minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 10 Packages Projects Releases Wiki Activity

[diffusion] CI: rebalance 2-gpu shards and cut the job timeout to 45m (#38239)

Browse Source
This commit is contained in:
Cheng Wan
2026-09-07 09:39:03 +08:00
committed by GitHub
parent b83f1bdd21
commit e3140fb9d4
3 changed files with 72 additions and 27 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/pr-test-multimodal-gen.yml
+2 -2
View File
@@ -292,7 +292,7 @@ jobs:
((github.event_name == 'schedule' || inputs.test_parallel_dispatch == 'true') || (inputs.caller_needs_failure != 'true' && !cancelled())) &&
inputs.multimodal_gen == 'true'
runs-on: 2-gpu-h100
timeout-minutes: 240
timeout-minutes: 45
strategy:
fail-fast: false
matrix: ${{ fromJson(needs.compute-diffusion-partitions.outputs.matrix-2gpu) }}
@@ -324,7 +324,7 @@ jobs:
CUSTOM_BUILD_SGL_KERNEL=${{inputs.sgl_kernel}} bash scripts/ci/cuda/ci_install_dependency.sh diffusion
- name: Run diffusion server tests
timeout-minutes: 240
timeout-minutes: 45
env:
HF_TOKEN: ${{ secrets.SGLANG_DIFFUSION_CI_HF_TOKEN || secrets.HF_TOKEN }}
HUGGING_FACE_HUB_TOKEN: ${{ secrets.SGLANG_DIFFUSION_CI_HF_TOKEN || secrets.HF_TOKEN }}
python/sglang/multimodal_gen/test/server/perf_baselines/h100.json
+49 -25
View File
@@ -177,7 +177,7 @@
"runtime_peak_vram_mb": 46938.0,
"load_peak_allocated_mb": 41798.0,
"runtime_peak_allocated_mb": 44855.0,
"estimated_full_test_time_s": 133.2
"estimated_full_test_time_s": 65.6
},
"qwen_image_t2i_2_gpus_extra_high": {
"stages_ms": {},
@@ -185,7 +185,31 @@
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 57.9
"estimated_full_test_time_s": 54.4
},
"flux2_modelopt_fp8_tp2_t2i": {
"stages_ms": {},
"denoise_step_ms": {},
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 75.4
},
"joy_echo_t2v_2gpu": {
"stages_ms": {},
"denoise_step_ms": {},
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 62.8
},
"ideogram4_fp8_tp2_t2i": {
"stages_ms": {},
"denoise_step_ms": {},
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 60.8
},
"ideogram4_fp8_t2i": {
"stages_ms": {
@@ -697,7 +721,7 @@
"runtime_peak_vram_mb": 28148.0,
"load_peak_allocated_mb": 23262.0,
"runtime_peak_allocated_mb": 25541.0,
"estimated_full_test_time_s": 125.5
"estimated_full_test_time_s": 44.9
},
"zimage_image_t2i": {
"stages_ms": {
@@ -813,7 +837,7 @@
"runtime_peak_vram_mb": 17602.0,
"load_peak_allocated_mb": 12852.0,
"runtime_peak_allocated_mb": 15131.0,
"estimated_full_test_time_s": 75.0
"estimated_full_test_time_s": 39.3
},
"qwen_image_edit_ti2i": {
"stages_ms": {
@@ -1225,7 +1249,7 @@
"runtime_peak_vram_mb": 11972.0,
"load_peak_allocated_mb": 5274.0,
"runtime_peak_allocated_mb": 9251.0,
"estimated_full_test_time_s": 127.6
"estimated_full_test_time_s": 45.5
},
"turbo_wan2_1_t2v_1.3b": {
"stages_ms": {
@@ -1321,7 +1345,7 @@
"runtime_peak_vram_mb": 57528.0,
"load_peak_allocated_mb": 43411.0,
