chenchenghao/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[AMD] rocm 7.2 image release, PR test, Nightly Test (#17799)

Browse Source 
Co-authored-by: Alan Kao <akao@amd.com>
Co-authored-by: bingxche <Bingxu.Chen@amd.com>
Co-authored-by: Michael <13900043+michaelzhang-ai@users.noreply.github.com>
This commit is contained in:
YC Tseng
2026-02-12 13:29:25 +08:00
committed by GitHub
parent 93ede0db19
commit 20554a0a4f
25 changed files with 2716 additions and 153 deletions
Download Patch File Download Diff File Expand all files Collapse all files

192
scripts/ci/amd/amd_ci_install_dependency.sh
View File

@@ -2,6 +2,27 @@
set -euo pipefail
HOSTNAME_VALUE=$(hostname)
GPU_ARCH="mi30x" # default
SKIP_TT_DEPS=""
SKIP_SGLANG_BUILD=""
SKIP_AITER_BUILD=""
while [[ $# -gt 0 ]]; do
case $1 in
--skip-aiter-build) SKIP_AITER_BUILD="1"; shift;;
--skip-sglang-build) SKIP_SGLANG_BUILD="1"; shift;;
--skip-test-time-deps) SKIP_TT_DEPS="1"; shift;;
-h|--help)
echo "Usage: $0 [OPTIONS] [OPTIONAL_DEPS]"
echo "Options:"
echo " --skip-sglang-build Don't build checkout sglang, use what was shipped with the image"
echo " --skip-aiter-build Don't build aiter, use what was shipped with the image"
echo " --skip-test-time-deps Don't build miscellaneous dependencies"
exit 0
;;
*) break ;;
esac
done
OPTIONAL_DEPS="${1:-}"
# Build python extras
@@ -23,15 +44,6 @@ fi
# Fix permissions on pip cache, ignore errors from concurrent access or missing temp files
docker exec ci_sglang chown -R root:root /sgl-data/pip-cache 2>/dev/null || true
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache --upgrade pip
docker exec ci_sglang pip uninstall sgl-kernel -y || true
docker exec ci_sglang pip uninstall sglang -y || true
# Clear Python cache to ensure latest code is used
docker exec ci_sglang find /opt/venv -name "*.pyc" -delete || true
docker exec ci_sglang find /opt/venv -name "__pycache__" -type d -exec rm -rf {} + || true
# Also clear cache in sglang-checkout
docker exec ci_sglang find /sglang-checkout -name "*.pyc" -delete || true
docker exec ci_sglang find /sglang-checkout -name "__pycache__" -type d -exec rm -rf {} + || true
docker exec -w /sglang-checkout/sgl-kernel ci_sglang bash -c "rm -f pyproject.toml && mv pyproject_rocm.toml pyproject.toml && python3 setup_rocm.py install"
# Helper function to install with retries and fallback PyPI mirror
install_with_retry() {
@@ -93,75 +105,82 @@ git_clone_with_retry() {
return 1
}
# Install checkout sglang
if [ -n "$SKIP_SGLANG_BUILD" ]; then
echo "Didn't build checkout SGLang"
else
docker exec ci_sglang pip uninstall sgl-kernel -y || true
docker exec ci_sglang pip uninstall sglang -y || true
# Clear Python cache to ensure latest code is used
docker exec ci_sglang find /opt/venv -name "*.pyc" -delete || true
docker exec ci_sglang find /opt/venv -name "__pycache__" -type d -exec rm -rf {} + || true
# Also clear cache in sglang-checkout
docker exec ci_sglang find /sglang-checkout -name "*.pyc" -delete || true
docker exec ci_sglang find /sglang-checkout -name "__pycache__" -type d -exec rm -rf {} + || true
docker exec -w /sglang-checkout/sgl-kernel ci_sglang bash -c "rm -f pyproject.toml && mv pyproject_rocm.toml pyproject.toml && python3 setup_rocm.py install"
docker exec ci_sglang bash -c 'rm -rf python/pyproject.toml && mv python/pyproject_other.toml python/pyproject.toml'
install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e "python[${EXTRAS}]"
fi
case "${GPU_ARCH}" in
mi35x)
echo "Runner uses ${GPU_ARCH}; will fetch mi35x image."
