[diffusion] CI: add minimax-h3 2-gpu consistency coverage (#33281)

python/sglang/multimodal_gen/test/server/gpu_cases.py

@@ -658,6 +658,61 @@ MINIMAX_H3_FOUR_GPU_H100_CASES = [
 ]
 
 TWO_GPU_CASES = [
+    DiffusionTestCase(
+        "minimax_h3_t2va_2gpu_h100",
+        DiffusionServerArgs(
+            model_path="MiniMaxAI/MiniMax-H3",
+            modality="video",
+            tp_size=2,
+            ulysses_degree=1,
+            extras=[
+                "--model-variant",
+                "fl2va",
+                "--performance-mode",
+                "memory",
+                "--layerwise-offload-components",
+                "dit,text_encoder,vae",
+                "--dit-offload-prefetch-size",
+                "1",
+                "--dit-layerwise-resident-layers",
+                "20",
+                "--enable-torch-compile",
+                "false",
+            ],
+        ),
+        DiffusionSamplingParams(
+            prompt=(
+                "A static night view of a narrow London alley in soft rain, wet "
+                "pavement reflecting a yellow streetlamp, the blue K. West sign "
+                "glowing above a doorway, cardboard boxes near the wall, a pale "
+                "parked car in the distance, and a slender glam-rock figure "
+                "holding a guitar under the lamp, brick storefronts, muted teal "
+                "and amber colors, subtle rain shimmer only."
+            ),
+            output_size="1344x768",
+            seconds=4,
+            output_format="mp4",
+            num_outputs_per_prompt=1,
+            extras={
+                "task": "t2va",
+                "conditions": [],
+                "target": {
+                    "short_edge": 768,
+                    "aspect_ratio": "16:9",
+                    "duration_seconds": 4.0,
+                },
+                "num_inference_steps": 8,
+                "flow_shift": 12.0,
+                "audio_flow_shift": 3.0,
+                "seed": 42,
+            },
+        ),
+        run_perf_check=True,
+        run_consistency_check=True,
+        run_component_accuracy_check=False,
+        run_models_api_check=False,
+        run_t2v_input_reference_check=False,
+    ),
     DiffusionTestCase(
         "flux2_modelopt_fp8_tp2_t2i",
         DiffusionServerArgs(

python/sglang/multimodal_gen/test/server/perf_baselines/h100.json

@@ -2790,6 +2790,33 @@
             "expected_median_denoise_ms": 0.0,
             "estimated_full_test_time_s": 55.4
         },
+        "minimax_h3_t2va_2gpu_h100": {
+            "stages_ms": {
+                "InputValidationStage": 0.05,
+                "MiniMaxH3PartitionAdmissionStage": 0.03,
+                "MiniMaxH3TextEncodingStage": 620.13,
+                "MiniMaxH3VisualEncodingStage": 0.11,
+                "MiniMaxH3AudioEncodingStage": 0.03,
+                "MiniMaxH3LatentPreparationStage": 18.72,
+                "MiniMaxH3TimestepPreparationStage": 0.34,
+                "MiniMaxH3DenoisingStage": 16839.0,
+                "MiniMaxH3DecodingStage": 29741.66,
+                "per_frame_generation": null
+            },
+            "denoise_step_ms": {
+                "0": 211.7,
+                "1": 2427.55,
+                "2": 2344.0,
+                "3": 2332.99,
+                "4": 2332.69,
+                "5": 2343.56,
+                "6": 2326.79
+            },
+            "expected_e2e_ms": 47538.57,
+            "expected_avg_denoise_ms": 2045.61,
+            "expected_median_denoise_ms": 2332.99,
+            "estimated_full_test_time_s": 235.9
+        },
         "mova_360p_tp2": {
             "stages_ms": {},
             "denoise_step_ms": {},

python/sglang/multimodal_gen/test/server/test_server_common.py

@@ -64,6 +64,7 @@ from sglang.multimodal_gen.test.test_utils import (
     load_action_consistency_gt,
     load_consistency_gt,
     save_consistency_failure_artifact,
+    save_missing_consistency_gt_artifact,
     wait_for_req_perf_record,
 )
 
@@ -610,6 +611,24 @@ class DiffusionServerBase:
         if not gt_exists(
             case.id, num_gpus, is_video=is_video, output_format=output_format
         ):
+            if is_video:
+                output_frames = pop_realtime_key_frames(case.id)
+                if output_frames is None:
+                    output_frames = extract_key_frames_from_video(content)
+            else:
+                output_frames = [image_bytes_to_numpy(content)]
+            artifact_path = save_missing_consistency_gt_artifact(
+                artifact_dir=os.environ.get("SGLANG_DIFFUSION_ARTIFACT_DIR"),
+                case_id=case.id,
+                num_gpus=num_gpus,
+                output_frames=output_frames,
+                is_video=is_video,
+                output_format=output_format,
+            )
+            if artifact_path is not None:
+                logger.info(
+                    "[Artifact] Saved missing consistency GT: %s", artifact_path
+                )
             if _get_consistency_gt_dir() is not None:
                 names = ", ".join(
                     get_consistency_gt_candidates(

python/sglang/multimodal_gen/test/test_utils.py

@@ -39,12 +39,12 @@ logger = init_logger(__name__)
 # NPU/ascend) is read from sgl-project/ci-data-diffusion, where the GT-gen workflows
 # publish.
 SGL_TEST_FILES_CI_DATA_REPO = "sgl-project/ci-data-diffusion"
-SGL_TEST_FILES_CI_DATA_REVISION = "6d689f4833e8c106ff0d853865f50496d1f2b453"
+SGL_TEST_FILES_CI_DATA_REVISION = "739c6c9b7cb972149cc3472cf19fe4bb29cf15c3"
 
 # The NPU pin is kept as a separate branch so ascend GT can be bumped independently
 # when it's regenerated on its own cadence.
 if current_platform.is_npu():
-    SGL_TEST_FILES_CI_DATA_REVISION = "6d689f4833e8c106ff0d853865f50496d1f2b453"
+    SGL_TEST_FILES_CI_DATA_REVISION = "d180ad38872dff3d1ad03e4610cffcda874d3eb8"
 
 SGL_TEST_FILES_CONSISTENCY_GT_ROOT = (
     "https://raw.githubusercontent.com/"
@@ -2003,6 +2003,30 @@ def _save_generated_artifact_images(
     return generated_files
 
 
+def save_missing_consistency_gt_artifact(
+    artifact_dir: str | Path | None,
+    case_id: str,
+    num_gpus: int,
+    output_frames: list[np.ndarray],
+    is_video: bool,
+    output_format: str | None = None,
+) -> Path | None:
+    if not artifact_dir:
+        return None
+
+    out_dir = Path(artifact_dir) / "missing_consistency_gt"
+    out_dir.mkdir(parents=True, exist_ok=True)
+    filenames = _consistency_gt_filenames(
+        case_id,
+        num_gpus,
+        is_video=is_video,
+        output_format=output_format,
+    )
+    for frame, filename in zip(output_frames, filenames):
+        Image.fromarray(_ensure_rgb_uint8_image(frame)).save(out_dir / filename)
+    return out_dir
+
+
 def _write_consistency_failure_index(
     out_dir: Path,
     records: list[dict[str, Any]],