[CI/Build] Share nightly CUDA YAML between H100/L4 and B200 (#7028)

Signed-off-by: wangyu <410167048@qq.com>
Signed-off-by: [Your Name] [Your Email]
Signed-off-by: [Your Name] <your.email@example.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: amy-why-3459 <wuhaiyan17@huawei.com>

tests/e2e/accuracy/helpers.py

@@ -135,6 +135,32 @@ def resolve_similarity_thresholds(
     return table.get(profile, table[DEFAULT_DEVICE_PROFILE])
 
 
+def resolve_device_threshold(
+    thresholds: float | int | Mapping[str, float],
+    *,
+    device_name: str | None = None,
+    label: str = "threshold",
+) -> tuple[str, float]:
+    """Resolve a scalar or per-GPU threshold for the current CUDA device.
+
+    Returns ``(profile_key, threshold)``. A bare number uses profile ``default``.
+    A mapping matches keys as substrings of the device name via
+    ``resolve_device_profile`` (for example ``{"H100": 0.3, "B200": 0.36}``).
+    Prefer an explicit ``default`` entry when unmatched GPUs should fall back;
+    otherwise unmatched devices raise.
+    """
+    if isinstance(thresholds, float | int):
+        return DEFAULT_DEVICE_PROFILE, float(thresholds)
+
+    profile = resolve_device_profile(device_name, profiles=thresholds)
+    if profile in thresholds:
+        return profile, float(thresholds[profile])
+    if DEFAULT_DEVICE_PROFILE in thresholds:
+        return DEFAULT_DEVICE_PROFILE, float(thresholds[DEFAULT_DEVICE_PROFILE])
+    known = ", ".join(str(key) for key in thresholds)
+    raise RuntimeError(f"no {label} entry for device profile {profile!r}; known: {known}")
+
+
 _SSIM_RE = re.compile(r"All:(?P<score>[0-9.]+)")
 _PSNR_RE = re.compile(r"average:(?P<score>[0-9.]+)")
 

tests/e2e/accuracy/hunyuanvideo15_i2v/test_hunyuanvideo15_i2v_video_similarity.py

@@ -183,7 +183,7 @@ def _generate_online_video(
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_hunyuanvideo15_i2v_diffusers_offline_generates_video(
     hunyuanvideo15_i2v_image_source: str | None,
 ) -> None:
@@ -203,7 +203,7 @@ def test_hunyuanvideo15_i2v_diffusers_offline_generates_video(
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", SERVER_CASES, indirect=True)
 def test_hunyuanvideo15_i2v_online_serving_generates_video(
     omni_server,
@@ -229,7 +229,7 @@ def test_hunyuanvideo15_i2v_online_serving_generates_video(
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_hunyuanvideo15_i2v_serving_matches_offline_video_similarity(
     hunyuanvideo15_i2v_image_source: str | None,
 ) -> None:

tests/e2e/accuracy/hunyuanvideo15_t2v/test_hunyuanvideo15_t2v_video_similarity.py

@@ -144,7 +144,7 @@ def _generate_online_video(*, omni_server, online_client, timeout_seconds: int)
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_hunyuanvideo15_t2v_diffusers_offline_generates_video() -> None:
     if not torch.cuda.is_available():
         pytest.skip("HunyuanVideo-1.5 T2V offline accuracy test requires CUDA.")
@@ -160,7 +160,7 @@ def test_hunyuanvideo15_t2v_diffusers_offline_generates_video() -> None:
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", SERVER_CASES, indirect=True)
 def test_hunyuanvideo15_t2v_online_serving_generates_video(
     omni_server,
@@ -182,7 +182,7 @@ def test_hunyuanvideo15_t2v_online_serving_generates_video(
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_hunyuanvideo15_t2v_serving_matches_offline_video_similarity() -> None:
     if not torch.cuda.is_available():
         pytest.skip("HunyuanVideo-1.5 T2V video similarity test requires CUDA.")

tests/e2e/accuracy/minicpmo_4_5/test_minicpmo_4_5.py

@@ -141,7 +141,7 @@ def _inline_daily_omni_media(monkeypatch: pytest.MonkeyPatch) -> None:
 
 
 @requires_daily_omni_deps
-@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"], "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", daily_test_params, indirect=True)
 def test_minicpmo_4_5_daily_omni_accuracy_bench(omni_server) -> None:
     _require_vllm_cli()
@@ -174,7 +174,7 @@ def test_minicpmo_4_5_daily_omni_accuracy_bench(omni_server) -> None:
     assert _acc_bench.run_acc_benchmark(_acc_bench.parse_acc_benchmark_args(argv)) == 0
 
 
-@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"], "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", seed_test_params, indirect=True)
 def test_minicpmo_4_5_seed_tts_wer_bench(omni_server) -> None:
     _require_vllm_cli()
@@ -200,7 +200,7 @@ def test_minicpmo_4_5_seed_tts_wer_bench(omni_server) -> None:
     assert _acc_bench.run_acc_benchmark(_acc_bench.parse_acc_benchmark_args(argv)) == 0
 
 
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", duplex_accuracy_test_params, indirect=True)
 def test_minicpmo_4_5_duplex_seed_tts_wer_bench(omni_server) -> None:
     """Gate Seed-TTS WER through the explicit Realtime TTS contract."""

