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

[AMD][CI] Add DeepSeek-V4-Flash FP8 accuracy coverage on MI30x (#36396)

Browse Source
This commit is contained in:
Michael
2026-08-26 19:30:07 -07:00
committed by GitHub
parent 07570202f0
commit 1c8f2b38cb
3 changed files with 201 additions and 99 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+61 -99
View File
@@ -59,11 +59,9 @@ on:
# 8-GPU Grok2 (MI30x + MI35x)
- nightly-8-gpu-grok2-rocm720
- nightly-8-gpu-mi35x-grok2-rocm720
# 8-GPU DeepSeek-V3.x (MI30x)
- nightly-8-gpu-deepseek-v31-rocm720
# 8-GPU DeepSeek-V3.2 (MI30x)
- nightly-8-gpu-deepseek-v32-rocm720
- nightly-8-gpu-deepseek-v32-mtp-rocm720
- nightly-8-gpu-deepseek-v3-kv-fp8-rocm720
# 8-GPU DeepSeek-V3.2 (MI35x)
- nightly-accuracy-8-gpu-mi35x-deepseek-v32-rocm720
- nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp-rocm720
@@ -75,7 +73,9 @@ on:
- nightly-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion-rocm720
- nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4-rocm720
- nightly-8-gpu-mi35x-deepseek-r1-hicache-rocm720
# 8-GPU DeepSeek-V4 (MI35x only)
# 8-GPU DeepSeek-V4-Flash (MI30x)
- nightly-8-gpu-deepseek-v4-flash-rocm720
# 8-GPU DeepSeek-V4 (MI35x)
- 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
@@ -856,60 +856,13 @@ jobs:
exit ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 8-GPU DeepSeek-V3.x (MI30x)
# 8-GPU DeepSeek-V3.2 (MI30x)
#
# V3-0324 and V3.1 are no longer scheduled here: V3.2 covers the same MI30x
# paths (aiter MLA, EAGLE MTP, multithread weight load) on a current
# checkpoint, and DeepSeek-R1/V4 carry the MI35x side.
# ==============================================================================
nightly-8-gpu-deepseek-v31-rocm720:
name: ${{ format('nightly-8-gpu-deepseek-v31 ({0}, linux-mi300-8gpu-sglang)', matrix.rocm_version) }}
strategy:
fail-fast: false
matrix:
rocm_version: ${{ fromJson(inputs.rocm_version && inputs.rocm_version != 'all' && format('["{0}"]', inputs.rocm_version) || '["rocm724", "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-deepseek-v31-rocm720,'))
runs-on: linux-mi300-8gpu-sglang
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Setup docker (${{ matrix.rocm_version }})
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ matrix.rocm_version }}
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
ENABLE_CACHE_HOST: "1"
- name: Install dependencies
run: bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
- name: Accuracy Test ROCm 7.2 (8-GPU DeepSeek-V3.1)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_AITER=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-deepseek-v31 --nightly --timeout-per-file 3600 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
- name: Performance Test ROCm 7.2 (8-GPU DeepSeek-V3.1)
timeout-minutes: 300
continue-on-error: true
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_USE_ROCM700A=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-perf-8-gpu-deepseek-v31 --nightly --timeout-per-file 18000 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
nightly-8-gpu-deepseek-v32-rocm720:
name: ${{ format('nightly-8-gpu-deepseek-v32 ({0}, linux-mi300-8gpu-sglang)', matrix.rocm_version) }}
strategy:
@@ -1008,44 +961,6 @@ jobs:
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
nightly-8-gpu-deepseek-v3-kv-fp8-rocm720:
name: ${{ format('nightly-8-gpu-deepseek-v3-kv-fp8 ({0}, linux-mi300-8gpu-sglang)', matrix.rocm_version) }}
strategy:
fail-fast: false
matrix:
rocm_version: ${{ fromJson(inputs.rocm_version && inputs.rocm_version != 'all' && format('["{0}"]', inputs.rocm_version) || '["rocm724", "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-deepseek-v3-kv-fp8-rocm720,'))
runs-on: linux-mi300-8gpu-sglang
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Setup docker (${{ matrix.rocm_version }})
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ matrix.rocm_version }}
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
ENABLE_CACHE_HOST: "1"
- name: Install dependencies
run: bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
- name: DeepSeek-V3 KV FP8 Test ROCm 7.2 (8-GPU Basic + MTP)
timeout-minutes: 120
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-8-gpu-deepseek-v3-kv-fp8 --nightly --timeout-per-file 3600 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 8-GPU DeepSeek-V3.2 (MI35x)
# ==============================================================================
@@ -1465,7 +1380,54 @@ jobs:
exit ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 8-GPU DeepSeek-V4 (MI35x only)