docker exec ci_sglang rm -rf python/pyproject.toml && mv python/pyproject_other.toml python/pyproject.toml
# Follow the same dependency installation flow as mi30x/mi300/mi325.
install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e "python[${EXTRAS}]"
# For lmms_evals evaluating MMMU
docker exec -w / ci_sglang git clone --branch v0.4.1 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
install_with_retry docker exec -w /lmms-eval ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e .
;;
mi30x|mi300|mi325)
echo "Runner uses ${GPU_ARCH}; will fetch mi30x image."
docker exec ci_sglang rm -rf python/pyproject.toml && mv python/pyproject_other.toml python/pyproject.toml
install_with_retry docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e "python[${EXTRAS}]"
# For lmms_evals evaluating MMMU
docker exec -w / ci_sglang git clone --branch v0.4.1 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
install_with_retry docker exec -w /lmms-eval ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e .
;;
*)
echo "Runner architecture '${GPU_ARCH}' unrecognised;" >&2
;;
esac
if [[ -n "${SKIP_TT_DEPS}" ]]; then
echo "Didn't build lmms_eval, human-eval, and others"
else
# For lmms_evals evaluating MMMU
docker exec -w / ci_sglang git clone --branch v0.4.1 --depth 1 https://github.com/EvolvingLMMs-Lab/lmms-eval.git
install_with_retry docker exec -w /lmms-eval ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e .
#docker exec -w / ci_sglang git clone https://github.com/merrymercy/human-eval.git
git_clone_with_retry https://github.com/merrymercy/human-eval.git human-eval
docker cp human-eval ci_sglang:/
# Ensure setuptools is installed (human-eval's setup.py imports pkg_resources)
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache setuptools
install_with_retry docker exec -w /human-eval ci_sglang pip install --cache-dir=/sgl-data/pip-cache --no-build-isolation -e .
git_clone_with_retry https://github.com/akao-amd/human-eval.git human-eval
docker cp human-eval ci_sglang:/
install_with_retry docker exec -w /human-eval ci_sglang pip install --cache-dir=/sgl-data/pip-cache -e .
docker exec -w / ci_sglang mkdir -p /dummy-grok
# Create dummy grok config inline (bypasses Azure blob storage which may have auth issues)
mkdir -p dummy-grok
cat > dummy-grok/config.json << 'EOF'
{
"architectures": [
"Grok1ModelForCausalLM"
],
"embedding_multiplier_scale": 78.38367176906169,
"output_multiplier_scale": 0.5773502691896257,
"vocab_size": 131072,
"hidden_size": 6144,
"intermediate_size": 32768,
"max_position_embeddings": 8192,
"num_experts_per_tok": 2,
"num_local_experts": 8,
"num_attention_heads": 48,
"num_hidden_layers": 64,
"num_key_value_heads": 8,
"head_dim": 128,
"rms_norm_eps": 1e-05,
"rope_theta": 10000.0,
"model_type": "mixtral",
"torch_dtype": "bfloat16"
}
docker exec -w / ci_sglang mkdir -p /dummy-grok
# Create dummy grok config inline (bypasses Azure blob storage which may have auth issues)
mkdir -p dummy-grok
cat > dummy-grok/config.json << 'EOF'
{
"architectures": [
"Grok1ModelForCausalLM"
],
"embedding_multiplier_scale": 78.38367176906169,
"output_multiplier_scale": 0.5773502691896257,
"vocab_size": 131072,
"hidden_size": 6144,
"intermediate_size": 32768,
"max_position_embeddings": 8192,
"num_experts_per_tok": 2,
"num_local_experts": 8,
"num_attention_heads": 48,
"num_hidden_layers": 64,
"num_key_value_heads": 8,
"head_dim": 128,
"rms_norm_eps": 1e-05,
"rope_theta": 10000.0,
"model_type": "mixtral",
"torch_dtype": "bfloat16"
}
EOF
docker cp ./dummy-grok ci_sglang:/
# docker exec -w / ci_sglang mkdir -p /dummy-grok
# mkdir -p dummy-grok && wget https://sharkpublic.blob.core.windows.net/sharkpublic/sglang/dummy_grok.json -O dummy-grok/config.json
# docker cp ./dummy-grok ci_sglang:/
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache huggingface_hub[hf_xet]
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache pytest
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache huggingface_hub[hf_xet]
