Skip to content

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.

Studio video: stream the Wan encode under model / sequential offload too (#15173)

Daniel Han43b83d515dd8
diff --git a/studio/backend/core/inference/video_stream_encode.py b/studio/backend/core/inference/video_stream_encode.py
index 8352e580527..8f643adfae2 100644
--- a/studio/backend/core/inference/video_stream_encode.py
+++ b/studio/backend/core/inference/video_stream_encode.py
@@ -561,6 +561,17 @@ def claim_stream(
return session.take(video_frames, fps, audio, audio_sample_rate)
+def _decodes_on_cuda(pipe: Any, decoder: Any, torch: Any) -> bool:
+ """True if resident on CUDA, or offloaded to host while the pipeline executes on CUDA."""
+ if next(decoder.parameters()).device.type == "cuda":
+ return True
+ try:
+ device = torch.device(getattr(pipe, "_execution_device", None) or "cpu")
+ except Exception: # noqa: BLE001 - no execution device: a CPU render
+ return False
+ return device.type == "cuda"
+
+
def open_stream_encode(
pipe: Any,
fps: int,
@@ -568,7 +579,7 @@ def open_stream_encode(
engage: bool,
logger: Any = None,
) -> Optional[StreamEncode]:
- """A session for a resident CUDA ``AutoencoderKLWan`` with libx264 frame threads, or None."""
+ """A session for an ``AutoencoderKLWan`` decoding on CUDA with libx264 frame threads, or None."""
if not engage or not stream_encode_enabled():
return None
from .video_encode import frame_threads_enabled
@@ -589,7 +600,9 @@ def open_stream_encode(
import av # noqa: F401
import torch
- if not torch.cuda.is_available() or next(decoder.parameters()).device.type != "cuda":
+ if not torch.cuda.is_available():
+ return None
+ if not _decodes_on_cuda(pipe, decoder, torch):
return None
return StreamEncode(vae, processor, fps, logger)
except Exception: # noqa: BLE001 - no PyAV, no CUDA or no parameters: encode afterwards
diff --git a/studio/backend/tests/test_video_stream_encode.py b/studio/backend/tests/test_video_stream_encode.py
index 1c8a67b70ea..5b2373244f1 100644
--- a/studio/backend/tests/test_video_stream_encode.py
+++ b/studio/backend/tests/test_video_stream_encode.py
@@ -199,6 +199,22 @@ def test_engagement_gates(monkeypatch):
assert vse.open_stream_encode(pipe, 8, engage = True) is None
+def test_an_offloaded_vae_engages_when_the_pipeline_executes_on_cuda(monkeypatch):
+ monkeypatch.setattr(torch.cuda, "is_available", lambda: True)
+ pipe = _Pipe(_vae())
+ assert vse.open_stream_encode(pipe, 8, engage = True) is None
+ pipe._execution_device = torch.device("cpu")
+ assert vse.open_stream_encode(pipe, 8, engage = True) is None
+ pipe._execution_device = torch.device("cuda", 0)
+ session = vse.open_stream_encode(pipe, 8, engage = True)
+ try:
+ assert isinstance(session, vse.StreamEncode)
+ finally:
+ session.close()
+ monkeypatch.setenv(vse.STREAM_ENCODE_ENV, "0")
+ assert vse.open_stream_encode(pipe, 8, engage = True) is None
+
+
def test_a_streamed_file_that_does_not_hold_the_clip_is_never_used(monkeypatch):
pipe = _Pipe(_vae())
session = vse.StreamEncode(pipe.vae, pipe.video_processor, 8)