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

ci: add 4-GPU mi35x runner and rebalance off the saturated 8-GPU pool (#28745)

Browse Source 
Co-authored-by: michaelzhang-ai <michaelzhang-ai@users.noreply.github.com>
This commit is contained in:
Michael
2026-06-19 15:41:19 -07:00
committed by GitHub
co-authored by michaelzhang-ai
parent 3a574846ff
commit 13aab2fc06
6 changed files with 109 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/nightly-test-amd-rocm720.yml
+4 -2
View File
@@ -1064,7 +1064,8 @@ jobs:
nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4-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-r1-mxfp4-tp4-rocm720,'))
runs-on: linux-mi35x-gpu-8
# TP=4 workload: runs on the dedicated 4-GPU mi35x slice, not a whole 8-GPU node.
runs-on: linux-mi35x-gpu-4
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -1760,7 +1761,8 @@ jobs:
nightly-4-gpu-mi35x-minimax-m25-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-4-gpu-mi35x-minimax-m25-rocm720,'))
runs-on: linux-mi35x-gpu-8
# TP=4 workload: runs on the dedicated 4-GPU mi35x slice, not a whole 8-GPU node.
runs-on: linux-mi35x-gpu-4
steps:
- name: Checkout code
uses: actions/checkout@v4
.github/workflows/nightly-test-amd.yml
+4 -2
View File
@@ -1068,7 +1068,8 @@ jobs:
nightly-8-gpu-mi35x-deepseek-r1-mxfp4-tp4:
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-r1-mxfp4-tp4,'))
runs-on: linux-mi35x-gpu-8
# TP=4 workload: runs on the dedicated 4-GPU mi35x slice, not a whole 8-GPU node.
runs-on: linux-mi35x-gpu-4
steps:
- name: Checkout code
uses: actions/checkout@v4
@@ -1616,7 +1617,8 @@ jobs:
nightly-4-gpu-mi35x-minimax-m25:
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-4-gpu-mi35x-minimax-m25,'))
runs-on: linux-mi35x-gpu-8
# TP=4 workload: runs on the dedicated 4-GPU mi35x slice, not a whole 8-GPU node.
runs-on: linux-mi35x-gpu-4
steps:
- name: Checkout code
uses: actions/checkout@v4
.github/workflows/pr-test-amd-rocm720.yml
+45 -2
View File
@@ -48,6 +48,7 @@ on:
- stage-c-test-large-8-gpu-amd-mi35x-rocm720
- stage-b-test-large-8-gpu-mi35x-disaggregation-amd-rocm720
- stage-c-test-4-gpu-amd-rocm720
- stage-c-test-4-gpu-amd-mi35x-rocm720
- dsv4-flash-fp4-fp8-amd-rocm720
- dsv4-pro-fp4-amd-rocm720
target_stage:
@@ -988,7 +989,48 @@ jobs:
fail-fast: false
matrix:
runner: [linux-mi35x-gpu-8]
part: [0, 1]
part: [0]
runs-on: ${{matrix.runner}}
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.pr_head_sha || inputs.ref || github.sha }}
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Start CI container
run: bash scripts/ci/amd/amd_ci_start_container.sh --rocm-version rocm720
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test
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-mi35x --auto-partition-id ${{ matrix.part }} --auto-partition-size 1 --timeout-per-file 3600 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
# MI35x stage-c tests that only need 4 GPUs (e.g. DeepSeek-R1-MXFP4 TP=4),
# split off the 8-GPU mi35x suite onto the dedicated 4-GPU mi35x slice.
stage-c-test-4-gpu-amd-mi35x-rocm720:
needs: [check-changes]
if: |
always() &&
(
(contains(format(',{0},', inputs.target_stage || inputs.target_stage_select), ',stage-c-test-4-gpu-amd-mi35x-rocm720,')) ||
(
!(inputs.target_stage || inputs.target_stage_select) &&
(!failure() && !cancelled()) &&
((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
)
)
strategy:
fail-fast: false
matrix:
runner: [linux-mi35x-gpu-4]
part: [0]
runs-on: ${{matrix.runner}}
steps:
- name: Checkout code
@@ -1009,7 +1051,7 @@ jobs:
- name: Run test
timeout-minutes: 60
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-mi35x --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --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-4-gpu-amd-mi35x --auto-partition-id ${{ matrix.part }} --auto-partition-size 1 --timeout-per-file 3600 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
# =============================================== Disaggregation ====================================================
stage-b-test-large-8-gpu-mi35x-disaggregation-amd-rocm720:
@@ -1255,6 +1297,7 @@ jobs:
stage-b-test-2-gpu-large-amd-rocm720,
stage-b-test-large-8-gpu-mi35x-disaggregation-amd-rocm720,
stage-c-test-4-gpu-amd-rocm720,
stage-c-test-4-gpu-amd-mi35x-rocm720,
stage-c-test-large-8-gpu-amd-rocm720,
stage-c-test-large-8-gpu-amd-mi35x-rocm720,
.github/workflows/pr-test-amd.yml
+47 -2
View File
@@ -35,6 +35,7 @@ on:
- multimodal-gen-test-1-gpu-amd
- multimodal-gen-test-2-gpu-amd
- stage-c-test-4-gpu-amd
- stage-c-test-4-gpu-amd-mi35x
- stage-c-test-large-8-gpu-amd
- stage-c-test-large-8-gpu-amd-mi35x
- stage-b-test-large-8-gpu-mi35x-disaggregation-amd
@@ -1036,7 +1037,50 @@ jobs:
fail-fast: false
matrix:
runner: [linux-mi35x-gpu-8]
part: [0, 1]
part: [0]
runs-on: ${{matrix.runner}}
steps:
- name: Checkout code
uses: actions/checkout@v4
with:
ref: ${{ inputs.pr_head_sha || inputs.ref || github.sha }}
- name: Ensure VRAM is clear
run: bash scripts/ci/amd/ensure_vram_clear.sh rocm
- name: Start CI container
run: bash scripts/ci/amd/amd_ci_start_container.sh
env:
GITHUB_WORKSPACE: ${{ github.workspace }}
- name: Install dependencies
run: bash scripts/ci/amd/amd_ci_install_dependency.sh
- name: Run test
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-mi35x --auto-partition-id ${{ matrix.part }} --auto-partition-size 1 --timeout-per-file 3600 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