tests/e2e/accuracy/minimax_h3/test_minimax_h3_i2va_ref2va_similarity.py

@@ -138,8 +138,6 @@ def _server_args() -> list[str]:
         "--vae-parallel-mode",
         "tile",
         "--vae-use-tiling",
-        "--diffusion-attention-backend",
-        "FLASH_ATTN",
         "--stage-init-timeout",
         "1800",
         "--init-timeout",
@@ -228,7 +226,7 @@ def _assert_official_video(
     )
 
 
-@hardware_test(res={"cuda": "H100"}, num_cards=4)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=4)
 def test_minimax_h3_i2va_matches_official_reference(
     accuracy_artifact_root: Path,
 ) -> None:
@@ -304,7 +302,7 @@ def test_minimax_h3_i2va_matches_official_reference(
     )
 
 
-@hardware_test(res={"cuda": "H100"}, num_cards=4)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=4)
 def test_minimax_h3_ref2va_matches_official_reference(
     accuracy_artifact_root: Path,
 ) -> None:

tests/e2e/accuracy/qwen3_omni/test_qwen3_omni.py

@@ -110,7 +110,7 @@ def _daily_omni_hub_inline_media(monkeypatch: pytest.MonkeyPatch) -> None:
     monkeypatch.setenv("VLLM_WORKER_MULTIPROC_METHOD", "spawn")
 
 
-@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@hardware_test(res={"cuda": ["H100", "B200"], "rocm": "MI325"}, num_cards=2)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_qwen3_omni_daily_omni_accuracy_bench(omni_server) -> None:
     _require_vllm_cli()
@@ -122,7 +122,7 @@ def test_qwen3_omni_daily_omni_accuracy_bench(omni_server) -> None:
     assert _acc_bench.run_acc_benchmark(ns) == 0
 
 
-@hardware_test(res={"cuda": "H100", "rocm": "MI325"}, num_cards=2)
+@hardware_test(res={"cuda": ["H100", "B200"], "rocm": "MI325"}, num_cards=2)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_qwen3_omni_seed_tts_wer_bench(omni_server) -> None:
     _require_vllm_cli()

tests/e2e/accuracy/sana_video/test_sana_video_golden.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -44,7 +44,7 @@ SSIM_THRESHOLDS = {
 }
 
 pytestmark = [
-    pytest.mark.full_model,
+    pytest.mark.slow,
     pytest.mark.diffusion,
     pytest.mark.skipif(
         not GOLDEN_BASE_URL,
@@ -169,6 +169,7 @@ def _validate_metadata(payload: Any, manifest: dict[str, Any], variant: str, tas
 
 def _download_case(variant: str, task: str, output_dir: Path, filenames: set[str]) -> dict[str, Any]:
     output_dir.mkdir(parents=True, exist_ok=True)
+    assert GOLDEN_BASE_URL is not None
     base_url = f"{GOLDEN_BASE_URL.rstrip('/')}/{variant}/{task}"
     response = requests.get(f"{base_url}/manifest.json", timeout=60)
     response.raise_for_status()

tests/e2e/accuracy/sana_video/test_sana_video_i2v_alignment.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -11,7 +11,7 @@ import torch
 from tests.helpers.mark import hardware_marks
 
 pytestmark = [
-    pytest.mark.full_model,
+    pytest.mark.slow,
     pytest.mark.diffusion,
     pytest.mark.benchmark,
 ]

tests/e2e/accuracy/sana_video/test_sana_video_pipeline_alignment.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -7,7 +7,7 @@ import gc
 import os
 from dataclasses import dataclass
 from pathlib import Path
-from typing import Any
+from typing import Any, TypedDict
 
 import pytest
 import torch
@@ -16,7 +16,7 @@ import torch.nn.functional as F
 from tests.helpers.mark import hardware_marks
 
 pytestmark = [
-    pytest.mark.full_model,
+    pytest.mark.slow,
     pytest.mark.diffusion,
     pytest.mark.benchmark,
 ]
@@ -30,7 +30,23 @@ HEIGHT = 192
 WIDTH = 320
 NUM_FRAMES = 9
 
-VARIANTS = {
+
+class _VariantConfig(TypedDict):
+    model: str
+    revision: str
+    model_env: str
+    max_abs: float
+    relative_l2: float
+    cosine: float
+    cfg_max_abs: float
+    cfg_relative_l2: float
+    cfg_cosine: float
+    step_max_abs: float
+    step_relative_l2: float
+    step_cosine: float
+
+
+VARIANTS: dict[str, _VariantConfig] = {
     "480p": {
         "model": "Efficient-Large-Model/SANA-Video_2B_480p_diffusers",
         "revision": "fed3bce411c58a0f688a31afe8f52e61acc2b15f",

tests/e2e/accuracy/test_cosmos3_similarity.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -35,6 +35,7 @@ def _model_name() -> str:
         pytest.skip(f"Set {MODEL_ENV_VAR} to run Cosmos3 full-model smoke tests.")
     if not torch.cuda.is_available():
         pytest.skip("Cosmos3 full-model smoke tests require CUDA.")
+    assert model
     return model
 
