rebase: adapt to vLLM v0.29 APIs

Signed-off-by: tzhouam <tzhouam@connect.ust.hk>

tests/diffusion/layers/test_rope_broadcast.py

@@ -16,7 +16,11 @@ import types
 import pytest
 import torch
 
-from vllm_omni.diffusion.layers.rope import RotaryEmbedding, apply_rotary_emb_mindiesd
+from vllm_omni.diffusion.layers.rope import (
+    RotaryEmbedding,
+    RotaryEmbeddingWan,
+    apply_rotary_emb_mindiesd,
+)
 
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
@@ -239,6 +243,25 @@ class TestRotaryEmbeddingNativeShapeRegression:
         assert not torch.isnan(output).any()
 
 
+def test_cuda_paths_fall_back_without_vendored_rotary_kernel() -> None:
+    """Source/CPU installs need not contain the wheel-vendored CUDA kernel."""
+    x = torch.randn(1, 5, 2, 8)
+
+    rope = RotaryEmbedding(is_neox_style=True)
+    rope.apply_rotary_emb_vllm_flash_attn = None
+    output = rope.forward_cuda(x, torch.randn(5, 4), torch.randn(5, 4))
+    assert output.shape == x.shape
+
+    wan_rope = RotaryEmbeddingWan(is_neox_style=False, half_head_dim=True)
+    wan_rope.apply_rotary_emb_vllm_flash_attn = None
+    output = wan_rope.forward_cuda(
+        x,
+        torch.randn(1, 5, 1, 4),
+        torch.randn(1, 5, 1, 4),
+    )
+    assert output.shape == x.shape
+
+
 def test_mindie_full_dim_rope_keeps_h3_cos_sin_layout(monkeypatch: pytest.MonkeyPatch) -> None:
     """H3 passes complete rotary dims to MindIE instead of the CUDA half-dim layout."""
     captured: dict[str, object] = {}