minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 13 Packages Projects Releases Wiki Activity

[AMD] prepare for MI300x PR runner pool: registry mirror, runner routing, threshold tuning (#23156)

Browse Source
This commit is contained in:
Bingxu Chen
2026-04-21 00:58:23 -07:00
committed by GitHub
parent 74fdf9cd77
commit 09b1d10d59
14 changed files with 145 additions and 117 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+2
View File
@@ -92,6 +92,8 @@ on:
env: env:
AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }} AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
concurrency: concurrency:
# When called via workflow_call with ref set, use a unique group per caller run to avoid # When called via workflow_call with ref set, use a unique group per caller run to avoid
.github/workflows/nightly-test-amd.yml
+2
View File
@@ -92,6 +92,8 @@ on:
env: env:
AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }} AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
concurrency: concurrency:
# When called via workflow_call with ref set, use a unique group per caller run to avoid # When called via workflow_call with ref set, use a unique group per caller run to avoid
.github/workflows/pr-test-amd-rocm720.yml
+2
View File
@@ -93,6 +93,8 @@ on:
env: env:
AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }} AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
concurrency: concurrency:
# When called via workflow_call with run_all_tests=true, use a unique group per run to # When called via workflow_call with run_all_tests=true, use a unique group per run to
.github/workflows/pr-test-amd.yml
+46 -51
View File
@@ -59,6 +59,14 @@ on:
required: false required: false
type: boolean type: boolean
default: false default: false
runner_arch:
description: 'AMD runner pool to dispatch GPU jobs to'
required: false
type: choice
default: mi300
options:
- mi300
- mi325
workflow_call: workflow_call:
inputs: inputs:
ref: ref:
@@ -84,6 +92,8 @@ on:
env: env:
AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }} AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
concurrency: concurrency:
# Scheduled and run_all_tests runs get unique groups (never cancel each other). # Scheduled and run_all_tests runs get unique groups (never cancel each other).
@@ -165,6 +175,7 @@ jobs:
# =============================================== sgl-kernel ==================================================== # =============================================== sgl-kernel ====================================================
sgl-kernel-unit-test-amd: sgl-kernel-unit-test-amd:
name: ${{ format('sgl-kernel-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
needs: [check-changes, call-gate] needs: [check-changes, call-gate]
if: | if: |
always() && !cancelled() && always() && !cancelled() &&
@@ -176,11 +187,7 @@ jobs:
needs.check-changes.outputs.sgl_kernel == 'true' needs.check-changes.outputs.sgl_kernel == 'true'
) )
) )
strategy: runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
runs-on: ${{matrix.runner}}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -226,6 +233,7 @@ jobs:
exit $failures exit $failures
sgl-kernel-unit-test-2-gpu-amd: sgl-kernel-unit-test-2-gpu-amd:
name: ${{ format('sgl-kernel-unit-test-2-gpu-amd (linux-{0}-2gpu-sglang)', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
needs: [check-changes, call-gate] needs: [check-changes, call-gate]
if: | if: |
always() && !cancelled() && always() && !cancelled() &&
@@ -237,11 +245,7 @@ jobs:
needs.check-changes.outputs.sgl_kernel == 'true' needs.check-changes.outputs.sgl_kernel == 'true'
) )
) )
strategy: runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
fail-fast: false
matrix:
runner: [linux-mi325-2gpu-sglang]
runs-on: ${{matrix.runner}}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -280,6 +284,7 @@ jobs:
# =============================================== primary ==================================================== # =============================================== primary ====================================================
stage-a-test-1-gpu-small-amd: stage-a-test-1-gpu-small-amd:
name: ${{ format('stage-a-test-1-gpu-small-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