 
@@ -56,7 +57,7 @@ def _image_data_url(image: Image.Image) -> str:
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_cosmos3_t2i_serving_smoke(accuracy_artifact_root: Path) -> None:
     output_dir = model_output_dir(accuracy_artifact_root, MODEL_ID)
     with OmniServer(_model_name(), _server_args(), use_omni=True) as server:
@@ -97,7 +98,7 @@ def test_cosmos3_t2i_serving_smoke(accuracy_artifact_root: Path) -> None:
     ],
 )
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_cosmos3_video_serving_smoke(
     accuracy_artifact_root: Path,
     name: str,

tests/e2e/accuracy/test_diffusers_backend_similarity.py

@@ -29,6 +29,7 @@ from tests.e2e.accuracy.helpers import (
     assert_similarity,
     env_to_apply_ftfy_mock_in_subproc,
     model_output_dir,
+    resolve_device_threshold,
 )
 from tests.e2e.accuracy.helpers import (
     parse_psnr_score as _parse_psnr_score,
@@ -49,16 +50,25 @@ from vllm_omni.diffusion.models.diffusers_adapter.pipeline_diffusers_adapter imp
 pytestmark = [pytest.mark.full_model, pytest.mark.diffusion]
 
 
-def _set_matched_attention_backend(pipe: DiffusionPipeline) -> None:
-    """Walk the same backend chain the omni server's diffusers adapter walks.
+def _flash_attempt_backends() -> list[str]:
+    """Backends in ``CUDA_FLASH_ATTENTION_BACKEND_ATTEMPTS`` that can actually launch.
 
-    The server side (--diffusion-load-format diffusers) resolves its attention
-    backend through the adapter's preference chain, skipping backends that are
-    unavailable on the image (e.g. the FA3 hub kernel has no build variant for
-    the image's torch — build 2953). The reference run must resolve to the
-    same backend or the similarity comparison measures kernel differences.
+    Those names are Hopper FA2/FA3. ``set_attention_backend`` only checks import,
+    so a newer GPU still "selects" them and then dies with "no kernel image".
+    Do not probe by launching: a missing cubin can leave the CUDA context dead.
+    Capability >= 10 is outside that wheel, including future SKUs in that family.
     """
-    for backend in CUDA_FLASH_ATTENTION_BACKEND_ATTEMPTS:
+    if not torch.cuda.is_available():
+        return []
+    major, _ = torch.cuda.get_device_capability()
+    if major >= 10:
+        return []
+    return list(CUDA_FLASH_ATTENTION_BACKEND_ATTEMPTS)
+
+
+def _set_matched_attention_backend(pipe: DiffusionPipeline) -> None:
+    """Pick a Diffusers attention backend the reference run can actually execute."""
+    for backend in _flash_attempt_backends():
         try:
             pipe.transformer.set_attention_backend(backend)
             return
@@ -297,7 +307,7 @@ def _run_diffusers_qwen_image(*, model: str, output_path: Path) -> tuple[Image.I
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("model_id", ["Qwen/Qwen-Image"])
 def test_diffusers_backend_t2i_matches_diffusers(model_id: str, accuracy_artifact_root: Path) -> None:
     output_dir = model_output_dir(accuracy_artifact_root, model_id + "-diffusers-backend")
@@ -336,7 +346,7 @@ def test_diffusers_backend_t2i_matches_diffusers(model_id: str, accuracy_artifac
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 @pytest.mark.parametrize(
     "model_id",
     [
@@ -363,7 +373,11 @@ def test_diffusers_backend_i2v_matches_diffusers(
         model=model_id, output_path=diffusers_path, conditioning_image=resized_image
     )
     diffusers_latency = diffusers_latency * 1000
-    latency_threshold_factor = 0.3
+    # H100 keeps the historical 30% slack. B200 measured ~35.8% (6008 vs 4425 ms).
+    gpu_key, latency_threshold_factor = resolve_device_threshold(
+        {"H100": 0.3, "B200": 0.36},
+        label="latency threshold factor",
+    )
     latency_threshold = diffusers_latency * (1 + latency_threshold_factor)
 
     ssim_output = _run_ffmpeg_similarity("ssim", vllm_path, diffusers_path)
@@ -372,7 +386,8 @@ def test_diffusers_backend_i2v_matches_diffusers(
     psnr_score = _parse_psnr_score(psnr_output)
     print(f"{model_id} latency metrics:")
     print(
-        f"  Latency={vllm_latency:.2f}ms, threshold<={latency_threshold:.2f}ms, diffusers latency={diffusers_latency:.2f}ms, lower is better"
+        f"  Latency={vllm_latency:.2f}ms, threshold<={latency_threshold:.2f}ms, "
+        f"diffusers latency={diffusers_latency:.2f}ms, slack={latency_threshold_factor:.0%} ({gpu_key}), lower is better"
     )
     print(f"{model_id} similarity metrics:")
     print(f"  SSIM: value={ssim_score:.6f}, threshold>={VIDEO_SSIM_THRESHOLD:.6f}, range=[-1, 1], higher is better")

tests/e2e/accuracy/test_hunyuan_image3.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 
@@ -12,6 +12,7 @@ import tempfile
 import time
 from dataclasses import dataclass
 from pathlib import Path
+from typing import Any, cast
 
 import pytest
 import requests
@@ -288,7 +289,7 @@ def _quant_accuracy_cases() -> list[pytest.ParameterSet]:
 
 
 def _make_config(enable_kv_reuse: bool, path: Path) -> None:
-    config = copy.deepcopy(_DEPLOY_CONFIG)
+    config = cast(dict[str, Any], copy.deepcopy(_DEPLOY_CONFIG))
     config["stages"][0]["omni_kv_config"] = {"need_send_cache": enable_kv_reuse}
     config["stages"][1]["omni_kv_config"] = {"need_recv_cache": enable_kv_reuse}
     path.write_text(yaml.dump(config, default_flow_style=False, sort_keys=False))
@@ -303,7 +304,7 @@ def _quant_tensor_parallel_size() -> int:
 
