103 lines
2.8 KiB
Python
103 lines
2.8 KiB
Python
"""
|
|
Tests of common diffusion feature combinations in online serving mode
|
|
for Z-Image.
|
|
|
|
Coverage is intentionally limited to the minimal 4xL4 cases that
|
|
exercise Z-Image's supported parallel feature combinations:
|
|
- CacheDiT + FP8 + Ring=2 + TP=2
|
|
- TeaCache + FP8 + Ulysses=2 + Ring=2
|
|
"""
|
|
|
|
import pytest
|
|
|
|
from tests.conftest import (
|
|
OmniServer,
|
|
OmniServerParams,
|
|
OpenAIClientHandler,
|
|
)
|
|
from tests.utils import hardware_marks
|
|
|
|
MODEL = "Tongyi-MAI/Z-Image-Turbo"
|
|
PROMPT = "A high-detail studio photo of an orange tabby cat sitting on a laptop keyboard."
|
|
|
|
FOUR_CARD_MARKS = hardware_marks(res={"cuda": "L4"}, num_cards=4)
|
|
|
|
|
|
def _get_diffusion_feature_cases():
|
|
"""Return the common Z-Image feature combinations that fit L4 CI."""
|
|
return [
|
|
pytest.param(
|
|
OmniServerParams(
|
|
model=MODEL,
|
|
server_args=[
|
|
"--cache-backend",
|
|
"cache_dit",
|
|
"--quantization",
|
|
"fp8",
|
|
"--ring",
|
|
"2",
|
|
"--tensor-parallel-size",
|
|
"2",
|
|
],
|
|
),
|
|
id="parallel_cachedit_fp8_ring2_tp2",
|
|
marks=FOUR_CARD_MARKS,
|
|
),
|
|
pytest.param(
|
|
OmniServerParams(
|
|
model=MODEL,
|
|
server_args=[
|
|
"--cache-backend",
|
|
"tea_cache",
|
|
"--quantization",
|
|
"fp8",
|
|
"--ulysses-degree",
|
|
"2",
|
|
"--ring",
|
|
"2",
|
|
],
|
|
),
|
|
id="parallel_teacache_fp8_ulysses2_ring2",
|
|
marks=FOUR_CARD_MARKS,
|
|
),
|
|
pytest.param(
|
|
OmniServerParams(
|
|
model=MODEL,
|
|
server_args=[
|
|
"--use-hsdp",
|
|
"--hsdp-shard-size",
|
|
"2",
|
|
],
|
|
),
|
|
id="parallel_hsdp",
|
|
marks=[*FOUR_CARD_MARKS, pytest.mark.skip(reason="issue #2435")],
|
|
),
|
|
]
|
|
|
|
|
|
@pytest.mark.advanced_model
|
|
@pytest.mark.diffusion
|
|
@pytest.mark.parametrize(
|
|
"omni_server",
|
|
_get_diffusion_feature_cases(),
|
|
indirect=True,
|
|
)
|
|
def test_zimage(
|
|
omni_server: OmniServer,
|
|
openai_client: OpenAIClientHandler,
|
|
):
|
|
"""Exercise supported Z-Image diffusion features in minimal CI cases."""
|
|
request_config = {
|
|
"model": omni_server.model,
|
|
"messages": [{"role": "user", "content": PROMPT}],
|
|
"extra_body": {
|
|
"height": 512,
|
|
"width": 512,
|
|
"num_inference_steps": 2,
|
|
"guidance_scale": 0.0,
|
|
"seed": 42,
|
|
},
|
|
}
|
|
|
|
openai_client.send_diffusion_request(request_config)
|