needs: [check-changes, call-gate] needs: [check-changes, call-gate]
if: | if: |
always() && !cancelled() && always() && !cancelled() &&
@@ -291,11 +296,7 @@ jobs:
((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true')) ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
) )
) )
strategy: runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
runs-on: ${{matrix.runner}}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -320,6 +321,7 @@ jobs:
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-a-test-1-gpu-small-amd ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }} bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-a-test-1-gpu-small-amd ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
jit-kernel-unit-test-amd: jit-kernel-unit-test-amd:
name: ${{ format('jit-kernel-unit-test-amd (linux-{0}-1gpu-sglang)', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
needs: [check-changes, call-gate] needs: [check-changes, call-gate]
if: | if: |
always() && !cancelled() && always() && !cancelled() &&
@@ -331,11 +333,7 @@ jobs:
needs.check-changes.outputs.jit_kernel == 'true' needs.check-changes.outputs.jit_kernel == 'true'
) )
) )
strategy: runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
runs-on: ${{matrix.runner}}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -386,6 +384,7 @@ jobs:
max-wait-minutes: '240' max-wait-minutes: '240'
stage-b-test-1-gpu-small-amd: stage-b-test-1-gpu-small-amd:
name: ${{ format('stage-b-test-1-gpu-small-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325'), matrix.part) }}
needs: [check-changes, wait-for-stage-a-amd] needs: [check-changes, wait-for-stage-a-amd]
if: | if: |
always() && always() &&
@@ -400,9 +399,8 @@ jobs:
strategy: strategy:
fail-fast: false fail-fast: false
matrix: matrix:
runner: [linux-mi325-1gpu-sglang]
part: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13] part: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]
runs-on: ${{matrix.runner}} runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -421,11 +419,12 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test - name: Run test
timeout-minutes: 30 timeout-minutes: 45
run: | run: |
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 14 --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }} bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 14 --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
stage-b-test-1-gpu-small-amd-nondeterministic: stage-b-test-1-gpu-small-amd-nondeterministic:
name: ${{ format('stage-b-test-1-gpu-small-amd-nondeterministic (linux-{0}-1gpu-sglang)', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
needs: [check-changes, wait-for-stage-a-amd] needs: [check-changes, wait-for-stage-a-amd]
if: | if: |
always() && always() &&
@@ -437,11 +436,7 @@ jobs:
((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true')) ((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
) )
) )
strategy: runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
runs-on: ${{matrix.runner}}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -460,7 +455,7 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test - name: Run test
timeout-minutes: 30 timeout-minutes: 45
run: | run: |
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd-nondeterministic --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }} bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd-nondeterministic --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
@@ -504,6 +499,7 @@ jobs:
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd-mi35x ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }} bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-small-amd-mi35x ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
stage-b-test-1-gpu-large-amd: stage-b-test-1-gpu-large-amd:
name: ${{ format('stage-b-test-1-gpu-large-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325'), matrix.part) }}
needs: [check-changes, wait-for-stage-a-amd] needs: [check-changes, wait-for-stage-a-amd]
if: | if: |
always() && always() &&
@@ -518,9 +514,8 @@ jobs:
strategy: strategy:
fail-fast: false fail-fast: false
matrix: matrix:
runner: [linux-mi325-1gpu-sglang]
part: [0, 1, 2] part: [0, 1, 2]