 
 def _make_quant_dit_config(path: Path) -> None:
-    config = copy.deepcopy(_QUANT_DIT_CONFIG)
+    config = cast(dict[str, Any], copy.deepcopy(_QUANT_DIT_CONFIG))
     config["stages"][0]["devices"] = _quant_devices()
     config["stages"][0]["parallel_config"]["tensor_parallel_size"] = _quant_tensor_parallel_size()
     path.write_text(yaml.dump(config, default_flow_style=False, sort_keys=False))
@@ -448,7 +449,7 @@ def test_image_to_image_alignment_online(accuracy_artifact_root: Path) -> None:
     """Online API test: same pipeline, same seed as offline → PSNR >= 10 dB."""
     if importlib.util.find_spec("FlagEmbedding") is None:
         raise ImportError("Missing dependency: FlagEmbedding\nInstall with: pip install FlagEmbedding")
-    from tabulate import tabulate
+    from tabulate import tabulate  # type: ignore[import-untyped]
 
     output_dir = model_output_dir(accuracy_artifact_root, MODEL_NAME + "-online")
 
@@ -556,7 +557,7 @@ def _run_dit_model(
             torch.accelerator.empty_cache()
 
 
-@hardware_test(res={"cuda": "H100"}, num_cards=8)
+@hardware_test(res={"cuda": "H200"}, num_cards=8)
 @pytest.mark.skipif(
     torch.accelerator.device_count() < AR_TP_SIZE + DIT_TP_SIZE,
     reason=f"Needs {AR_TP_SIZE + DIT_TP_SIZE}+ GPUs ({AR_TP_SIZE} AR + {DIT_TP_SIZE} DiT)",
@@ -564,7 +565,7 @@ def _run_dit_model(
 def test_image_to_image_alignment(accuracy_artifact_root: Path) -> None:
     if importlib.util.find_spec("FlagEmbedding") is None:
         raise ImportError("Missing dependency: FlagEmbedding\nInstall with: pip install FlagEmbedding")
-    from tabulate import tabulate  # lazy import
+    from tabulate import tabulate  # type: ignore[import-untyped]
 
     """KV reuse ON vs OFF: same pipeline, same seed → PSNR >= 10 dB."""
     output_dir = model_output_dir(accuracy_artifact_root, MODEL_NAME + "-offline-kv-reuse")

tests/e2e/accuracy/test_hunyuan_image3_pixel_accuracy.py

@@ -299,7 +299,7 @@ def _assert_against_baseline(image: Image.Image, label: str) -> None:
 
 
 @pytest.mark.full_model
-@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=4)
+@hardware_test(res={"cuda": ["H100", "B200"], "npu": "A3"}, num_cards=4)
 def test_hunyuan_image3_pixel_accuracy_online(accuracy_artifact_root: Path) -> None:
     model = _model_name()
     output_dir = model_output_dir(accuracy_artifact_root, MODEL_NAME)
@@ -314,7 +314,7 @@ def test_hunyuan_image3_pixel_accuracy_online(accuracy_artifact_root: Path) -> N
 
 
 @pytest.mark.full_model
-@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=4)
+@hardware_test(res={"cuda": ["H100", "B200"], "npu": "A3"}, num_cards=4)
 def test_hunyuan_image3_pixel_accuracy_offline(accuracy_artifact_root: Path) -> None:
     model = _model_name()
     output_dir = model_output_dir(accuracy_artifact_root, MODEL_NAME)

tests/e2e/accuracy/test_qwen_image.py

@@ -233,7 +233,7 @@ def _run_diffusers_qwen_image_2512(*, model: str, output_path: Path) -> Image.Im
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_qwen_image_matches_diffusers(accuracy_artifact_root: Path) -> None:
     model = _model_name()
     output_dir = model_output_dir(accuracy_artifact_root, MODEL_ID)
@@ -253,7 +253,7 @@ def test_qwen_image_matches_diffusers(accuracy_artifact_root: Path) -> None:
 
 
 @pytest.mark.benchmark
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_qwen_image_2512_matches_diffusers_pixelwise(accuracy_artifact_root: Path) -> None:
     model = _model_2512_name()
     output_dir = model_output_dir(accuracy_artifact_root, MODEL_2512_ID)

tests/e2e/accuracy/wan22_i2v/test_wan22_i2v_video_similarity.py

@@ -195,7 +195,7 @@ def _generate_offline_video(*, image_source: str) -> tuple[Path, Path]:
 
 @pytest.mark.benchmark
 @pytest.mark.full_model
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_wan22_i2v_diffusers_offline_generates_video(
     wan22_i2v_image_source: str | None,
 ) -> None:
@@ -217,7 +217,7 @@ def test_wan22_i2v_diffusers_offline_generates_video(
 
 @pytest.mark.benchmark
 @pytest.mark.full_model
-@hardware_test(res={"cuda": "H100"}, num_cards=2)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=2)
 @pytest.mark.parametrize("omni_server", SERVER_CASES, indirect=True)
 def test_wan22_i2v_online_serving_generates_video(
     omni_server,
@@ -244,7 +244,7 @@ def test_wan22_i2v_online_serving_generates_video(
 
 @pytest.mark.benchmark
 @pytest.mark.full_model
-@hardware_test(res={"cuda": "H100"}, num_cards=2)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=2)
 def test_wan22_i2v_serving_matches_diffusers_video_similarity(
     wan22_i2v_image_source: str | None,
 ) -> None:

tests/e2e/features/custom_pipeline/test_async_omni_collective_rpc.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """
 Regression tests for AsyncOmni collective_rpc in inline diffusion mode.

