[Perf] Support request-level batching for Wan2.2 pipelines (#5676)

Signed-off-by: nagisa-kun <1434936049@qq.com>
Signed-off-by: nagisa <1434936049@qq.com>

tests/diffusion/models/wan2_2/test_wan22_pipeline_helpers.py

@@ -10,6 +10,7 @@ import torch
 import vllm_omni.diffusion.models.wan2_2.pipeline_wan2_2 as wan22_module
 from vllm_omni.diffusion.models.wan2_2.pipeline_wan2_2 import (
     create_transformer_from_config,
+    get_wan22_post_process_func,
     load_transformer_config,
     retrieve_latents,
 )
@@ -26,6 +27,17 @@ class _LatentDist:
         return torch.tensor([2.0])
 
 
+def test_wan22_postprocess_honors_request_output_type() -> None:
+    video = torch.zeros(1, 4, 1, 2, 2)
+
+    output = get_wan22_post_process_func(SimpleNamespace())(
+        video,
+        sampling_params=SimpleNamespace(output_type="latent"),
+    )
+
+    assert output is video
+
+
 def test_retrieve_latents_supports_sample_mode_argmax_and_direct_latents() -> None:
     generator = torch.Generator(device="cpu")