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

[AMD][CI] Update v4 CI setting and move the task to main branch (#26662)

Browse Source
This commit is contained in:
Thomas Wang
2026-05-29 00:12:40 -07:00
committed by GitHub
parent b2eed9e16d
commit 6e9bd82714
5 changed files with 109 additions and 172 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+9 -64
View File
@@ -1162,52 +1162,20 @@ jobs:
with:
ref: ${{ inputs.ref || github.sha }}
- name: Resolve DSv4 image tag
id: dsv4_image
run: |
# Pick the latest Docker Hub tag matching rocm720-mi35x-<sha7>-<YYYYMMDD>-DSv4.
# Docker Hub returns results sorted by last_updated DESC by default, so the
# first regex match is the most recent daily build.
AUTH_HEADER=()
if [[ -n "${DOCKERHUB_AMD_USERNAME:-}" && -n "${DOCKERHUB_AMD_TOKEN:-}" ]]; then
TOKEN=$(curl -s -H "Content-Type: application/json" \
-X POST -d "{\"username\":\"${DOCKERHUB_AMD_USERNAME}\",\"password\":\"${DOCKERHUB_AMD_TOKEN}\"}" \
https://hub.docker.com/v2/users/login/ | python3 -c "import json,sys; print(json.load(sys.stdin).get('token',''))")
if [[ -n "$TOKEN" ]]; then
AUTH_HEADER=(-H "Authorization: JWT $TOKEN")
fi
fi
TAG=$(curl -s "${AUTH_HEADER[@]}" \
"https://hub.docker.com/v2/repositories/rocm/sgl-dev/tags?page_size=100&name=DSv4" \
| grep -oE '"name":"rocm720-mi35x-[a-f0-9]{7}-[0-9]{8}-DSv4"' \
| head -n 1 | cut -d'"' -f4)
if [ -z "$TAG" ]; then
echo "::error::No DSv4 image found matching rocm720-mi35x-<sha7>-<YYYYMMDD>-DSv4 on Docker Hub"
exit 1
fi
echo "image=rocm/sgl-dev:$TAG" >> "$GITHUB_OUTPUT"
echo "Resolved DSv4 image: rocm/sgl-dev:$TAG"
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Setup docker (ROCm 7.2 DSv4)
- name: Setup docker (ROCm 7.2)
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --custom-image ${{ steps.dsv4_image.outputs.image }}
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies (preserve DSv4 sglang/aiter from image)
- name: Install dependencies
run: |
# --skip-sglang-build: keep the image's pre-installed DSv4 sglang
# (default would `pip install -e /sglang-checkout/python` and clobber it with main's source).
# --skip-aiter-build: keep the image's DSv4-tuned aiter
# (default reads /sglang-checkout/docker/rocm.Dockerfile from main and rebuilds aiter to that commit).
# --skip-test-time-deps: GSM8K + bench_one_batch_server don't need lmms-eval / human-eval.
bash scripts/ci/amd/amd_ci_install_dependency.sh \
--skip-sglang-build --skip-aiter-build --skip-test-time-deps
# tabulate is the only thing run_suite.py imports that may not be in the DSv4 image.
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
- name: Accuracy + Performance Test MI35x ROCm 7.2 (8-GPU DeepSeek-V4-Flash FP8 + FP4)
@@ -1229,43 +1197,20 @@ jobs:
with:
ref: ${{ inputs.ref || github.sha }}
- name: Resolve DSv4 image tag
id: dsv4_image
run: |
AUTH_HEADER=()
if [[ -n "${DOCKERHUB_AMD_USERNAME:-}" && -n "${DOCKERHUB_AMD_TOKEN:-}" ]]; then
TOKEN=$(curl -s -H "Content-Type: application/json" \
-X POST -d "{\"username\":\"${DOCKERHUB_AMD_USERNAME}\",\"password\":\"${DOCKERHUB_AMD_TOKEN}\"}" \
https://hub.docker.com/v2/users/login/ | python3 -c "import json,sys; print(json.load(sys.stdin).get('token',''))")
if [[ -n "$TOKEN" ]]; then
AUTH_HEADER=(-H "Authorization: JWT $TOKEN")
fi
fi
TAG=$(curl -s "${AUTH_HEADER[@]}" \
"https://hub.docker.com/v2/repositories/rocm/sgl-dev/tags?page_size=100&name=DSv4" \
| grep -oE '"name":"rocm720-mi35x-[a-f0-9]{7}-[0-9]{8}-DSv4"' \
| head -n 1 | cut -d'"' -f4)
if [ -z "$TAG" ]; then
echo "::error::No DSv4 image found matching rocm720-mi35x-<sha7>-<YYYYMMDD>-DSv4 on Docker Hub"
exit 1
fi
echo "image=rocm/sgl-dev:$TAG" >> "$GITHUB_OUTPUT"
echo "Resolved DSv4 image: rocm/sgl-dev:$TAG"
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Setup docker (ROCm 7.2 DSv4)
- name: Setup docker (ROCm 7.2)
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --custom-image ${{ steps.dsv4_image.outputs.image }}
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies (preserve DSv4 sglang/aiter from image)
- name: Install dependencies
run: |
bash scripts/ci/amd/amd_ci_install_dependency.sh \
--skip-sglang-build --skip-aiter-build --skip-test-time-deps
# --skip-test-time-deps: GSM8K + bench_one_batch_server don't need lmms-eval / human-eval.
bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
bash scripts/ci/amd/amd_ci_exec.sh pip install tabulate
- name: Accuracy + Performance Test MI35x ROCm 7.2 (8-GPU DeepSeek-V4-Pro FP8 + FP4)