tests/e2e/features/custom_pipeline/test_async_omni_qwen_image_generate.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """E2E tests for AsyncOmni Qwen-Image generation with trajectory_* fields.
 

tests/e2e/features/custom_pipeline/test_worker_extension.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from __future__ import annotations
 

tests/e2e/offline_inference/test_device_postprocess_equivalence.py

@@ -83,7 +83,7 @@ def _as_uint8_frames(video: np.ndarray) -> np.ndarray:
 _MAX_DIFFERING_FRACTION = 0.35
 
 
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_device_postprocess_produces_equivalent_frames() -> None:
     float_path = _generate(False)
     device_path = _generate(True)

tests/e2e/offline_inference/test_dots_tts_expansion.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """E2E tests for dots.tts native AR offline inference."""
 
 from collections.abc import Mapping

tests/e2e/offline_inference/test_gepard_tts.py

@@ -246,7 +246,7 @@ def _transcribe_for(wav: torch.Tensor, tmp_dir: str, name: str, accepts) -> str:
         pytest.param(
             "The quick brown fox jumps over the lazy dog.",
             None,
-            marks=pytest.mark.full_model,
+            marks=pytest.mark.slow,
             id="second_text",
         ),
     ],
@@ -274,7 +274,7 @@ def test_gepard_offline_zero_shot(omni_runner, run_level: str, text: str, keywor
     assert keyword in transcript, f"expected {keyword!r} in transcript, got {transcript!r}"
 
 
-@pytest.mark.full_model
+@pytest.mark.slow
 @hardware_test(res={"cuda": "L4"}, num_cards=1)
 def test_gepard_offline_long_text_skips_repetition(omni_runner, run_level: str) -> None:
     """The other branch of the prompt layout, end to end.

tests/e2e/offline_inference/test_hunyuanimage3_i2t_expansion.py

@@ -57,7 +57,7 @@ def omni() -> Generator[Omni, None, None]:
         yield runner.omni
 
 
-@hardware_test(res={"cuda": "H100"}, num_cards=4)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=4)
 def test_i2t_generates_text(omni: Omni) -> None:
     """Verify I2T output's first 20 token IDs match the HF greedy baseline."""
     # Solid-color image keeps the input self-contained and reproducible.

tests/e2e/offline_inference/test_lingbot_world_v2.py

@@ -47,7 +47,7 @@ pytestmark = [
 ]
 
 
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 def test_lingbot_world_v2_tp1_one_block(omni_runner_handler: OmniRunnerHandler) -> None:
     """Load the real checkpoint and generate one causal block on one GPU."""
 

tests/e2e/offline_inference/test_longcat_video_avatar.py

@@ -1,7 +1,8 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 from collections.abc import Generator
+from typing import Any
 
 import numpy as np
 import pytest
@@ -105,15 +106,17 @@ def _generate_avatar_frames(
     num_frames: int = 5,
     sampling_extra_args: dict | None = None,
 ):
+    extra_info: dict[str, Any] = dict(additional_information or {})
+    extra_info.update({"stage": stage, "resolution": "480p"})
+    multi_modal_data: dict[str, Any] = {"audio": audio_path}
+    if image_path is not None:
+        multi_modal_data["image"] = image_path
     prompt = {
         "prompt": PROMPT,
         "negative_prompt": NEGATIVE_PROMPT,
-        "multi_modal_data": {"audio": audio_path},
-        "additional_information": additional_information or {},
+        "multi_modal_data": multi_modal_data,
+        "additional_information": extra_info,
     }
-    prompt["additional_information"].update({"stage": stage, "resolution": "480p"})
-    if image_path is not None:
-        prompt["multi_modal_data"]["image"] = image_path
 
     outputs = omni_runner.generate(
         [prompt],

tests/e2e/offline_inference/test_qwen_image_autoround_w4a16_expansion.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 """E2E test for Qwen-Image AutoRound W4A16 quantized inference.
 
 Verifies that the W4A16 quantized Qwen-Image checkpoint loads end-to-end
@@ -47,8 +47,9 @@ def _first_request_images(outputs) -> list:
     return req_out.images
 
 
+@pytest.mark.full_model
 @pytest.mark.diffusion
-@hardware_test(res={"cuda": "L4"})
+@hardware_test(res={"cuda": ["L4", "B200"]})
 def test_qwen_image_autoround_w4a16_load():
     """Load the W4A16 quantized Qwen-Image model and run a minimal generation.
 
@@ -68,8 +69,9 @@ def test_qwen_image_autoround_w4a16_load():
         assert arr.std() > 1.0, "Generated image appears blank (std ≈ 0)"
 
 
+@pytest.mark.full_model
 @pytest.mark.diffusion
-@hardware_test(res={"cuda": "L4"})
+@hardware_test(res={"cuda": ["L4", "B200"]})
 def test_qwen_image_autoround_w4a16_generate():
     """Full generation: 512×512, 20 steps, CFG=5.0.
 

tests/e2e/offline_inference/test_voxcpm2_tts.py

@@ -1,3 +1,6 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
+
 """E2E test for VoxCPM2 native AR offline inference."""
 
 import os

tests/e2e/offline_inference/test_wan22_autoround_w4a16_expansion.py

@@ -156,7 +156,7 @@ def _generate_t2v_video(offline_client, prompt: str = "A cat sitting on a table"
 # ------------------------------------------------------------------
 
