7 +19,7 @@ av==15.0.0
 fastapi==0.120.1
 ftfy==6.3.1
 genai-perf==0.0.15
-grpcio-tools<=1.66.0  # May have platform-specific builds
+grpcio-tools<=1.76.0  # May have platform-specific builds
 httpx==0.28.1
 kr8s==0.20.13
 kubernetes==32.0.1
@@ -38,7 +38,7 @@ pre-commit==4.5.0
 prometheus-api-client==0.6.0
 prometheus_client==0.23.1
 prophet==1.2.1
-protobuf==5.29.5
+protobuf>=5.29.5,<7.0.0
 pydantic>=2.11.4,<2.13  # vllm==0.12.0 depends on pydantic>=2.12.0
 pyright==1.1.407
 PyYAML==6.0.3
@@ -48,12 +48,12 @@ sentencepiece==0.2.1
 # Required by kr8s
 # https://github.com/kr8s-org/kr8s/blob/750022c3ebbb7988cddb5a979aca2ee8074a1069/examples/kubectl-ng/uv.lock#L988
 sniffio==1.3.1
-tensorboard==2.19.0
+tensorboard>=2.19.0,<2.21.0
 tensorboardX==2.6.2.2
 # Transformers version constraint for container builds
 # - vLLM 0.11.0: >=4.55.2, vLLM 0.11.2: >=4.56.0,<5
 # - TensorRT-LLM 1.2.0rc5: ==4.56.0
-# - SGLang 0.5.6: ==4.57.1
+# - SGLang 0.5.6.post2: ==4.57.1
 # Using >=4.56.0 and <=4.57.1 to satisfy all frameworks
 transformers>=4.56.0,<=4.57.1
 types-aiofiles==25.1.0.20251011

docs/reference/support-matrix.md

@@ -62,7 +62,7 @@ The following table shows the dependency versions included with each Dynamo rele
 
 | **Dependency** | **main (ToT)** | **v0.8.0 (unreleased)** | **v0.7.1** | **v0.7.0.post1** | **v0.7.0** |
 | :------------- | :------------- | :---------------------- | :--------- | :--------------- | :--------- |
-| SGLang         | 0.5.6          | 0.5.6.post1             | 0.5.3.post4| 0.5.3.post4      | 0.5.3.post4|
+| SGLang         | 0.5.6.post2    | 0.5.6.post2             | 0.5.3.post4| 0.5.3.post4      | 0.5.3.post4|
 | TensorRT-LLM   | 1.2.0rc5       | 1.2.0rc6                | 1.2.0rc3   | 1.2.0rc3         | 1.2.0rc2   |
 | vLLM           | 0.12.0         | 0.12.0                  | 0.11.0     | 0.11.0           | 0.11.0     |
 | NIXL           | 0.8.0          | 0.8.0                   | 0.8.0      | 0.8.0            | 0.8.0      |

pyproject.toml

@@ -61,8 +61,8 @@ vllm = [
 
 sglang = [
     "uvloop",
+    "sglang==0.5.6.post2",
     "nixl[cu12]<=0.8.0",
-    "sglang==0.5.6",
 ]
 
 [project.entry-points.pytest11]