# MI35x stage-c tests that only need 4 GPUs (e.g. DeepSeek-R1-MXFP4 TP=4).
# Split off the 8-GPU mi35x suite so they no longer hold a whole 8-GPU node
# and instead run on the dedicated 4-GPU mi35x slice.
stage-c-test-4-gpu-amd-mi35x:
needs: [check-changes, call-gate, wait-for-stage-b-amd]
if: |
always() &&
(
(contains(format(',{0},', inputs.target_stage || inputs.target_stage_select), ',stage-c-test-4-gpu-amd-mi35x,')) ||
(
!(inputs.target_stage || inputs.target_stage_select) &&
((github.event_name == 'schedule') || (!failure() && !cancelled())) &&
((needs.check-changes.outputs.main_package == 'true') || (needs.check-changes.outputs.sgl_kernel == 'true'))
)
)
strategy:
fail-fast: false
matrix:
runner: [linux-mi35x-gpu-4]
part: [0]
runs-on: ${{matrix.runner}}
steps:
- name: Checkout code
@@ -1058,7 +1102,7 @@ jobs:
- name: Run test
timeout-minutes: 60
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-mi35x --auto-partition-id ${{ matrix.part }} --auto-partition-size 2 --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-4-gpu-amd-mi35x --auto-partition-id ${{ matrix.part }} --auto-partition-size 1 --timeout-per-file 3600 ${{ needs.check-changes.outputs.continue_on_error == 'true' && '--continue-on-error' || '' }}
# =============================================== Disaggregation ====================================================
stage-b-test-large-8-gpu-mi35x-disaggregation-amd:
@@ -1194,6 +1238,7 @@ jobs:
stage-b-test-2-gpu-large-amd,
stage-b-test-large-8-gpu-mi35x-disaggregation-amd,
stage-c-test-4-gpu-amd,
stage-c-test-4-gpu-amd-mi35x,
stage-c-test-large-8-gpu-amd,
stage-c-test-large-8-gpu-amd-mi35x,
]
test/registered/amd/test_deepseek_r1_mxfp4_8gpu.py → test/registered/amd/test_deepseek_r1_mxfp4_4gpu.py
+8 -6
View File
@@ -16,7 +16,7 @@ from sglang.test.test_utils import (
write_github_step_summary,
)
register_amd_ci(est_time=3600, suite="stage-c-test-large-8-gpu-amd-mi35x")
register_amd_ci(est_time=3600, suite="stage-c-test-4-gpu-amd-mi35x")
DEEPSEEK_R1_MODEL_PATH = "amd/DeepSeek-R1-MXFP4-Preview"
SERVER_LAUNCH_TIMEOUT = 1800
@@ -30,12 +30,12 @@ class TestDeepseekR1MXFP4(CustomTestCase):
# Workaround: AITER custom all-gather corrupts CUDA-graph IPC buffer
# registration and triggers a decode-time "Memory access fault" on
# MI35x TP=8. Disable until the AITER-side fix lands (see PR body).
# MI35x. Disable until the AITER-side fix lands (see PR body).
envs.SGLANG_USE_AITER_AG.set(False)
other_args = [
"--tp",
"8",
"4",
"--chunked-prefill-size",
"131072",
"--model-loader-extra-config",
@@ -90,7 +90,8 @@ class TestDeepseekR1MXFP4(CustomTestCase):
write_github_step_summary(
f"### test_bs_1_speed (deepseek-r1-mxfp4)\n" f"{speed=:.2f} token/s\n"
)
self.assertGreater(speed, 75)
# Report-only: the previous >75 tok/s gate was calibrated for TP=8.
# Decode throughput at TP=4 differs; re-calibrate before re-enabling.
class TestDeepseekR1MXFP4MTP(CustomTestCase):
@@ -105,7 +106,7 @@ class TestDeepseekR1MXFP4MTP(CustomTestCase):
other_args = [
"--tp",
"8",
"4",
"--chunked-prefill-size",
"131072",
"--speculative-algorithm",
@@ -175,7 +176,8 @@ class TestDeepseekR1MXFP4MTP(CustomTestCase):
f"{speed=:.2f} token/s\n"
)
self.assertGreater(acc_length, 2.04)
self.assertGreater(speed, 150)
# Report-only: the previous >150 tok/s gate was calibrated for TP=8.
# Decode throughput at TP=4 differs; re-calibrate before re-enabling.
if __name__ == "__main__":
test/run_suite.py
+1
View File
@@ -41,6 +41,7 @@ PER_COMMIT_SUITES = {
"jit-kernel-unit-test-amd",
"sgl-kernel-unit-test-2-gpu-amd",
"stage-c-test-4-gpu-amd",
"stage-c-test-4-gpu-amd-mi35x",
"stage-c-test-large-8-gpu-amd",
"stage-c-test-large-8-gpu-amd-mi35x",
# extra-a: label-gated PR opt-in suites in pr-test-amd-extra.yml