"runtime_peak_allocated_mb": 54736.0,
"estimated_full_test_time_s": 345.4
"estimated_full_test_time_s": 100.4
},
"wan2_2_ti2v_5b": {
"stages_ms": {
@@ -1632,7 +1656,7 @@
"runtime_peak_vram_mb": 26322.0,
"load_peak_allocated_mb": 6177.0,
"runtime_peak_allocated_mb": 19978.0,
"estimated_full_test_time_s": 264.1
"estimated_full_test_time_s": 168.7
},
"wan2_1_i2v_14b_480P_2gpu": {
"stages_ms": {
@@ -1705,7 +1729,7 @@
"warmup_peak_vram_mb": 45782.0,
"load_peak_allocated_mb": 34033.0,
"runtime_peak_allocated_mb": 38569.0,
"estimated_full_test_time_s": 243.0
"estimated_full_test_time_s": 128.8
},
"wan2_1_i2v_14b_720P_2gpu": {
"stages_ms": {
@@ -1777,7 +1801,7 @@
"runtime_peak_vram_mb": 51230.0,
"load_peak_allocated_mb": 34033.0,
"runtime_peak_allocated_mb": 44644.0,
"estimated_full_test_time_s": 248.9
"estimated_full_test_time_s": 182.2
},
"wan2_2_t2v_a14b_2gpu": {
"stages_ms": {
@@ -1838,7 +1862,7 @@
"warmup_peak_vram_mb": 26902.0,
"load_peak_allocated_mb": 6175.0,
"runtime_peak_allocated_mb": 15472.0,
"estimated_full_test_time_s": 204.3
"estimated_full_test_time_s": 188.0
},
"wan2_1_t2v_14b_2gpu": {
"stages_ms": {
@@ -1908,7 +1932,7 @@
"runtime_peak_vram_mb": 37788.0,
"load_peak_allocated_mb": 29869.0,
"runtime_peak_allocated_mb": 33843.0,
"estimated_full_test_time_s": 173.2
"estimated_full_test_time_s": 96.3
},
"wan2_2_t2v_a14b_lora_2gpu": {
"stages_ms": {
@@ -1968,7 +1992,7 @@
"runtime_peak_vram_mb": 21450.0,
"load_peak_allocated_mb": 6499.0,
"runtime_peak_allocated_mb": 14775.0,
"estimated_full_test_time_s": 181.8
"estimated_full_test_time_s": 576.3
},
"wan2_1_t2v_1_3b_lora_1gpu": {
"stages_ms": {
@@ -2110,7 +2134,7 @@
"runtime_peak_vram_mb": 51132.0,
"load_peak_allocated_mb": 34407.0,
"runtime_peak_allocated_mb": 44992.0,
"estimated_full_test_time_s": 248.2
"estimated_full_test_time_s": 509.3
},
"flux_2_image_t2i_2_gpus": {
"stages_ms": {
@@ -2181,7 +2205,7 @@
"runtime_peak_vram_mb": 38262.0,
"load_peak_allocated_mb": 32851.0,
"runtime_peak_allocated_mb": 34824.0,
"estimated_full_test_time_s": 135.3
"estimated_full_test_time_s": 75.0
},
"qwen_image_edit_2511_ti2i": {
"stages_ms": {
@@ -2271,7 +2295,7 @@
"runtime_peak_vram_mb": 12634.0,
"load_peak_allocated_mb": 7033.0,
"runtime_peak_allocated_mb": 9311.0,
"estimated_full_test_time_s": 122.7
"estimated_full_test_time_s": 61.3
},
"hunyuan3d_shape_gen": {
"stages_ms": {
@@ -2676,7 +2700,7 @@
"runtime_peak_vram_mb": 49390.0,
"load_peak_allocated_mb": 45460.0,
"runtime_peak_allocated_mb": 47725.0,
"estimated_full_test_time_s": 144.2
"estimated_full_test_time_s": 167.9
},
"ltx_2.3_two_stage_t2v_2gpus": {
"stages_ms": {
@@ -2737,7 +2761,7 @@
"runtime_peak_vram_mb": 59544.0,
"load_peak_allocated_mb": 46411.0,
"runtime_peak_allocated_mb": 56951.0,
"estimated_full_test_time_s": 160.0
"estimated_full_test_time_s": 186.0
},
"ltx_2_3_two_stage_ti2v_2gpus": {
"stages_ms": {
@@ -2798,7 +2822,7 @@