 
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 @pytest.mark.parametrize("omni_runner", quant_i2v_params, indirect=True)
 def test_wan22_i2v_autoround_w4a16_generates_video(omni_runner, offline_client):
     """Load the W4A16 quantized Wan2.2 I2V model and verify it produces a valid video."""
@@ -180,7 +180,7 @@ def test_wan22_i2v_autoround_w4a16_generates_video(omni_runner, offline_client):
 # ------------------------------------------------------------------
 
 
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 @pytest.mark.parametrize("omni_runner", quant_t2v_params, indirect=True)
 def test_wan22_t2v_autoround_w4a16_generates_video(omni_runner, offline_client):
     """Load the W4A16 quantized Wan2.2 T2V model and verify it produces a valid video."""
@@ -202,7 +202,7 @@ def test_wan22_t2v_autoround_w4a16_generates_video(omni_runner, offline_client):
 # ------------------------------------------------------------------
 
 
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 @pytest.mark.parametrize("omni_runner", quant_i2v_params, indirect=True)
 def test_wan22_i2v_autoround_w4a16_quant_peak(omni_runner, offline_client):
     """Measure peak GPU memory of W4A16 quantized I2V model."""
@@ -218,7 +218,7 @@ def test_wan22_i2v_autoround_w4a16_quant_peak(omni_runner, offline_client):
 # ------------------------------------------------------------------
 
 
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 @pytest.mark.parametrize("omni_runner", baseline_i2v_params, indirect=True)
 def test_wan22_i2v_autoround_w4a16_baseline_peak(omni_runner, offline_client):
     """Measure peak GPU memory of BF16 baseline I2V model."""
@@ -234,7 +234,7 @@ def test_wan22_i2v_autoround_w4a16_baseline_peak(omni_runner, offline_client):
 # ------------------------------------------------------------------
 
 
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 def test_wan22_i2v_autoround_w4a16_memory_savings():
     """Assert quantized I2V model uses meaningfully less memory than BF16 baseline."""
     quant_peak = _memory_results["quant_i2v"]
@@ -259,7 +259,7 @@ def test_wan22_i2v_autoround_w4a16_memory_savings():
 # ------------------------------------------------------------------
 
 
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 @pytest.mark.parametrize("omni_runner", quant_t2v_params, indirect=True)
 def test_wan22_t2v_autoround_w4a16_quant_peak(omni_runner, offline_client):
     """Measure peak GPU memory of W4A16 quantized T2V model."""
@@ -275,7 +275,7 @@ def test_wan22_t2v_autoround_w4a16_quant_peak(omni_runner, offline_client):
 # ------------------------------------------------------------------
 
 
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 @pytest.mark.parametrize("omni_runner", baseline_t2v_params, indirect=True)
 def test_wan22_t2v_autoround_w4a16_baseline_peak(omni_runner, offline_client):
     """Measure peak GPU memory of BF16 baseline T2V model."""
@@ -291,7 +291,7 @@ def test_wan22_t2v_autoround_w4a16_baseline_peak(omni_runner, offline_client):
 # ------------------------------------------------------------------
 
 
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 def test_wan22_t2v_autoround_w4a16_memory_savings():
     """Assert quantized T2V model uses meaningfully less memory than BF16 baseline."""
     quant_peak = _memory_results["quant_t2v"]

tests/e2e/online_serving/test_hunyuan_video_15_expansion.py

@@ -33,19 +33,13 @@ NEGATIVE_PROMPT = "low quality, blurry, distorted"
 
 MODEL = "hunyuanvideo-community/HunyuanVideo-1.5-Diffusers-480p_t2v"
 
-SINGLE_CARD_MARKS = hardware_marks(res={"cuda": "H100"})
-PARALLEL_MARKS = hardware_marks(res={"cuda": "H100"}, num_cards=2)
-
-# Only the most basic single-card deployment row is cheap enough for PR (L2) / merge (L3)
-# CI; the CacheDiT / parallel feature combinations run nightly (L4).
-_CORE_MARKS = [pytest.mark.core_model, pytest.mark.advanced_model]
-_FULL_MARKS = [pytest.mark.full_model]
-
 
 def _get_diffusion_feature_cases(model: str):
     """Return diffusion feature cases for HunyuanVideo-1.5.
 
     Designed for 2x H100 environment per issue #1832.
+    Only the CPU-offload row is cheap enough for PR (L2) / merge (L3);
+    CacheDiT / parallel combinations run nightly (L4).
     """
     return [
         # (1 GPU) CPU offload
@@ -57,7 +51,11 @@ def _get_diffusion_feature_cases(model: str):
                 ],
             ),
             id="single_card_cpu_offload",
-            marks=SINGLE_CARD_MARKS + _CORE_MARKS,
+            marks=[
+                *hardware_marks(res={"cuda": "H100"}),
+                pytest.mark.core_model,
+                pytest.mark.advanced_model,
+            ],
         ),
         # (1 GPU) CacheDiT + Layerwise CPU offloading
         pytest.param(
@@ -70,7 +68,10 @@ def _get_diffusion_feature_cases(model: str):
                 ],
             ),
             id="single_card_cachedit_layerwise",
-            marks=SINGLE_CARD_MARKS + _FULL_MARKS,
+            marks=[
+                *hardware_marks(res={"cuda": ["H100", "B200"]}),
+                pytest.mark.full_model,
+            ],
         ),
         # (2 GPUs) CacheDiT + TP=2 + VAE patch parallel=2
         pytest.param(
@@ -87,7 +88,10 @@ def _get_diffusion_feature_cases(model: str):
                 ],
             ),
             id="parallel_cachedit_tp2_vae2",
-            marks=PARALLEL_MARKS + _FULL_MARKS,
+            marks=[
+                *hardware_marks(res={"cuda": ["H100", "B200"]}, num_cards=2),
+                pytest.mark.full_model,
+            ],
         ),
     ]
 