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache pytest
# Install tvm-ffi for JIT kernel support (QK-norm, etc.)
echo "Installing tvm-ffi for JIT kernel support..."
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache git+https://github.com/apache/tvm-ffi.git || echo "tvm-ffi installation failed, JIT kernels will use fallback"
# Install tvm-ffi for JIT kernel support (QK-norm, etc.)
echo "Installing tvm-ffi for JIT kernel support..."
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache git+https://github.com/apache/tvm-ffi.git || echo "tvm-ffi installation failed, JIT kernels will use fallback"
# Install cache-dit for qwen_image_t2i_cache_dit_enabled test (added in PR 16204)
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache cache-dit || echo "cache-dit installation failed"
# Install cache-dit for qwen_image_t2i_cache_dit_enabled test (added in PR 16204)
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache cache-dit || echo "cache-dit installation failed"
# Install accelerate for distributed training and inference support
docker exec ci_sglang pip install --cache-dir=/sgl-data/pip-cache accelerate || echo "accelerate installation failed"
fi
if [[ -n "${SKIP_AITER_BUILD}" ]]; then
exit 0
fi
# Detect AITER version
#############################################
@@ -215,16 +234,17 @@ echo "[CI-AITER-CHECK] AITER version inside CI image: ${IMAGE_AITER_VERSION}"
#############################################
NEED_REBUILD="false"
if [[ "${IMAGE_AITER_VERSION}" == "none" ]]; then
echo "[CI-AITER-CHECK] No AITER found in image"
NEED_REBUILD="true"
elif [[ "${IMAGE_AITER_VERSION}" != "${REPO_AITER_COMMIT}" ]]; then
echo "[CI-AITER-CHECK] Version mismatch:"
echo " Image: ${IMAGE_AITER_VERSION}"
echo " Repo : ${REPO_AITER_COMMIT}"
if [[ "${IMAGE_AITER_VERSION}" == "vnone" || "${IMAGE_AITER_VERSION}" == "v" ]]; then
echo "[CI-AITER-CHECK] No AITER found in image → rebuild needed"
NEED_REBUILD="true"
elif [[ "${IMAGE_AITER_VERSION}" == "${REPO_AITER_COMMIT}" ]]; then
echo "[CI-AITER-CHECK] AITER version matches"
elif [[ "${IMAGE_AITER_VERSION}" =~ (dev|\+g[0-9a-f]+) ]]; then
# Dev/patched version (contains 'dev' or git hash) → preserve it
echo "[CI-AITER-CHECK] Dev/patched version detected: ${IMAGE_AITER_VERSION} → skipping rebuild"
else
echo "[CI-AITER-CHECK] AITER version matches → using image's version."
echo "[CI-AITER-CHECK] Version mismatch: image=${IMAGE_AITER_VERSION}, repo=${REPO_AITER_COMMIT}"
NEED_REBUILD="true"
fi
@@ -270,12 +290,12 @@ fi
echo "[CI-AITER-CHECK] === AITER VERSION CHECK END ==="
# Clear pre-built AITER kernels from Docker image to avoid segfaults
# The Docker image may contain pre-compiled kernels incompatible with the current environment
echo "Clearing pre-built AITER kernels from Docker image..."
