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

fix(test): stabilize nightly precision regression (#34668)

Browse Source 
Co-authored-by: Alison Shao <54658187+alisonshao@users.noreply.github.com>
Co-authored-by: Alison Shao <a.shao@wustl.edu>
This commit is contained in:
Xinyuan Tong
2026-08-25 20:04:52 -07:00
committed by GitHub
co-authored by Alison Shao Alison Shao
parent 2e4773aadd
commit 34de1fb47f
6 changed files with 271 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/rerun-test.yml
+24 -4
View File
@@ -56,6 +56,11 @@ on:
required: false
type: boolean
default: false
refresh_precision_baseline:
description: "Force-refresh the nightly precision baseline. Restricted by the slash-command handler to the precision test on trusted in-repo PRs."
required: false
type: boolean
default: false
# Mirrors _pr-test-stage.yml's env, so a rerun validates what CI actually ran.
env:
@@ -68,9 +73,11 @@ env:
HF_HUB_ETAG_TIMEOUT: 300
SGLANG_JIT_KERNEL_RUN_FULL_TESTS: ${{ inputs.full_jit_kernel_tests && '1' || '0' }}
IS_H200: ${{ inputs.runs_on == '8-gpu-h200' && '1' || '0' }}
# SGLANG_PRECISION_* stays out: that test pushes to the shared baseline on
# every run, so a rerun holding the write token could become everyone's next
# comparison baseline. Without the repo var it fails fast instead.
SGLANG_PRECISION_HF_REPO: ${{ vars.SGLANG_PRECISION_HF_REPO }}
SGLANG_PRECISION_HF_REVISION: ${{ vars.SGLANG_PRECISION_HF_REVISION || 'main' }}
SGLANG_PRECISION_HF_READ_ONLY: ${{ inputs.refresh_precision_baseline && '0' || '1' }}
SGLANG_PRECISION_FORCE_UPDATE: ${{ inputs.refresh_precision_baseline && '1' || '0' }}
SGLANG_PRECISION_COMMIT: ${{ inputs.pr_head_sha || github.sha }}
# Every job below sets its own `permissions`, which replaces rather than merges
# with a workflow-level block -- so keep the floor here minimal and grant per job.
@@ -138,7 +145,20 @@ jobs:
- name: Run test
timeout-minutes: 60
env:
TEST_COMMAND: ${{ inputs.test_command }}
SGLANG_PRECISION_HF_TOKEN: ${{ inputs.refresh_precision_baseline && inputs.mode == 'cuda' && inputs.runs_on == '8-gpu-h200' && inputs.test_command == 'registered/debug_utils/test_nightly_precision_regression.py' && inputs.pr_head_sha == '' && secrets.HF_TOKEN_PRECISION_STORE || '' }}
run: |
if [[ "${{ inputs.refresh_precision_baseline }}" == "true" ]]; then
expected="registered/debug_utils/test_nightly_precision_regression.py"
if [[ "${{ inputs.mode }}" != "cuda" \
|| "${{ inputs.runs_on }}" != "8-gpu-h200" \
|| -n "${{ inputs.pr_head_sha }}" \
|| "$TEST_COMMAND" != "$expected" ]]; then
echo "::error::Precision baseline refresh only accepts $expected on 8-gpu-h200"
exit 1
fi
fi
if [[ "${{ inputs.runs_on }}" == "1-gpu-5090" ]]; then
source /etc/profile.d/sglang-ci.sh
fi
@@ -146,7 +166,7 @@ jobs:
while IFS= read -r cmd; do
[ -z "$cmd" ] && continue
cmds+=("$cmd")
done <<< "${{ inputs.test_command }}"
done <<< "$TEST_COMMAND"
total=${#cmds[@]}
suite_start=$SECONDS
for idx in "${!cmds[@]}"; do