runs-on: ${{matrix.runner}} runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -539,11 +534,12 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test - name: Run test
timeout-minutes: 30 timeout-minutes: 45
run: | run: |
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }} bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-1-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 2700 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
stage-b-test-2-gpu-large-amd: stage-b-test-2-gpu-large-amd:
name: ${{ format('stage-b-test-2-gpu-large-amd (linux-{0}-2gpu-sglang, {1})', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325'), matrix.part) }}
needs: [check-changes, wait-for-stage-a-amd] needs: [check-changes, wait-for-stage-a-amd]
if: | if: |
always() && always() &&
@@ -558,9 +554,8 @@ jobs:
strategy: strategy:
fail-fast: false fail-fast: false
matrix: matrix:
runner: [linux-mi325-2gpu-sglang]
part: [0, 1] part: [0, 1]
runs-on: ${{matrix.runner}} runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -579,11 +574,12 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test - name: Run test
timeout-minutes: 30 timeout-minutes: 45
run: | run: |
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-2-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --timeout-per-file 1800 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }} bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-b-test-2-gpu-large-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --timeout-per-file 2700 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
multimodal-gen-test-1-gpu-amd: multimodal-gen-test-1-gpu-amd:
name: ${{ format('multimodal-gen-test-1-gpu-amd (linux-{0}-1gpu-sglang, {1})', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325'), matrix.part) }}
needs: [check-changes, call-gate] needs: [check-changes, call-gate]
if: | if: |
always() && !cancelled() && always() && !cancelled() &&
@@ -599,9 +595,8 @@ jobs:
fail-fast: false fail-fast: false
max-parallel: 1 # Run one at a time to avoid eviction from resource exhaustion during AITER kernel JIT max-parallel: 1 # Run one at a time to avoid eviction from resource exhaustion during AITER kernel JIT
matrix: matrix:
runner: [linux-mi325-1gpu-sglang] part: [0, 1, 2, 3]
part: [0, 1, 2, 3] # 2 partitions: 11 tests ÷ 2 = ~5-6 tests each runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
runs-on: ${{matrix.runner}}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -715,6 +710,7 @@ jobs:
free -h free -h
multimodal-gen-test-2-gpu-amd: multimodal-gen-test-2-gpu-amd:
name: ${{ format('multimodal-gen-test-2-gpu-amd (linux-{0}-2gpu-sglang, {1})', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325'), matrix.part) }}
needs: [check-changes, call-gate] needs: [check-changes, call-gate]
if: | if: |
always() && !cancelled() && always() && !cancelled() &&
@@ -730,9 +726,8 @@ jobs:
fail-fast: false fail-fast: false
max-parallel: 1 # Run one at a time to avoid eviction from resource exhaustion during AITER kernel JIT max-parallel: 1 # Run one at a time to avoid eviction from resource exhaustion during AITER kernel JIT
matrix: matrix:
runner: [linux-mi325-2gpu-sglang]
part: [0, 1, 2] # 3 partitions: 2 parametrized + 1 standalone (test_disagg_server.py) part: [0, 1, 2] # 3 partitions: 2 parametrized + 1 standalone (test_disagg_server.py)
runs-on: ${{matrix.runner}} runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -812,7 +807,7 @@ jobs:
docker exec ci_sglang rocm-smi --showmeminfo vram 2>/dev/null || echo "rocm-smi not available" docker exec ci_sglang rocm-smi --showmeminfo vram 2>/dev/null || echo "rocm-smi not available"
- name: Run diffusion server tests (2-GPU) - name: Run diffusion server tests (2-GPU)
timeout-minutes: 80 timeout-minutes: 90
run: | run: |
# AMD CI: All 2-GPU tests including LoRA # AMD CI: All 2-GPU tests including LoRA
# Tests: T2V, T2I, I2V, LoRA # Tests: T2V, T2I, I2V, LoRA
@@ -871,6 +866,7 @@ jobs:
max-wait-minutes: '480' max-wait-minutes: '480'
stage-c-test-4-gpu-amd: stage-c-test-4-gpu-amd:
name: ${{ format('stage-c-test-4-gpu-amd (linux-{0}-4gpu-sglang, {1})', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325'), matrix.part) }}