tests/e2e/online_serving/test_lingbot_video.py

@@ -20,7 +20,7 @@ PROMPT = "a robotic arm picks up a red block"
 NEGATIVE_PROMPT = "low quality, blurry, watermark, text"
 DEFAULT_SAMPLING_PARAMS = '{"0":{"num_frames":81,"num_inference_steps":40,"guidance_scale":6.0}}'
 
-SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"})
+SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": ["H100", "B200"]})
 
 
 def _get_diffusion_feature_cases(model: str):

tests/e2e/online_serving/test_lingbot_video_expansion.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """
 L4 expansion coverage for ``robbyant/lingbot-video-dense-1.3b``.
@@ -26,7 +26,7 @@ MODEL = "robbyant/lingbot-video-dense-1.3b"
 PROMPT = "a robotic arm picks up a red block"
 NEGATIVE_PROMPT = "low quality, blurry, watermark, text"
 
-SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"})
+SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": ["H100", "B200"]})
 
 
 def _get_diffusion_feature_cases(model: str):

tests/e2e/online_serving/test_lingbot_video_moe.py

@@ -20,7 +20,7 @@ PROMPT = "a robotic arm picks up a red block"
 NEGATIVE_PROMPT = "low quality, blurry, watermark, text"
 SMOKE_DEFAULT_SAMPLING_PARAMS = '{"0":{"num_frames":81,"num_inference_steps":2,"guidance_scale":3.0}}'
 
-SINGLE_CARD_MARKS = hardware_marks(res={"cuda": "H100"})
+SINGLE_CARD_MARKS = hardware_marks(res={"cuda": ["H100", "B200"]})
 
 
 def _get_server_cases(model: str):

tests/e2e/online_serving/test_minicpmo_4_5_duplex_expansion.py

@@ -51,7 +51,7 @@ def _cached_server_vad_artifact() -> None:
     )
 
 
-@hardware_test(res={"cuda": "H100", "npu": "A3"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"], "npu": "A3"}, num_cards=1)
 @pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
 def test_duplex_admission_and_expiry_reaper(omni_server, tmp_path: Path) -> None:
     args = multi_session_args(
@@ -79,7 +79,7 @@ def test_duplex_admission_and_expiry_reaper(omni_server, tmp_path: Path) -> None
 # still talking, which only holds when the duplex pipeline sustains real-time
 # throughput. The current NPU stack runs several times slower than real time,
 # so it never reaches a mid-stream decision point.
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
 def test_duplex_soft_interrupt(omni_server, tmp_path: Path) -> None:
     input_wav = validated_soft_interrupt_wav()
@@ -112,7 +112,7 @@ def test_duplex_soft_interrupt(omni_server, tmp_path: Path) -> None:
     assert result["followup_response_transcript_ok"] is True
 
 
-@hardware_test(res={"cuda": "H100"}, num_cards=1)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", SERVER_PARAMS, indirect=True)
 def test_duplex_server_vad_hard_interrupt(_cached_server_vad_artifact, omni_server) -> None:
     result = asyncio.run(

tests/e2e/online_serving/test_pi0_expansion.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """E2E online serving test for π0 (Pi-Zero) OpenPI websocket serving.
 
@@ -58,7 +58,7 @@ test_params = [
 
 @pytest.mark.full_model
 @pytest.mark.diffusion
-@hardware_test(res={"cuda": "H100"})
+@hardware_test(res={"cuda": ["H100", "B200"]})
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_pi0_openpi_online(omni_server):
     try:

tests/e2e/online_serving/test_qwen3_omni_multi_replicas.py

@@ -91,7 +91,7 @@ def _assert_audio_outputs(responses: list[OmniResponse], *, expect_text: bool) -
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100"}, num_cards=4)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=4)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_text_only_batch_uses_multi_replica_talker(omni_server, online_client) -> None:
     request_config = {
@@ -108,7 +108,7 @@ def test_text_only_batch_uses_multi_replica_talker(omni_server, online_client) -
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100"}, num_cards=4)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=4)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_text_to_audio_stream_batch_uses_multi_replica_vocoder(omni_server, online_client) -> None:
     request_config = {
@@ -125,7 +125,7 @@ def test_text_to_audio_stream_batch_uses_multi_replica_vocoder(omni_server, onli
 
 @pytest.mark.full_model
 @pytest.mark.omni
-@hardware_test(res={"cuda": "H100"}, num_cards=4)
+@hardware_test(res={"cuda": ["H100", "B200"]}, num_cards=4)
 @pytest.mark.parametrize("omni_server", test_params, indirect=True)
 def test_mixed_modal_stream_batch_generates_text_and_audio(omni_server, online_client) -> None:
     request_config = {

tests/e2e/online_serving/test_qwen3_tts_base_expansion.py

@@ -69,7 +69,7 @@ tts_no_async_chunk_server_params = [
 DEFAULT_AUDIO_SPEECH_TIMEOUT_S = 180.0
 