# 8-GPU DeepSeek-V4-Flash (MI30x)
#
# Accuracy only: the MI35x jobs below carry the DSV4 perf numbers, and gfx942
# has no DSV4 perf baseline to regress against yet. Runs the cookbook's
# MI300X Flash FP8 low-latency cell, so it also covers MLA + KV-FP8 on gfx942.
# ==============================================================================
nightly-8-gpu-deepseek-v4-flash-rocm720:
name: ${{ format('nightly-8-gpu-deepseek-v4-flash ({0}, linux-mi300-8gpu-sglang)', matrix.rocm_version) }}
strategy:
fail-fast: false
matrix:
rocm_version: ${{ fromJson(inputs.rocm_version && inputs.rocm_version != 'all' && format('["{0}"]', inputs.rocm_version) || '["rocm724", "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-deepseek-v4-flash-rocm720,'))
runs-on: linux-mi300-8gpu-sglang
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.ref || github.sha }}
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Setup docker (${{ matrix.rocm_version }})
run: |
touch github_summary.md
bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version ${{ matrix.rocm_version }}
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
ENABLE_CACHE_HOST: "1"
- name: Install dependencies
run: bash scripts/ci/amd/amd_ci_install_dependency.sh --skip-test-time-deps
- name: Accuracy Test ROCm 7.2 (8-GPU DeepSeek-V4-Flash FP8)
timeout-minutes: 180
run: |
> github_summary.md # Clear summary file
bash scripts/ci/amd/amd_ci_exec.sh -w /sglang-checkout/test \
-e SGLANG_MOE_COPY_WEIGHT_VIEWS_BEFORE_H2D=1 \
-e GITHUB_STEP_SUMMARY="/sglang-checkout/github_summary.md" \
python3 run_suite.py --hw amd --suite nightly-amd-accuracy-8-gpu-deepseek-v4-flash --nightly --timeout-per-file 5400 ${{ (github.event_name == 'schedule' || inputs.continue_on_error) && '--continue-on-error' || '' }} || TEST_EXIT_CODE=$?
echo "$(<github_summary.md )" >> $GITHUB_STEP_SUMMARY || true
exit ${TEST_EXIT_CODE:-0}
# ==============================================================================
# 8-GPU DeepSeek-V4 (MI35x)
# ==============================================================================
nightly-8-gpu-mi35x-deepseek-v4-flash-rocm720:
@@ -2392,11 +2354,9 @@ jobs:
# 8-GPU Grok2 (MI30x + MI35x)
- nightly-8-gpu-grok2-rocm720
- nightly-8-gpu-mi35x-grok2-rocm720
# 8-GPU DeepSeek-V3.x (MI30x)
- nightly-8-gpu-deepseek-v31-rocm720
# 8-GPU DeepSeek-V3.2 (MI30x)
- nightly-8-gpu-deepseek-v32-rocm720
- nightly-8-gpu-deepseek-v32-mtp-rocm720
- nightly-8-gpu-deepseek-v3-kv-fp8-rocm720
# 8-GPU DeepSeek-V3.2 (MI35x)
- nightly-accuracy-8-gpu-mi35x-deepseek-v32-rocm720
- nightly-accuracy-8-gpu-mi35x-deepseek-v32-mtp-rocm720
@@ -2408,7 +2368,9 @@ jobs:
- nightly-8-gpu-mi35x-deepseek-r1-mxfp4-ar-fusion-rocm720
- nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4-rocm720
- nightly-8-gpu-mi35x-deepseek-r1-hicache-rocm720
# 8-GPU DeepSeek-V4 (MI35x only)
# 8-GPU DeepSeek-V4-Flash (MI30x)
- nightly-8-gpu-deepseek-v4-flash-rocm720
# 8-GPU DeepSeek-V4 (MI35x)
- 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