needs: [check-changes, call-gate, wait-for-stage-b-amd] needs: [check-changes, call-gate, wait-for-stage-b-amd]
if: | if: |
always() && always() &&
@@ -885,9 +881,8 @@ jobs:
strategy: strategy:
fail-fast: false fail-fast: false
matrix: matrix:
runner: [linux-mi325-4gpu-sglang]
part: [0] part: [0]
runs-on: ${{matrix.runner}} runs-on: ${{ format('linux-{0}-4gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -906,7 +901,7 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test - name: Run test
timeout-minutes: 60 timeout-minutes: 90
run: | run: |
bash scripts/ci/amd/amd_ci_exec.sh \ bash scripts/ci/amd/amd_ci_exec.sh \
-e NCCL_CUMEM_ENABLE=0 \ -e NCCL_CUMEM_ENABLE=0 \
@@ -919,7 +914,7 @@ jobs:
--suite stage-c-test-4-gpu-amd \ --suite stage-c-test-4-gpu-amd \
--auto-partition-id ${{ matrix.part }} \ --auto-partition-id ${{ matrix.part }} \
--auto-partition-size 1 \ --auto-partition-size 1 \
--timeout-per-file 1800 \ --timeout-per-file 5400 \
--enable-retry \ --enable-retry \
--max-attempts 2 \ --max-attempts 2 \
--retry-wait-seconds 120 \ --retry-wait-seconds 120 \
@@ -927,6 +922,7 @@ jobs:
${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }} ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
stage-c-test-large-8-gpu-amd: stage-c-test-large-8-gpu-amd:
name: ${{ format('stage-c-test-large-8-gpu-amd (linux-{0}-8gpu-sglang, {1})', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325'), matrix.part) }}
needs: [check-changes, call-gate, wait-for-stage-b-amd] needs: [check-changes, call-gate, wait-for-stage-b-amd]
if: | if: |
always() && always() &&
@@ -939,13 +935,12 @@ jobs:
) )
) )
env: env:
RUNNER_LABELS: linux-mi325-8gpu-sglang RUNNER_LABELS: ${{ format('linux-{0}-8gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
strategy: strategy:
fail-fast: false fail-fast: false
matrix: matrix:
runner: [linux-mi325-8gpu-sglang]
part: [0, 1, 2] part: [0, 1, 2]
runs-on: ${{matrix.runner}} runs-on: ${{ format('linux-{0}-8gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps: steps:
- name: Checkout code - name: Checkout code
uses: actions/checkout@v4 uses: actions/checkout@v4
@@ -970,7 +965,7 @@ jobs:
docker exec ci_sglang bash -c "cd /sglang-checkout && NCCL_DEBUG=INFO RCCL_DEBUG=INFO torchrun --nproc_per_node=8 scripts/ci/amd/test_rccl_multi_gpu.py" docker exec ci_sglang bash -c "cd /sglang-checkout && NCCL_DEBUG=INFO RCCL_DEBUG=INFO torchrun --nproc_per_node=8 scripts/ci/amd/test_rccl_multi_gpu.py"
- name: Run test - name: Run test
timeout-minutes: 60 timeout-minutes: 120
run: | run: |
bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-c-test-large-8-gpu-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 3600 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }} bash scripts/ci/amd/amd_ci_exec.sh -w "/sglang-checkout/test" python3 run_suite.py --hw amd --suite stage-c-test-large-8-gpu-amd --auto-partition-id ${{ matrix.part }} --auto-partition-size 3 --timeout-per-file 3600 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
scripts/ci/amd/amd_ci_start_container.sh
+19 -19
View File
@@ -154,18 +154,15 @@ find_latest_image() {
fi fi
done done
# Then try the local registry. # If not found locally, fall back to pulling from public registry.
for days_back in {0..6}; do # We intentionally do not probe ${LOCAL_DOCKER_REGISTRY} here with
image_tag="${base_tag}-$(date -d "${days_back} days ago" +%Y%m%d)" # `docker manifest inspect --insecure` because that command runs in the
echo "Checking for image: ${LOCAL_DOCKER_REGISTRY}/rocm/sgl-dev:${image_tag}" >&2 # runner pod's network namespace, which on every observed AMD scale set
if docker manifest inspect --insecure "${LOCAL_DOCKER_REGISTRY}/rocm/sgl-dev:${image_tag}" >/dev/null 2>&1; then # cannot reach 172.29.8.23:5000 (every probe either fast-fails with TLS
echo "Found available image: ${LOCAL_DOCKER_REGISTRY}/rocm/sgl-dev:${image_tag}" >&2 # reject or hits a 30s TCP timeout, multiplied across 7 daily candidates).
echo "${LOCAL_DOCKER_REGISTRY}/rocm/sgl-dev:${image_tag}" # The actual local-registry pull still happens in the call site below via
return 0 # `docker pull "${LOCAL_DOCKER_REGISTRY}/${IMAGE}"`, which goes through the
fi # docker daemon on the host and inherits its insecure-registries config.
