[Feat] Minimax H3 support ref and tgt video sparse attention on NPU (#6518)

Signed-off-by: fan2956 <zhoufan53@huawei.com>
Co-authored-by: WeiQing Chen <40507679+david6666666@users.noreply.github.com>

tests/diffusion/attention/test_rainfusion_plan.py

@@ -11,13 +11,16 @@ to the always-kept prefix segment before generating the block mask.
 """
 
 import dataclasses
+import sys
+import types
 
 import pytest
 import torch
 
-from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata, VideoTokenLayout
+from vllm_omni.diffusion.attention.backends.abstract import AttentionMetadata, VideoTokenLayout, VideoTokenSpan
 from vllm_omni.diffusion.attention.backends.rainfusion_attn import (
     _BLOCK_SIZE,
+    RainFusionAttentionBackend,
     RainFusionAttentionImpl,
     RainFusionPlan,
 )
@@ -97,6 +100,62 @@ def test_video_segment_must_be_the_tail_of_packed_document_zero():
     assert make_impl()._resolve_plan(metadata) is None
 
 
+def test_ref2va_multi_video_spans_resolve():
+    layout = VideoTokenLayout(
+        used_len=12000,
+        video_spans=(
+            VideoTokenSpan(start=128, latent_grid=(4, 16, 64), role="reference"),
+            VideoTokenSpan(start=5000, latent_grid=(4, 16, 64), role="target"),
+        ),
+    )
+    plan = make_impl()._resolve_plan(AttentionMetadata(extra={"max_seqlen_q": 12000}, video_layout=layout))
+
+    assert plan is not None
+    assert plan.used_len == 12000
+    assert plan.video_spans == [
+        {"start": 128, "latent_shape": [4, 16, 64]},
+        {"start": 5000, "latent_shape": [4, 16, 64]},
+    ]
+
+
+def test_invalid_ref2va_video_spans_fall_back_to_dense():
+    layout = VideoTokenLayout(
+        used_len=12000,
+        video_spans=(
+            VideoTokenSpan(start=128, latent_grid=(4, 16, 64), role="reference"),
+            VideoTokenSpan(start=4000, latent_grid=(4, 16, 64), role="target"),
+        ),
+    )
+    assert make_impl()._resolve_plan(AttentionMetadata(extra={"max_seqlen_q": 12000}, video_layout=layout)) is None
+
+
+def test_validate_available_rejects_legacy_mindiesd(monkeypatch):
+    mindiesd = types.ModuleType("mindiesd")
+
+    def sparse_attention(query, key, value, **kwargs):
+        return query
+
+    mindiesd.sparse_attention = sparse_attention
+    monkeypatch.setitem(sys.modules, "mindiesd", mindiesd)
+    monkeypatch.setattr("importlib.util.find_spec", lambda _: object())
+
+    with pytest.raises(ValueError, match="video_spans"):
+        RainFusionAttentionBackend.validate_available()
+
+
+def test_validate_available_accepts_new_mindiesd(monkeypatch):
+    mindiesd = types.ModuleType("mindiesd")
+
+    def sparse_attention(query, key, value, *, video_spans=None, **kwargs):
+        return query
+
+    mindiesd.sparse_attention = sparse_attention
+    monkeypatch.setitem(sys.modules, "mindiesd", mindiesd)
+    monkeypatch.setattr("importlib.util.find_spec", lambda _: object())
+
+    RainFusionAttentionBackend.validate_available()
+
+
 @pytest.mark.parametrize("grid", [(4, 24, 40), (1, 24, 40)])
 def test_short_video_stays_dense(grid):
     assert make_impl()._resolve_plan(make_metadata(grid)) is None

tests/diffusion/models/minimax_h3/test_minimax_h3_packing.py

@@ -74,6 +74,8 @@ def test_t2va_and_fl2va_packing_keep_update_rows_separate():
     assert fl2va["update_mask"].sum().item() == 12
     assert (~fl2va["update_mask"]).sum().item() == 6
     assert t2va["audio_pos"].numel() == fl2va["audio_pos"].numel() == 6
+    assert t2va["video_spans"] == ({"start": 10, "latent_grid": (2, 2, 3), "role": "target"},)
+    assert fl2va["video_spans"] == ({"start": 16, "latent_grid": (2, 2, 3), "role": "target"},)
 
 
 def test_ref2va_packing_tracks_video_and_audio_update_masks():
@@ -101,6 +103,33 @@ def test_ref2va_packing_tracks_video_and_audio_update_masks():
     assert packed["cu_seqlens"].tolist() == [0, 30, 64]
 
 
+def test_ref2va_packing_publishes_physical_video_spans_only():
+    from vllm_omni.diffusion.models.minimax_h3.packed_sequence import (
+        minimax_h3_packed_sequence_ref2va_blocks,
+    )
+
+    packed = minimax_h3_packed_sequence_ref2va_blocks(
+        text_len=4,
+        latent_t=2,
+        latent_h=4,
+        latent_w=6,
+        audio_t=3,
+        ref_blocks=[
+            {"kind": "image", "latent_h": 4, "latent_w": 4},
+            {"kind": "video_audio", "ref_audio_t": 2, "latent_t": 2, "latent_h": 4, "latent_w": 6},
+            {"kind": "audio", "ref_audio_t": 1},
+            {"kind": "video", "ref_audio_t": 0, "latent_t": 1, "latent_h": 4, "latent_w": 4},
+        ],
+    )
+
+    # Image and all audio rows intentionally stay out of the sparse spans.
+    assert packed["video_spans"] == (
+        {"start": 12, "latent_grid": (2, 2, 3), "role": "reference"},
+        {"start": 26, "latent_grid": (1, 2, 2), "role": "reference"},
+        {"start": 36, "latent_grid": (2, 2, 3), "role": "target"},
+    )
+
+
 def test_condition_noise_is_seeded_and_keeps_clean_anchor_at_timestep_one():
     from vllm_omni.diffusion.models.minimax_h3.condition_noise import (
         minimax_h3_audio_cond_noise_aug_rows,