some fix

Author: H1yori233

fastvideo/pipelines/basic/wan/wan_svi_i2v_pipeline.py

@@ -33,6 +33,21 @@ def _tensor_to_pil_list(frames: torch.Tensor) -> list[PIL.Image.Image]:
     return [PIL.Image.fromarray(a) for a in arr]
 
 
+def _validate_multiclip_frames(num_motion: int, num_frames: int) -> None:
+    """num_motion must be < num_frames, else follow-up clips stitch to empty."""
+    if num_motion >= num_frames:
+        raise ValueError(f"svi_num_motion_frames ({num_motion}) must be smaller than num_frames ({num_frames}) "
+                         "for multi-clip generation; otherwise stitched follow-up clips would be empty.")
+
+
+def _stitch_clip_outputs(clip_outputs: list[torch.Tensor], num_motion: int) -> torch.Tensor:
+    """Concat per-clip (B,C,T,H,W) videos on time, dropping each follow-up's leading num_motion frames."""
+    concatenated = [clip_outputs[0]]
+    for video in clip_outputs[1:]:
+        concatenated.append(video[:, :, num_motion:, :, :])
+    return torch.cat(concatenated, dim=2)
+
+
 class WanSVIImageToVideoPipeline(WanImageToVideoPipeline):
     """
     Pipeline for Stable-Video-Infinity multi-clip I2V generation on Wan 2.1.
@@ -110,6 +125,9 @@ def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> Forward
                 batch = stage(batch, fastvideo_args)
             return batch
 
+        num_frames = int(batch.num_frames) if batch.num_frames is not None else 0
+        _validate_multiclip_frames(num_motion, num_frames)
+
         # Multi-clip needs the reference image up front to construct motion frames
         # for clip 0. Run InputValidationStage now to resolve image_path -> pil_image.
         self.input_validation_stage(batch, fastvideo_args)
@@ -156,12 +174,7 @@ def forward(self, batch: ForwardBatch, fastvideo_args: FastVideoArgs) -> Forward
                 tail = clip_batch.output[0, :, -num_motion:, :, :]
                 motion_frames = _tensor_to_pil_list(tail)
 
-        # Drop the first num_motion frames of each follow-up clip to avoid duplicating
-        # the previous clip's tail in the stitched output.
-        concatenated = [clip_outputs[0]]
-        for video in clip_outputs[1:]:
-            concatenated.append(video[:, :, num_motion:, :, :])
-        batch.output = torch.cat(concatenated, dim=2)
+        batch.output = _stitch_clip_outputs(clip_outputs, num_motion)
         return batch
 
 

fastvideo/pipelines/lora_pipeline.py

@@ -325,9 +325,12 @@ def set_lora_adapter(self,
             for name, weight in lora_state_dict.items():
                 # Extract weights (lora_A, lora_B, and lora_alpha)
                 name = name.replace("diffusion_model.", "")
-                name = name.replace("pipe.dit.", "")
+                # Guarded so non-SVI / non-PEFT adapters are provably untouched.
+                if "pipe.dit." in name:
+                    name = name.replace("pipe.dit.", "")
                 name = name.replace(".weight", "")
-                name = name.replace(".default", "")
+                if ".default" in name:
+                    name = name.replace(".default", "")
 
                 if "lora_alpha" in name:
                     # Store alpha with minimal mapping - same processing as lora_A/lora_B