done
# Finally, try the public registry.
for days_back in {0..6}; do for days_back in {0..6}; do
image_tag="${base_tag}-$(date -d "${days_back} days ago" +%Y%m%d)" image_tag="${base_tag}-$(date -d "${days_back} days ago" +%Y%m%d)"
echo "Checking for image: rocm/sgl-dev:${image_tag}" >&2 echo "Checking for image: rocm/sgl-dev:${image_tag}" >&2
@@ -260,14 +257,17 @@ elif [[ -n "${BUILD_FROM_DOCKERFILE}" ]]; then
else else
# Find the latest pre-built image # Find the latest pre-built image
IMAGE=$(find_latest_image "${GPU_ARCH}") IMAGE=$(find_latest_image "${GPU_ARCH}")
echo "Pulling Docker image: ${IMAGE}" # Try the local docker registry first (avoids Docker Hub rate limits and is
if [[ "${IMAGE}" == "${LOCAL_DOCKER_REGISTRY}/"* ]]; then # faster on the LAN); if that fails for any reason, fall back to the
# Local registry is on-LAN; no need to retry. # public registry with exponential-backoff retries. Capture stderr so the
docker pull "${IMAGE}" # real failure reason (TLS handshake, 404, connection refused, etc.) is
docker tag "${IMAGE}" "${IMAGE#${LOCAL_DOCKER_REGISTRY}/}" # visible in the job log instead of being silently swallowed.
IMAGE="${IMAGE#${LOCAL_DOCKER_REGISTRY}/}" if local_pull_output=$(docker pull "${LOCAL_DOCKER_REGISTRY}/${IMAGE}" 2>&1); then
echo "Pulled from local docker registry: ${LOCAL_DOCKER_REGISTRY}/${IMAGE}"
docker tag "${LOCAL_DOCKER_REGISTRY}/${IMAGE}" "${IMAGE}"
else else
# Public registry pulls can hit rate limits; retry with backoff. echo "Local docker registry pull failed; falling back to public registry: ${IMAGE}" >&2
printf '%s\n' "${local_pull_output}" | sed 's/^/ [local-pull] /' >&2
retry_with_backoff 6 docker pull "${IMAGE}" retry_with_backoff 6 docker pull "${IMAGE}"
fi fi
fi fi
scripts/ci/amd/amd_ci_start_container_disagg.sh
+15 -19
View File
@@ -141,18 +141,11 @@ find_latest_image() {
fi fi
done done
# Then try the local registry. # If not found locally, fall back to pulling from public registry.
for days_back in {0..6}; do # See amd_ci_start_container.sh for why we don't probe
image_tag="${base_tag}-$(date -d "${days_back} days ago" +%Y%m%d)" # ${LOCAL_DOCKER_REGISTRY} with `docker manifest inspect --insecure` from
echo "Checking for image: ${LOCAL_DOCKER_REGISTRY}/rocm/sgl-dev:${image_tag}" >&2 # the runner pod's network namespace; the actual local-registry pull
if docker manifest inspect --insecure "${LOCAL_DOCKER_REGISTRY}/rocm/sgl-dev:${image_tag}" >/dev/null 2>&1; then # happens at the call site below via the docker daemon on the host.
echo "Found available image: ${LOCAL_DOCKER_REGISTRY}/rocm/sgl-dev:${image_tag}" >&2
echo "${LOCAL_DOCKER_REGISTRY}/rocm/sgl-dev:${image_tag}"
