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

[diffusion] CI: tighten NVIDIA perf baselines (#35016)

Browse Source
This commit is contained in:
Mick
2026-08-16 20:54:50 +08:00
committed by GitHub
parent 41abbb0d32
commit d3589a7251
3 changed files with 1388 additions and 1401 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/multimodal_gen/test/server/perf_baselines/5090.json
+29 -29
View File
@@ -2,8 +2,8 @@
"metadata": {
"model": "Diffusion Server",
"hardware": "CI RTX 5090 pool",
"description": "Reference numbers captured from real 5090 CI runner history (PR 29791 run 28492141274, job 84450974283).",
"last_updated": "2026-07-01"
"description": "Reference numbers tightened from RTX 5090 multimodal_gen CI run 31935214180 at source revision 2ee0d38a8553d186faad9c50893480e3c1d79270.",
"last_updated": "2026-08-16"
},
"tolerances": {
"long_term": {
@@ -38,68 +38,68 @@
"flux_2_klein_base_image_t2i": {
"stages_ms": {
"DecodingStage": 18.19,
"DenoisingStage": 18213.48,
"DenoisingStage": 17612.9,
"ImageVAEEncodingStage": 0.01,
"InputValidationStage": 0.08,
"LatentPreparationStage": 4.87,
"LatentPreparationStage": 1.67,
"TextEncodingStage": 104.46,
"TimestepPreparationStage": 142.18
},
"denoise_step_ms": {
"0": 219.58,
"10": 359.79,
"20": 361.55,
"29": 361.99,
"39": 362.57,
"49": 362.89
"0": 95.07,
"10": 344.0,
"20": 345.85,
"29": 346.2,
"39": 346.9,
"49": 347.31
},
"expected_e2e_ms": 18641.09,
"expected_avg_denoise_ms": 357.93,
"expected_median_denoise_ms": 361.8,
"expected_e2e_ms": 18255.39,
"expected_avg_denoise_ms": 338.78,
"expected_median_denoise_ms": 345.95,
"estimated_full_test_time_s": 94.0
},
"wan2_1_t2v_1.3b": {
"stages_ms": {
"DecodingStage": 1202.43,
"DecodingStage": 495.01,
"DenoisingStage": 21452.59,
"InputValidationStage": 0.09,
"LatentPreparationStage": 0.23,
"TextEncodingStage": 1210.22,
"TextEncodingStage": 1025.11,
"TimestepPreparationStage": 3.91
},
"denoise_step_ms": {
"0": 480.39,
"0": 169.02,
"10": 426.11,
"20": 427.32,
"29": 428.09,
"39": 428.04,
"49": 422.31
},
"expected_e2e_ms": 23877.73,
"expected_avg_denoise_ms": 428.74,
"expected_e2e_ms": 23418.73,
"expected_avg_denoise_ms": 426.59,
"expected_median_denoise_ms": 427.94,
"estimated_full_test_time_s": 160.9
},
"zimage_image_t2i": {
"stages_ms": {
"DecodingStage": 7.11,
"DenoisingStage": 2229.48,
"DenoisingStage": 2178.8,
"InputValidationStage": 0.04,
"LatentPreparationStage": 0.17,
"TextEncodingStage": 252.73,
"TimestepPreparationStage": 38.57
"TextEncodingStage": 252.2,
"TimestepPreparationStage": 36.04
},
"denoise_step_ms": {
"0": 39.34,
"2": 277.35,
"3": 267.94,
"5": 279.01,
"6": 273.01,
"8": 275.12
"2": 270.3,
"3": 263.6,
"5": 264.28,
"6": 265.1,
"8": 265.56
},
"expected_e2e_ms": 2533.95,
"expected_avg_denoise_ms": 246.97,
"expected_median_denoise_ms": 273.01,
"expected_e2e_ms": 2492.09,
"expected_avg_denoise_ms": 241.53,
"expected_median_denoise_ms": 265.1,
"estimated_full_test_time_s": 329.8
},
"lingbot_video_moe_t2v": {
python/sglang/multimodal_gen/test/server/perf_baselines/h100.json
+1359 -1356
View File
File diff suppressed because it is too large Load Diff
python/sglang/multimodal_gen/test/server/test_server_common.py
-16
View File
@@ -158,22 +158,6 @@ def diffusion_server(case: DiffusionTestCase) -> ServerContext:
# picking another one (which causes the test client to connect to the wrong server).
extra_args += " --strict-ports"
# Shape-only mesh cases (e.g. hunyuan3d_shape_gen) validate geometry via
# mesh-correctness and must NOT run the paint/texture stages, whose
# verification checks texture artifacts (paint_mesh/normal_maps/renderer)
# that the shape-only path never produces. Inject a pipeline-config override
# disabling paint for these cases.
if server_args.custom_validator == "mesh":
import json as _json
import tempfile as _tempfile
_paint_off_cfg = os.path.join(
_tempfile.gettempdir(), f"{case.id}_paint_off.json"
)
with open(_paint_off_cfg, "w") as _f:
_json.dump({"paint_enable": False}, _f)
extra_args += f" --config {_paint_off_cfg}"
for arg in server_args.extras:
extra_args += f" {arg}"