 
-@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@hardware_test(res={"cuda": ["L4", "B200"]}, num_cards=1)
 @pytest.mark.parametrize(
     "omni_server",
     tts_async_chunk_server_params + tts_no_async_chunk_server_params,
@@ -99,7 +99,7 @@ def test_voice_clone_streaming_001(omni_server, online_client) -> None:
     online_client.send_audio_speech_request(request_config, request_num=get_max_batch_size("few"))
 
 
-@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@hardware_test(res={"cuda": ["L4", "B200"]}, num_cards=1)
 @pytest.mark.parametrize(
     "omni_server",
     tts_async_chunk_server_params + tts_no_async_chunk_server_params,
@@ -127,7 +127,7 @@ def test_response_format_001(omni_server, online_client) -> None:
     online_client.send_audio_speech_request(request_config)
 
 
-@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@hardware_test(res={"cuda": ["L4", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", tts_async_chunk_server_params, indirect=True)
 def test_xvector_then_icl_same_ref_audio_keeps_engine_alive(omni_server, online_client) -> None:
     """Regression for #5049: x-vector-only then ICL on the same ``ref_audio``.

tests/e2e/online_serving/test_qwen3_tts_customvoice_expansion.py

@@ -60,7 +60,7 @@ tts_server_params = [
 ]
 
 
-@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@hardware_test(res={"cuda": ["L4", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
 def test_voice_001(omni_server, online_client) -> None:
     """
@@ -92,7 +92,7 @@ def test_voice_001(omni_server, online_client) -> None:
             raise
 
 
-@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@hardware_test(res={"cuda": ["L4", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
 def test_voice_002(omni_server, online_client) -> None:
     """
@@ -115,7 +115,7 @@ def test_voice_002(omni_server, online_client) -> None:
     online_client.send_audio_speech_request(request_config)
 
 
-@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@hardware_test(res={"cuda": ["L4", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
 def test_voice_003(omni_server, online_client) -> None:
     """
@@ -138,7 +138,7 @@ def test_voice_003(omni_server, online_client) -> None:
     online_client.send_audio_speech_request(request_config)
 
 
-@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@hardware_test(res={"cuda": ["L4", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
 def test_sample_rate_001(omni_server, online_client) -> None:
     """The speech API resamples Qwen3-TTS output from native 24 kHz to 8 kHz."""
@@ -155,7 +155,7 @@ def test_sample_rate_001(omni_server, online_client) -> None:
     online_client.send_audio_speech_request(request_config)
 
 
-@hardware_test(res={"cuda": "L4"}, num_cards=1)
+@hardware_test(res={"cuda": ["L4", "B200"]}, num_cards=1)
 @pytest.mark.parametrize("omni_server", tts_server_params, indirect=True)
 def test_language_001(omni_server, online_client) -> None:
     """

tests/e2e/online_serving/test_qwen_image_expansion.py

@@ -21,8 +21,8 @@ pytestmark = [pytest.mark.diffusion, pytest.mark.full_model]
 
 T2I_PROMPT = "A photo of a cat sitting on a laptop keyboard, digital art style."
 NEGATIVE_PROMPT = "blurry, low quality"
-SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"})
-PARALLEL_FEATURE_MARKS = hardware_marks(res={"cuda": "H100"}, num_cards=2)
+SINGLE_CARD_FEATURE_MARKS = hardware_marks(res={"cuda": ["H100", "B200"]})
+PARALLEL_FEATURE_MARKS = hardware_marks(res={"cuda": ["H100", "B200"]}, num_cards=2)
 
 MODEL_IMAGE = "Qwen/Qwen-Image"
 MODEL_2512 = "Qwen/Qwen-Image-2512"

tests/e2e/online_serving/test_sana_wm.py

@@ -1,5 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM-Omni project
 
 """
 Online serving smoke for SANA-WM image-to-video (first-frame I2V via ``/v1/videos``).

tests/e2e/online_serving/test_wan22_expansion.py

@@ -30,8 +30,8 @@ PROMPT = "Two anthropomorphic cats in comfy boxing gear and bright gloves fight
 NEGATIVE_PROMPT = "low quality, blurry, distorted face, extra limbs, bad anatomy, watermark, logo, text, ugly, deformed, mutated, jpeg artifacts"
 
 # CUDA marks (original matrix, unchanged)
-CUDA_SINGLE_CARD_MARKS = hardware_marks(res={"cuda": "H100"})
-CUDA_PARALLEL_MARKS = hardware_marks(res={"cuda": "H100"}, num_cards=2)
+CUDA_SINGLE_CARD_MARKS = hardware_marks(res={"cuda": ["H100", "B200"]})
+CUDA_PARALLEL_MARKS = hardware_marks(res={"cuda": ["H100", "B200"]}, num_cards=2)
 
 # NPU marks
 NPU_TWO_CARD_MARKS = hardware_marks(res={"npu": "A2"}, num_cards=2)

tests/e2e/online_serving/test_zimage_expansion.py

@@ -19,7 +19,7 @@ import pytest
 from tests.helpers.mark import hardware_marks
 from tests.helpers.runtime import OmniServer, OmniServerParams, OnlineOmniClient
 
-pytestmark = [pytest.mark.diffusion, pytest.mark.full_model]
+pytestmark = [pytest.mark.diffusion, pytest.mark.slow]
 
 MODEL = "Tongyi-MAI/Z-Image-Turbo"
 PROMPT = "A high-detail studio photo of an orange tabby cat sitting on a laptop keyboard."