return 0
fi
done
# Finally, try the public registry.
for days_back in {0..6}; do for days_back in {0..6}; do
image_tag="${base_tag}-$(date -d "${days_back} days ago" +%Y%m%d)" image_tag="${base_tag}-$(date -d "${days_back} days ago" +%Y%m%d)"
echo "Checking for image: rocm/sgl-dev:${image_tag}" >&2 echo "Checking for image: rocm/sgl-dev:${image_tag}" >&2
@@ -210,14 +203,17 @@ find_latest_image() {
# Pull and run the latest image # Pull and run the latest image
IMAGE=$(find_latest_image "${GPU_ARCH}") IMAGE=$(find_latest_image "${GPU_ARCH}")
echo "Pulling Docker image: ${IMAGE}" # Try the local docker registry first (avoids Docker Hub rate limits and is
if [[ "${IMAGE}" == "${LOCAL_DOCKER_REGISTRY}/"* ]]; then # faster on the LAN); if that fails for any reason, fall back to the
# Local registry is on-LAN; no need to retry. # public registry with exponential-backoff retries. Capture stderr so the
docker pull "${IMAGE}" # real failure reason (TLS handshake, 404, connection refused, etc.) is
docker tag "${IMAGE}" "${IMAGE#${LOCAL_DOCKER_REGISTRY}/}" # visible in the job log instead of being silently swallowed.
IMAGE="${IMAGE#${LOCAL_DOCKER_REGISTRY}/}" if local_pull_output=$(docker pull "${LOCAL_DOCKER_REGISTRY}/${IMAGE}" 2>&1); then
echo "Pulled from local docker registry: ${LOCAL_DOCKER_REGISTRY}/${IMAGE}"
docker tag "${LOCAL_DOCKER_REGISTRY}/${IMAGE}" "${IMAGE}"
else else
# Public registry pulls can hit rate limits; retry with backoff. echo "Local docker registry pull failed; falling back to public registry: ${IMAGE}" >&2
printf '%s\n' "${local_pull_output}" | sed 's/^/ [local-pull] /' >&2
retry_with_backoff 6 docker pull "${IMAGE}" retry_with_backoff 6 docker pull "${IMAGE}"
fi fi
test/registered/amd/test_deepseek_v32_basic.py
+1 -1
View File
@@ -15,7 +15,7 @@ from sglang.test.test_utils import (
write_github_step_summary, write_github_step_summary,
) )
register_amd_ci(est_time=1800, suite="stage-c-test-large-8-gpu-amd") register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd")
DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2" DEEPSEEK_V32_MODEL_PATH = "deepseek-ai/DeepSeek-V3.2"
test/registered/amd/test_deepseek_v3_mtp.py
+2 -1
View File
@@ -84,7 +84,8 @@ class TestDeepseekV3MTP(CustomTestCase):
f'{metrics["accuracy"]=:.3f}\n' f'{metrics["accuracy"]=:.3f}\n'
f"{avg_spec_accept_length=:.2f}\n" f"{avg_spec_accept_length=:.2f}\n"
) )
self.assertGreater(metrics["accuracy"], 0.935) # relax for mi300x
self.assertGreaterEqual(metrics["accuracy"], 0.93)
if is_in_amd_ci(): if is_in_amd_ci():
self.assertGreater(avg_spec_accept_length, 2.8) self.assertGreater(avg_spec_accept_length, 2.8)
else: else:
test/registered/amd/test_moriep_small.py
+1 -1
View File
@@ -60,7 +60,7 @@ common_args = [
"1", "1",
"--enable-dp-lm-head", "--enable-dp-lm-head",
"--mem-fraction-static", "--mem-fraction-static",
"0.6", "0.72", # relax for mi300x
"--chunked-prefill-size", "--chunked-prefill-size",
"32768", "32768",
"--max-running-requests", "--max-running-requests",
test/registered/moe/test_torch_compile_moe.py
+4
View File
@@ -12,6 +12,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_amd_ci,
popen_launch_server, popen_launch_server,
) )
@@ -73,6 +74,9 @@ class TestTorchCompileMoe(CustomTestCase):
throughput = max_tokens / (tok - tic) throughput = max_tokens / (tok - tic)
if is_cuda(): if is_cuda():
self.assertGreaterEqual(throughput, 285) self.assertGreaterEqual(throughput, 285)
elif is_in_amd_ci():
