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

[AMD] Fix deepseek-v4 mtp accept length issue (#28520)

Browse Source 
Co-authored-by: Bingxu Chen <bingxche@amd.com>
This commit is contained in:
Thomas Wang
2026-06-17 11:13:55 -07:00
committed by GitHub
co-authored by Bingxu Chen
parent 8aaca72c21
commit f5b041622b
3 changed files with 249 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+50
View File
@@ -68,6 +68,7 @@ on:
# 8-GPU DeepSeek-V4 (MI35x only)
- nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720
- nightly-8-gpu-mi35x-deepseek-v4-pro-rocm720
- nightly-8-gpu-mi35x-deepseek-v4-pro-mtp-rocm720
# 8-GPU Kimi-K2.6 (MI30x + MI35x)
- nightly-8-gpu-kimi-k26-rocm720
- nightly-8-gpu-mi35x-kimi-k26-rocm720
@@ -1321,6 +1322,54 @@ jobs:
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-deepseek-v4-pro --nightly --timeout-per-file 14400 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }}
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
nightly-8-gpu-mi35x-deepseek-v4-pro-mtp-rocm720:
if: (github.repository == 'sgl-project/sglang' || github.event_name == 'pull_request') && (!(inputs.job_filter || inputs.job_select) || (inputs.job_filter || inputs.job_select) == 'all' || contains(format(',{0},', inputs.job_filter || inputs.job_select), ',nightly-8-gpu-mi35x-deepseek-v4-pro-mtp-rocm720,'))
runs-on: linux-mi35x-gpu-8
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Setup docker (ROCm 7.2)
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies
run: |
# --skip-test-time-deps: GSM8K + send_one 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 + MTP Accept Length Test MI35x ROCm 7.2 (8-GPU DeepSeek-V4-Pro FP8 MTP, unified_kv_triton)
timeout-minutes: 240
run: |
> github_summary.md # Clear summary file
echo "## SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton" >> github_summary.md
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_HACK_FLASHMLA_BACKEND=unified_kv_triton \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-deepseek-v4-pro-mtp --nightly --timeout-per-file 7200 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }}
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
- name: Accuracy + MTP Accept Length Test MI35x ROCm 7.2 (8-GPU DeepSeek-V4-Pro FP8 MTP, triton)
if: ${{ !cancelled() }}
timeout-minutes: 240
run: |
> github_summary.md # Clear summary file
echo "## SGLANG_HACK_FLASHMLA_BACKEND=triton" >> github_summary.md
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_HACK_FLASHMLA_BACKEND=triton \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-mi35x-deepseek-v4-pro-mtp --nightly --timeout-per-file 7200 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }}
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
# ==============================================================================
# 8-GPU Kimi-K2.6 (MI30x + MI35x)
# ==============================================================================
@@ -1879,6 +1928,7 @@ jobs:
# 8-GPU DeepSeek-V4 (MI35x only)
- nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720
- nightly-8-gpu-mi35x-deepseek-v4-pro-rocm720
- nightly-8-gpu-mi35x-deepseek-v4-pro-mtp-rocm720
# 8-GPU Kimi-K2.6 (MI30x + MI35x)
- nightly-8-gpu-kimi-k26-rocm720
- nightly-8-gpu-mi35x-kimi-k26-rocm720