docker exec ci_sglang find /sgl-workspace/aiter/aiter/jit -name "*.so" -delete 2>/dev/null || true
docker exec ci_sglang ls -la /sgl-workspace/aiter/aiter/jit/ 2>/dev/null || echo "jit dir empty or not found"
# # Clear pre-built AITER kernels from Docker image to avoid segfaults
# # The Docker image may contain pre-compiled kernels incompatible with the current environment
# echo "Clearing pre-built AITER kernels from Docker image..."
# docker exec ci_sglang find /sgl-workspace/aiter/aiter/jit -name "*.so" -delete 2>/dev/null || true
# docker exec ci_sglang ls -la /sgl-workspace/aiter/aiter/jit/ 2>/dev/null || echo "jit dir empty or not found"
# Pre-build AITER kernels to avoid timeout during tests
echo "Warming up AITER JIT kernels..."
docker exec -e SGLANG_USE_AITER=1 ci_sglang python3 /sglang-checkout/scripts/ci/amd/amd_ci_warmup_aiter.py || echo "AITER warmup completed (some kernels may not be available)"
# # Pre-build AITER kernels to avoid timeout during tests
# echo "Warming up AITER JIT kernels..."
# docker exec -e SGLANG_USE_AITER=1 ci_sglang python3 /sglang-checkout/scripts/ci/amd/amd_ci_warmup_aiter.py || echo "AITER warmup completed (some kernels may not be available)"

99
scripts/ci/amd/amd_ci_start_container.sh
View File

@@ -27,13 +27,32 @@ DEFAULT_MI35X_BASE_TAG="${SGLANG_VERSION}-${ROCM_VERSION}-mi35x"
# Parse command line arguments
MI30X_BASE_TAG="${DEFAULT_MI30X_BASE_TAG}"
MI35X_BASE_TAG="${DEFAULT_MI35X_BASE_TAG}"
CUSTOM_IMAGE=""
BUILD_FROM_DOCKERFILE=""
GPU_ARCH_BUILD=""
while [[ $# -gt 0 ]]; do
case $1 in
--mi30x-base-tag) MI30X_BASE_TAG="$2"; shift 2;;
--mi35x-base-tag) MI35X_BASE_TAG="$2"; shift 2;;
--custom-image) CUSTOM_IMAGE="$2"; shift 2;;
--build-from-dockerfile) BUILD_FROM_DOCKERFILE="1"; shift;;
--gpu-arch) GPU_ARCH_BUILD="$2"; shift 2;;
--rocm-version)
ROCM_VERSION="$2"
MI30X_BASE_TAG="${SGLANG_VERSION}-${ROCM_VERSION}-mi30x"
MI35X_BASE_TAG="${SGLANG_VERSION}-${ROCM_VERSION}-mi35x"
echo "Using ROCm version override: ${ROCM_VERSION}"
shift 2;;
-h|--help)
echo "Usage: $0 [--mi30x-base-tag TAG] [--mi35x-base-tag TAG]"
echo "Usage: $0 [OPTIONS]"
echo "Options:"
echo " --mi30x-base-tag TAG Override MI30x base image tag"
echo " --mi35x-base-tag TAG Override MI35x base image tag"
echo " --custom-image IMAGE Use a specific Docker image directly"
echo " --build-from-dockerfile Build image from docker/rocm.Dockerfile"
echo " --gpu-arch ARCH GPU architecture for Dockerfile build (e.g., gfx950-rocm720)"
echo " --rocm-version VERSION Override ROCm version for image lookup (e.g., rocm720)"
exit 0
;;
*) echo "Unknown option $1"; exit 1;;
@@ -54,7 +73,7 @@ else
echo "Warning: could not parse GPU architecture from '${HOSTNAME_VALUE}', defaulting to ${GPU_ARCH}"
fi
# Normalise / collapse architectures we don’t yet build specifically for
# Normalise / collapse architectures we don't yet build specifically for
case "${GPU_ARCH}" in
mi35x)
echo "Runner uses ${GPU_ARCH}; will fetch mi35x image."