# relax for mi300x
self.assertGreaterEqual(throughput, 240)
else: else:
self.assertGreaterEqual(throughput, 270) self.assertGreaterEqual(throughput, 270)
test/registered/perf/test_bench_serving_1gpu_part1.py
+10 -8
View File
@@ -142,9 +142,6 @@ class TestBenchServing1GPUPart1(CustomTestCase):
self.assertLess(res["median_itl_ms"], 10) self.assertLess(res["median_itl_ms"], 10)
def test_online_lora_latency(self): def test_online_lora_latency(self):
if is_in_amd_ci():
pass
res = self._run_lora_latency_test(enable_background_task=False) res = self._run_lora_latency_test(enable_background_task=False)
if is_in_ci(): if is_in_ci():
@@ -154,12 +151,13 @@ class TestBenchServing1GPUPart1(CustomTestCase):
f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n" f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
) )
self.assertLess(res["median_e2e_latency_ms"], 2400) self.assertLess(res["median_e2e_latency_ms"], 2400)
self.assertLess(res["median_ttft_ms"], 58) # relax for mi300x (LoRA TTFT ~2x slower than mi325)
if is_in_amd_ci():
self.assertLess(res["median_ttft_ms"], 100)
else:
self.assertLess(res["median_ttft_ms"], 58)
def test_online_lora_latency_with_concurrent_adapter_updates(self): def test_online_lora_latency_with_concurrent_adapter_updates(self):
if is_in_amd_ci():
pass
res = self._run_lora_latency_test(enable_background_task=True) res = self._run_lora_latency_test(enable_background_task=True)
if is_in_ci(): if is_in_ci():
@@ -169,7 +167,11 @@ class TestBenchServing1GPUPart1(CustomTestCase):
f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n" f"median_ttft_ms: {res['median_ttft_ms']:.2f} ms\n"
) )
self.assertLess(res["median_e2e_latency_ms"], 4000) self.assertLess(res["median_e2e_latency_ms"], 4000)
self.assertLess(res["median_ttft_ms"], 80) # relax for mi300x (LoRA TTFT ~2x slower than mi325)
if is_in_amd_ci():
self.assertLess(res["median_ttft_ms"], 130)
else:
self.assertLess(res["median_ttft_ms"], 80)
def _run_lora_latency_test(self, enable_background_task: bool): def _run_lora_latency_test(self, enable_background_task: bool):
""" """
test/registered/perf/test_bench_serving_1gpu_part2.py
+31 -16
View File
@@ -41,8 +41,9 @@ class TestBenchServing1GPUPart2(CustomTestCase):
f"### test_vlm_offline_throughput\n" f"### test_vlm_offline_throughput\n"
f"Output throughput: {res['output_throughput']:.2f} token/s\n" f"Output throughput: {res['output_throughput']:.2f} token/s\n"
) )
# relax for mi300x
if is_in_amd_ci(): if is_in_amd_ci():
self.assertGreater(res["output_throughput"], 2000) self.assertGreater(res["output_throughput"], 900)
else: else:
self.assertGreater(res["output_throughput"], 2500) self.assertGreater(res["output_throughput"], 2500)
@@ -116,12 +117,16 @@ class TestBenchServing1GPUPart2(CustomTestCase):
) )
self.assertEqual(res["successful_requests"], res["total_requests"]) self.assertEqual(res["successful_requests"], res["total_requests"])
bounds = { # relax for mi300x
10: (45, 50), if is_in_amd_ci():
25: (50, 60), bounds = {10: (60, 65), 25: (70, 80), 50: (80, 90)}
50: (60, 65), default_bounds = (90, 90)
} else:
avg_latency_bound, p95_latency_bound = bounds.get(batch_size, (60, 65)) bounds = {10: (45, 50), 25: (50, 60), 50: (60, 65)}
default_bounds = (60, 65)
avg_latency_bound, p95_latency_bound = bounds.get(
batch_size, default_bounds
)
self.assertLess(res["avg_latency_ms"], avg_latency_bound) self.assertLess(res["avg_latency_ms"], avg_latency_bound)
self.assertLess(res["p95_latency_ms"], p95_latency_bound) self.assertLess(res["p95_latency_ms"], p95_latency_bound)