fastvideo/tests/inference/lora/test_maybe_download_lora.py

@@ -0,0 +1,73 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Path-routing tests for fastvideo.utils.maybe_download_lora."""
+from __future__ import annotations
+
+import huggingface_hub
+
+import fastvideo.utils as fv_utils
+from fastvideo.utils import maybe_download_lora
+
+
+def test_triple_slash_downloads_single_file(monkeypatch):
+    """org/repo/sub/file.safetensors -> hf_hub_download(repo_id=org/repo, filename=sub/file)."""
+    calls: list[dict] = []
+
+    def fake_hf_hub_download(*, repo_id, filename, local_dir=None, **kwargs):
+        calls.append({"repo_id": repo_id, "filename": filename, "local_dir": local_dir})
+        return f"/cache/{repo_id}/{filename}"
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", fake_hf_hub_download)
+
+    result = maybe_download_lora("vita-video-gen/svi-model/version-1.0/svi-shot.safetensors")
+
+    assert len(calls) == 1
+    assert calls[0]["repo_id"] == "vita-video-gen/svi-model"
+    assert calls[0]["filename"] == "version-1.0/svi-shot.safetensors"
+    assert result == "/cache/vita-video-gen/svi-model/version-1.0/svi-shot.safetensors"
+
+
+def test_triple_slash_keeps_only_first_two_segments_as_repo(monkeypatch):
+    """A deeper nesting still maps repo_id to the first two segments."""
+    captured: dict = {}
+
+    def fake_hf_hub_download(*, repo_id, filename, local_dir=None, **kwargs):
+        captured["repo_id"] = repo_id
+        captured["filename"] = filename
+        return "ok"
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", fake_hf_hub_download)
+
+    maybe_download_lora("org/repo/a/b/c/weights.safetensors")
+
+    assert captured["repo_id"] == "org/repo"
+    assert captured["filename"] == "a/b/c/weights.safetensors"
+
+
+def test_local_file_short_circuits(tmp_path, monkeypatch):
+    """An existing local .safetensors file is returned verbatim, no download."""
+
+    def boom(*args, **kwargs):
+        raise AssertionError("hf_hub_download must not be called for a local file")
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", boom)
+
+    f = tmp_path / "version-1.0" / "svi-shot.safetensors"
+    f.parent.mkdir(parents=True)
+    f.write_bytes(b"\x00")
+
+    assert maybe_download_lora(str(f)) == str(f)
+
+
+def test_plain_repo_id_falls_through_to_repo_download(monkeypatch):
+    """A two-segment HF id (no .safetensors suffix) does NOT hit the single-file branch."""
+
+    def boom(*args, **kwargs):
+        raise AssertionError("two-segment repo id must not take the single-file branch")
+
+    monkeypatch.setattr(huggingface_hub, "hf_hub_download", boom)
+    monkeypatch.setattr(fv_utils, "maybe_download_model", lambda *a, **k: "/cache/org/repo")
+    monkeypatch.setattr(fv_utils, "_best_guess_weight_name", lambda *a, **k: "adapter.safetensors")
+
+    result = maybe_download_lora("org/repo")
+
+    assert result == "/cache/org/repo/adapter.safetensors"

fastvideo/tests/pipelines/test_svi_multiclip_stitch.py

@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Tests for SVI multi-clip stitch index math and the frame-budget guard."""
+from __future__ import annotations
+
+import pytest
+import torch
+
+from fastvideo.pipelines.basic.wan.wan_svi_i2v_pipeline import (_stitch_clip_outputs, _validate_multiclip_frames)
+
+
+def _clip(num_frames: int, fill: float) -> torch.Tensor:
+    # (B=1, C=3, T=num_frames, H=2, W=2)
+    return torch.full((1, 3, num_frames, 2, 2), fill, dtype=torch.float32)
+
+
+def test_stitch_drops_motion_overlap_from_followups():
+    num_motion = 2
+    clips = [_clip(9, 0.0), _clip(9, 1.0), _clip(9, 2.0)]
+
+    out = _stitch_clip_outputs(clips, num_motion)
+
+    expected_t = 9 + (9 - num_motion) + (9 - num_motion)
+    assert out.shape == (1, 3, expected_t, 2, 2)
+
+    # Provenance: 9 from clip0, then 7 from clip1, then 7 from clip2.
+    assert torch.all(out[:, :, :9] == 0.0)
+    assert torch.all(out[:, :, 9:9 + 7] == 1.0)
+    assert torch.all(out[:, :, 9 + 7:] == 2.0)
+
+
+def test_stitch_single_clip_is_identity():
+    clip = _clip(5, 3.0)
+    out = _stitch_clip_outputs([clip], num_motion=2)
+    assert torch.equal(out, clip)
+
+
+def test_stitch_num_motion_one_drops_single_frame():
+    clips = [_clip(4, 0.0), _clip(4, 1.0)]
+    out = _stitch_clip_outputs(clips, num_motion=1)
+    assert out.shape[2] == 4 + 3
+
+
+def test_validate_rejects_motion_ge_frames():
+    with pytest.raises(ValueError, match="must be smaller than num_frames"):
+        _validate_multiclip_frames(num_motion=5, num_frames=5)
+    with pytest.raises(ValueError, match="must be smaller than num_frames"):
+        _validate_multiclip_frames(num_motion=8, num_frames=5)
+
+
+def test_validate_accepts_motion_lt_frames():
+    # Should not raise.
+    _validate_multiclip_frames(num_motion=1, num_frames=81)
+    _validate_multiclip_frames(num_motion=5, num_frames=6)