[diffusion] CI: warm up CI cases at the request's quality level (#42958)

This commit is contained in:
WenhaoZhang
2026-10-07 23:26:36 +08:00
committed by GitHub
parent 00b21d6415
commit 0b635266d4
2 changed files with 91 additions and 2 deletions
@@ -29,6 +29,7 @@ from pathlib import Path
from typing import Any, Sequence
from sglang.multimodal_gen.configs.pipeline_configs.base import ModelTaskType
from sglang.multimodal_gen.configs.sample.sampling_params import normalize_quality
from sglang.multimodal_gen.registry import (
get_model_info,
get_pipeline_config_classes,
@@ -381,6 +382,21 @@ class DiffusionTestCase:
extras={**self.sampling_params.extras, "quality": "exact"},
),
)
# Warmup must run the request's level: a request below the server default
# unmounts the fusions warmup mounted, so its timed run is the first pass
# over an unwarmed path.
request_quality = self.sampling_params.extras.get("quality")
if request_quality is not None:
object.__setattr__(
self,
"server_args",
replace(
self.server_args,
extras=_with_warmup_quality(
self.server_args.extras, normalize_quality(request_quality)
),
),
)
has_startup_lora = self.server_args.lora_path is not None
has_dynamic_lora = self.server_args.dynamic_lora_path is not None
@@ -875,6 +891,37 @@ HUNYUAN3D_SHAPE_sampling_params = DiffusionSamplingParams(
)
def _with_warmup_quality(extras: Sequence[str], quality: str) -> list[str]:
"""Return ``extras`` with ``quality`` in ``--warmup-sampling-params``.
Merges into an existing JSON object and keeps a quality the case set itself.
"""
option = "--warmup-sampling-params"
def with_quality(raw: str) -> str:
params = json.loads(raw)
params.setdefault("quality", quality)
return json.dumps(params)
merged: list[str] = []
found = False
for item in extras:
tokens = shlex.split(item)
changed = False
for index, token in enumerate(tokens):
if token.startswith(f"{option}="):
tokens[index] = f"{option}={with_quality(token[len(option) + 1 :])}"
changed = True
elif token == option and index + 1 < len(tokens):
tokens[index + 1] = with_quality(tokens[index + 1])
changed = True
merged.append(shlex.join(tokens) if changed else item)
found = found or changed
if not found:
merged.append(f"{option} {shlex.quote(json.dumps({'quality': quality}))}")
return merged
def _get_extra_arg_value(extras: Sequence[str], option_name: str) -> str | None:
tokens: list[str] = []
for item in extras:
@@ -13,6 +13,7 @@ Covers warmup and cfg-parallel guard paths introduced alongside this file:
All tests are CPU-only; no model loading, no distributed init.
"""
import json
import unittest
from collections import deque
from types import SimpleNamespace
@@ -32,7 +33,10 @@ from sglang.multimodal_gen.configs.pipeline_configs.sana_wm import SanaWMPipelin
from sglang.multimodal_gen.configs.sample.longlive2 import LongLive2SamplingParams
from sglang.multimodal_gen.configs.sample.ltx_2_5 import LTX25SamplingParams
from sglang.multimodal_gen.configs.sample.minimax_h3 import MiniMaxH3SamplingParams
from sglang.multimodal_gen.configs.sample.sampling_params import SamplingParams
from sglang.multimodal_gen.configs.sample.sampling_params import (
SamplingParams,
normalize_quality,
)
from sglang.multimodal_gen.configs.sample.sana_wm import SanaWMSamplingParams
from sglang.multimodal_gen.runtime.entrypoints.control_requests import (
SetLoraReq,
@@ -65,7 +69,10 @@ from sglang.multimodal_gen.runtime.warmup_request_builder import (
supports_synthetic_warmup,
)
from sglang.multimodal_gen.test.server.gpu_cases import ONE_GPU_CASES, TWO_GPU_CASES
from sglang.multimodal_gen.test.server.testcase_configs import _get_extra_arg_value
from sglang.multimodal_gen.test.server.testcase_configs import (
_get_extra_arg_value,
_with_warmup_quality,
)
def _make_bare_scheduler(enable_cfg_parallel: bool) -> Scheduler:
@@ -707,6 +714,41 @@ class TestWarmupReqCfgParallel(unittest.TestCase):
self.assertEqual((reqs[0].width, reqs[0].height), (384, 640))
self.assertEqual(reqs[0].num_frames, case.sampling_params.num_frames)
def test_ci_warmup_runs_the_request_quality(self):
"""A request below the server default level unmounts the fusions warmup
mounted, so its timed run takes an unwarmed path; warmup must match."""
for case in ONE_GPU_CASES + TWO_GPU_CASES:
request_quality = case.sampling_params.extras.get("quality")
if request_quality is None:
continue
warmup = _get_extra_arg_value(
case.server_args.extras, "--warmup-sampling-params"
)
self.assertIsNotNone(warmup, case.id)
self.assertEqual(
json.loads(warmup)["quality"],
normalize_quality(request_quality),
case.id,
)
def test_warmup_quality_merges_into_existing_warmup_params(self):
merged = _with_warmup_quality(
["--load-x", """--warmup-sampling-params '{"num_frames": 9}'"""],
"exact",
)
self.assertEqual(
json.loads(_get_extra_arg_value(merged, "--warmup-sampling-params")),
{"num_frames": 9, "quality": "exact"},
)
self.assertEqual(merged[0], "--load-x")
kept = _with_warmup_quality(
["""--warmup-sampling-params='{"quality": "lossless"}'"""], "exact"
)
self.assertEqual(
json.loads(_get_extra_arg_value(kept, "--warmup-sampling-params")),
{"quality": "lossless"},
)
def test_ltx25_ci_warmup_matches_formal_decoder_and_shape(self):
case = next(
case