@@ -146,9 +151,15 @@ class TestBenchServing1GPUPart2(CustomTestCase):
) )
self.assertEqual(res["successful_requests"], res["total_requests"]) self.assertEqual(res["successful_requests"], res["total_requests"])
self.assertLess(res["avg_latency_ms"], 20) # relax for mi300x
self.assertLess(res["p95_latency_ms"], 25) if is_in_amd_ci():
self.assertGreater(res["throughput"], 60) self.assertLess(res["avg_latency_ms"], 35)
self.assertLess(res["p95_latency_ms"], 40)
self.assertGreater(res["throughput"], 30)
else:
self.assertLess(res["avg_latency_ms"], 20)
self.assertLess(res["p95_latency_ms"], 25)
self.assertGreater(res["throughput"], 60)
def test_embeddings_api_batch_scaling(self): def test_embeddings_api_batch_scaling(self):
"""Test embeddings API performance with different batch sizes""" """Test embeddings API performance with different batch sizes"""
@@ -173,12 +184,16 @@ class TestBenchServing1GPUPart2(CustomTestCase):
) )
self.assertEqual(res["successful_requests"], res["total_requests"]) self.assertEqual(res["successful_requests"], res["total_requests"])
bounds = { # relax for mi300x
10: (60, 65), if is_in_amd_ci():
25: (115, 120), bounds = {10: (80, 90), 25: (140, 150), 50: (230, 240)}
50: (190, 195), default_bounds = (300, 300)
} else:
avg_latency_bound, p95_latency_bound = bounds.get(batch_size, (250, 250)) bounds = {10: (60, 65), 25: (115, 120), 50: (190, 195)}
default_bounds = (250, 250)
avg_latency_bound, p95_latency_bound = bounds.get(
batch_size, default_bounds
)
self.assertLess(res["avg_latency_ms"], avg_latency_bound) self.assertLess(res["avg_latency_ms"], avg_latency_bound)
self.assertLess(res["p95_latency_ms"], p95_latency_bound) self.assertLess(res["p95_latency_ms"], p95_latency_bound)
test/registered/sampling/test_pytorch_sampling_backend.py
+7
View File
@@ -11,6 +11,7 @@ from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH, DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST, DEFAULT_URL_FOR_TEST,
CustomTestCase, CustomTestCase,
is_in_amd_ci,
popen_launch_server, popen_launch_server,
) )
@@ -47,6 +48,12 @@ class TestPyTorchSamplingBackend(CustomTestCase):
metrics = run_eval(args) metrics = run_eval(args)
self.assertGreaterEqual(metrics["score"], 0.65) self.assertGreaterEqual(metrics["score"], 0.65)
@unittest.skipIf(
is_in_amd_ci(),
"Skip on MI300x: greedy decode is not bit-exact across runs on MI300x "
"(kernel-level numerical jitter), so the assertEqual on identical "
"regenerated text is flaky on this runner pool.",
)
def test_greedy(self): def test_greedy(self):
first_text = None first_text = None
test/registered/tokenizer/test_multi_tokenizer.py
+3 -1
View File
@@ -10,6 +10,7 @@ from sglang.test.test_utils import (
CustomTestCase, CustomTestCase,
auto_config_device, auto_config_device,
get_benchmark_args, get_benchmark_args,
is_in_amd_ci,
is_in_ci, is_in_ci,
popen_launch_server, popen_launch_server,
run_benchmark, run_benchmark,
@@ -70,7 +71,8 @@ class TestMultiTokenizer(CustomTestCase, MMLUMixin):
f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n" f"median_e2e_latency_ms: {res['median_e2e_latency_ms']:.2f} ms\n"
) )
self.assertLess(res["median_e2e_latency_ms"], 11000) self.assertLess(res["median_e2e_latency_ms"], 11000)
self.assertLess(res["median_ttft_ms"], 86) # relax for mi300x
self.assertLess(res["median_ttft_ms"], 130 if is_in_amd_ci() else 86)
self.assertLess(res["median_itl_ms"], 10) self.assertLess(res["median_itl_ms"], 10)