@@ -134,18 +153,73 @@ find_latest_image() {
fi
echo "Error: no ${gpu_arch} image found in the last 7 days for base ${base_tag}" >&2
echo "Using hard-coded fallback…" >&2
if [[ "${gpu_arch}" == "mi35x" ]]; then
echo "rocm/sgl-dev:v0.5.5-rocm700-mi35x-20251110"
else
echo "rocm/sgl-dev:v0.5.5-rocm700-mi30x-20251110"
fi
echo "Using hard-coded fallback for ${ROCM_VERSION}…" >&2
case "${ROCM_VERSION}" in
rocm720)
if [[ "${gpu_arch}" == "mi35x" ]]; then
echo "rocm/sgl-dev:v0.5.8.post1-rocm720-mi35x-20260211-preview"
else
echo "rocm/sgl-dev:v0.5.8.post1-rocm720-mi30x-20260211-preview"
fi
;;
rocm700)
if [[ "${gpu_arch}" == "mi35x" ]]; then
echo "rocm/sgl-dev:v0.5.8.post1-rocm700-mi35x-20260211"
else
echo "rocm/sgl-dev:v0.5.8.post1-rocm700-mi30x-20260211"
fi
;;
*)
echo "Error: no hard-coded fallback available for ${ROCM_VERSION}" >&2
return 1
;;
esac
}
# Pull and run the latest image
IMAGE=$(find_latest_image "${GPU_ARCH}")
echo "Pulling Docker image: ${IMAGE}"
docker pull "${IMAGE}"
# Determine which image to use
if [[ -n "${CUSTOM_IMAGE}" ]]; then
# Use explicitly provided custom image
IMAGE="${CUSTOM_IMAGE}"
echo "Using custom image: ${IMAGE}"
docker pull "${IMAGE}"
elif [[ -n "${BUILD_FROM_DOCKERFILE}" ]]; then
# Build image from Dockerfile
if [[ -z "${GPU_ARCH_BUILD}" ]]; then
echo "Error: --gpu-arch is required when using --build-from-dockerfile" >&2
exit 1
fi
DOCKERFILE_DIR="${GITHUB_WORKSPACE:-$PWD}/docker"
# Use rocm720.Dockerfile for ROCm 7.2 builds, otherwise use rocm.Dockerfile
if [[ "${GPU_ARCH_BUILD}" == *"rocm720"* ]]; then
DOCKERFILE="${DOCKERFILE_DIR}/rocm720.Dockerfile"
else
DOCKERFILE="${DOCKERFILE_DIR}/rocm.Dockerfile"
fi
if [[ ! -f "${DOCKERFILE}" ]]; then
echo "Error: Dockerfile not found at ${DOCKERFILE}" >&2
exit 1
fi
IMAGE="sglang-ci:${GPU_ARCH_BUILD}-$(date +%Y%m%d)"
echo "Building Docker image from ${DOCKERFILE} with GPU_ARCH=${GPU_ARCH_BUILD}..."
# Pass full GPU_ARCH (e.g., gfx950-rocm720) - Dockerfile handles stripping suffix
docker build \
--build-arg GPU_ARCH="${GPU_ARCH_BUILD}" \
--build-arg SGL_BRANCH="main" \
-t "${IMAGE}" \
-f "${DOCKERFILE}" \
"${DOCKERFILE_DIR}"
echo "Successfully built image: ${IMAGE}"
else
# Find the latest pre-built image
IMAGE=$(find_latest_image "${GPU_ARCH}")
echo "Pulling Docker image: ${IMAGE}"
docker pull "${IMAGE}"
fi
CACHE_HOST=/home/runner/sgl-data
if [[ -d "$CACHE_HOST" ]]; then
@@ -156,6 +230,7 @@ fi
echo "Launching container: ci_sglang"
docker run -dt --user root --device=/dev/kfd ${DEVICE_FLAG} \
--ulimit nofile=65536:65536 \
-v "${GITHUB_WORKSPACE:-$PWD}:/sglang-checkout" \
$CACHE_VOLUME \
--group-add video \