"runtime_peak_vram_mb": 60286.0,
"load_peak_allocated_mb": 46411.0,
"runtime_peak_allocated_mb": 56951.0,
"estimated_full_test_time_s": 170.0
"estimated_full_test_time_s": 101.4
},
"longlive2_t2v": {
"stages_ms": {
@@ -3002,7 +3026,7 @@
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 253.9
"estimated_full_test_time_s": 186.4
},
"wan2_1_t2v_1_3b_cache_dit_sp_only_2gpu": {
"stages_ms": {},
@@ -3010,7 +3034,7 @@
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 55.4
"estimated_full_test_time_s": 52.2
},
"minimax_h3_ref2va_video_audio_2gpu_h100": {
"stages_ms": {},
@@ -3018,7 +3042,7 @@
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 266.7
"estimated_full_test_time_s": 170.9
},
"minimax_h3_t2va_2gpu_h100": {
"stages_ms": {
@@ -3049,7 +3073,7 @@
"runtime_peak_vram_mb": 63312.0,
"load_peak_allocated_mb": 15139.0,
"runtime_peak_allocated_mb": 22887.0,
"estimated_full_test_time_s": 208.0
"estimated_full_test_time_s": 103.9
},
"mova_360p_tp2": {
"stages_ms": {},
@@ -3057,7 +3081,7 @@
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 160.2
"estimated_full_test_time_s": 141.9
},
"mova_360p_ring1_uly2": {
"stages_ms": {},
@@ -3065,7 +3089,7 @@
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 236.8
"estimated_full_test_time_s": 211.8
},
"zimage_image_t2i_2_gpus_non_square": {
"stages_ms": {},
@@ -3073,7 +3097,7 @@
"expected_e2e_ms": 0.0,
"expected_avg_denoise_ms": 0.0,
"expected_median_denoise_ms": 0.0,
"estimated_full_test_time_s": 67.9
"estimated_full_test_time_s": 40.5
},
"flux1_modelopt_fp8_t2i": {
"stages_ms": {},
python/sglang/multimodal_gen/test/unit/test_suite_partitioning.py
+21
View File
@@ -5,6 +5,8 @@ suite when standalone files outnumber the shards, and the shards must agree on
who runs what without talking to each other.
"""
import json
from pathlib import Path
from types import SimpleNamespace
import pytest
@@ -21,6 +23,10 @@ from sglang.multimodal_gen.test.server.gpu_cases import (
TWO_GPU_CASES,
)
_H100_BASELINE_PATH = (
Path(__file__).resolve().parents[1] / "server/perf_baselines/h100.json"
)
def _items(*est_times: float) -> list[PartitionItem]:
return [
@@ -111,6 +117,21 @@ def test_failing_cases_do_not_skip_the_shards_standalone_files(monkeypatch, tmp_
assert exit_code == 1
def test_two_gpu_cases_have_h100_full_test_time_estimates():
"""Every 2-gpu case must have an h100 estimated_full_test_time_s.
NVIDIA LPT sharding falls back to 300s when the field is missing, which
unbalances the three 2-gpu partitions.
"""
scenarios = json.loads(_H100_BASELINE_PATH.read_text())["scenarios"]
missing = [
case.id
for case in TWO_GPU_CASES
if scenarios.get(case.id, {}).get("estimated_full_test_time_s") is None
]
assert missing == []
def test_qwen_quality_variants_use_the_same_generation_request():
cases = {case.id: case for case in TWO_GPU_CASES}
lossless = cases["qwen_image_t2i_2_gpus"].sampling_params