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

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
+46 -51
View File
@@ -59,6 +59,14 @@ on:
required: false
type: boolean
default: false
runner_arch:
description: 'AMD runner pool to dispatch GPU jobs to'
required: false
type: choice
default: mi300
options:
- mi300
- mi325
workflow_call:
inputs:
ref:
@@ -84,6 +92,8 @@ on:
env:
AITER_COMMIT_OVERRIDE: ${{ inputs.aiter_ref }}
DOCKERHUB_AMD_USERNAME: ${{ secrets.DOCKERHUB_AMD_USERNAME }}
DOCKERHUB_AMD_TOKEN: ${{ secrets.DOCKERHUB_AMD_TOKEN }}
concurrency:
# Scheduled and run_all_tests runs get unique groups (never cancel each other).
@@ -165,6 +175,7 @@ jobs:
# =============================================== sgl-kernel ====================================================
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]
if: |
always() && !cancelled() &&
@@ -176,11 +187,7 @@ jobs:
needs.check-changes.outputs.sgl_kernel == 'true'
)
)
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
runs-on: ${{matrix.runner}}
runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -226,6 +233,7 @@ jobs:
exit $failures
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]
if: |
always() && !cancelled() &&
@@ -237,11 +245,7 @@ jobs:
needs.check-changes.outputs.sgl_kernel == 'true'
)
)
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-2gpu-sglang]
runs-on: ${{matrix.runner}}
runs-on: ${{ format('linux-{0}-2gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -280,6 +284,7 @@ jobs:
# =============================================== primary ====================================================
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]
if: |
always() && !cancelled() &&
@@ -291,11 +296,7 @@ jobs:
((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
)
)
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
runs-on: ${{matrix.runner}}
runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps:
- name: Checkout code
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' || '' }}
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]
if: |
always() && !cancelled() &&
@@ -331,11 +333,7 @@ jobs:
needs.check-changes.outputs.jit_kernel == 'true'
)
)
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
runs-on: ${{matrix.runner}}
runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -386,6 +384,7 @@ jobs:
max-wait-minutes: '240'
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]
if: |
always() &&
@@ -400,9 +399,8 @@ jobs:
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
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:
- name: Checkout code
uses: actions/checkout@v4
@@ -421,11 +419,12 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test
timeout-minutes: 30
timeout-minutes: 45
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' || '' }}
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]
if: |
always() &&
@@ -437,11 +436,7 @@ jobs:
((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
)
)
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
runs-on: ${{matrix.runner}}
runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -460,7 +455,7 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test
timeout-minutes: 30
timeout-minutes: 45
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' || '' }}
@@ -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' || '' }}
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]
if: |
always() &&
@@ -518,9 +514,8 @@ jobs:
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-1gpu-sglang]
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:
- name: Checkout code
uses: actions/checkout@v4
@@ -539,11 +534,12 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test
timeout-minutes: 30
timeout-minutes: 45
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:
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]
if: |
always() &&
@@ -558,9 +554,8 @@ jobs:
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-2gpu-sglang]
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:
- name: Checkout code
uses: actions/checkout@v4
@@ -579,11 +574,12 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test
timeout-minutes: 30
timeout-minutes: 45
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:
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]
if: |
always() && !cancelled() &&
@@ -599,9 +595,8 @@ jobs:
fail-fast: false
max-parallel: 1 # Run one at a time to avoid eviction from resource exhaustion during AITER kernel JIT
matrix:
runner: [linux-mi325-1gpu-sglang]
part: [0, 1, 2, 3] # 2 partitions: 11 tests ÷ 2 = ~5-6 tests each
runs-on: ${{matrix.runner}}
part: [0, 1, 2, 3]
runs-on: ${{ format('linux-{0}-1gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -715,6 +710,7 @@ jobs:
free -h
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]
if: |
always() && !cancelled() &&
@@ -730,9 +726,8 @@ jobs:
fail-fast: false
max-parallel: 1 # Run one at a time to avoid eviction from resource exhaustion during AITER kernel JIT
matrix:
runner: [linux-mi325-2gpu-sglang]
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:
- name: Checkout code
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"
- name: Run diffusion server tests (2-GPU)
timeout-minutes: 80
timeout-minutes: 90
run: |
# AMD CI: All 2-GPU tests including LoRA
# Tests: T2V, T2I, I2V, LoRA
@@ -871,6 +866,7 @@ jobs:
max-wait-minutes: '480'
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]
if: |
always() &&
@@ -885,9 +881,8 @@ jobs:
strategy:
fail-fast: false
matrix:
runner: [linux-mi325-4gpu-sglang]
part: [0]
runs-on: ${{matrix.runner}}
runs-on: ${{ format('linux-{0}-4gpu-sglang', inputs.runner_arch || (github.event_name == 'pull_request' && 'mi300' || 'mi325')) }}
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -906,7 +901,7 @@ jobs:
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test
timeout-minutes: 60
timeout-minutes: 90
run: |
bash scripts/ci/amd/amd_ci_exec.sh \
-e NCCL_CUMEM_ENABLE=0 \
@@ -919,7 +914,7 @@ jobs:
--suite stage-c-test-4-gpu-amd \
--auto-partition-id ${{ matrix.part }} \
--auto-partition-size 1 \
--timeout-per-file 1800 \
--timeout-per-file 5400 \
--enable-retry \
--max-attempts 2 \
--retry-wait-seconds 120 \
@@ -927,6 +922,7 @@ jobs:
${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
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]
if: |
always() &&
@@ -939,13 +935,12 @@ jobs:
)
)
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:
fail-fast: false
matrix:
runner: [linux-mi325-8gpu-sglang]
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:
- name: Checkout code
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"
- name: Run test
timeout-minutes: 60
timeout-minutes: 120
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' || '' }}