Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Studio video: keep the stream encode hook out of a compiled VAE decode graph (#14822)
Daniel Han8cbcf05f66e5
diff --git a/studio/backend/core/inference/video_stream_encode.py b/studio/backend/core/inference/video_stream_encode.pyindex 82ac41ae875..22780964ca3 100644--- a/studio/backend/core/inference/video_stream_encode.py+++ b/studio/backend/core/inference/video_stream_encode.py@@ -183,7 +183,11 @@ class StreamEncode:try:_wrap("decode", _decode_wrapper)_wrap("tiled_decode", _tiled_wrapper)- handle = vae.decoder.register_forward_hook(self._on_chunk, with_kwargs = True)+ # Disabled for Dynamo: under a compiled decode (UNSLOTH_DIFFUSION_COMPILE_VAE=1) the hook's thread, queue and+ # CUDA stream code broke the graph at each of them; as one eager island the chunks still stream.+ handle = vae.decoder.register_forward_hook(+ self._torch.compiler.disable(self._on_chunk), with_kwargs = True+ )except Exception: # noqa: BLE001 - a VAE that refuses the hooks just encodes afterwardsself.invalidate("hooks refused")try:diff --git a/studio/backend/tests/test_video_stream_encode.py b/studio/backend/tests/test_video_stream_encode.pyindex 23f7c67be90..b12d0446df1 100644--- a/studio/backend/tests/test_video_stream_encode.py+++ b/studio/backend/tests/test_video_stream_encode.py@@ -100,6 +100,40 @@ def test_streamed_frames_are_the_exported_frames(patch, tmp_path):assert "decode" not in vars(pipe.vae) and not pipe.vae.decoder._forward_hooks[email protected]("device", ["cpu", "cuda"])+def test_a_compiled_decode_runs_the_chunk_hook_as_one_eager_island(device):+ """UNSLOTH_DIFFUSION_COMPILE_VAE=1 compiles ``vae.decode`` (fullgraph=False, dynamic=True): Dynamo must not trace+ into the hook's thread, queue and CUDA stream code, and the streamed frames are still the exported ones."""+ if device == "cuda" and not torch.cuda.is_available():+ pytest.skip("needs CUDA")+ from torch._dynamo.utils import counters++ torch._dynamo.reset()+ counters.clear()+ pipe = _Pipe(_vae().to(device))+ z = torch.randn(1, 4, 4, 8, 8, device = device)+ with torch.inference_mode():+ reference = pipe.video_processor.postprocess_video(+ pipe.vae.decode(z, return_dict = False)[0], output_type = "np"+ )[0]+ pipe.vae.decode = torch.compile(+ pipe.vae.decode, backend = "eager", fullgraph = False, dynamic = True+ )+ try:+ for _ in range(2):+ session = vse.StreamEncode(pipe.vae, pipe.video_processor, 8)+ frames = _render(pipe, session, z)+ assert len(session.chunks) == 4+ assert np.array_equal(frames.cpu().numpy(), (reference * 255).round().astype(np.uint8))+ assert session.take(frames, 8) is not None+ reasons = {k.splitlines()[0] for k in counters["graph_break"]}+ assert all("torch.compiler.disable" in r for r in reasons), reasons+ assert sum(counters["graph_break"].values()) <= 2+ finally:+ del pipe.vae.decode+ torch._dynamo.reset()++def test_tiled_decode_falls_back():vae = _vae()vae.enable_tiling(tile_sample_min_height = 32, tile_sample_min_width = 32)