mstar.model.cosmos3.tests.test_video_capture

CPU checks for the video / windowed denoise CUDA-graph capture: one graph per latent shape built with every frame declared noisy, the request’s clean/noisy layout carried as a per-token mask input. Covers the bucket declaration, the capture key (video, chained windows; never kv windows), the captured inputs a windowed request stages, the captured-step tail in postprocess (masked velocity, pinned frames, window boundary), and the mask’s exact equivalence to the eager layout on a tiny transformer.

Functions

test_masked_capture_matches_eager_layout()	On a tiny transformer: the all-frames-noisy graph layout with the clean/noisy mask reproduces the eager i2v layout's velocity on the noisy frames exactly (the clean frame's output is what postprocess zeroes).
test_postprocess_captured_video_step(monkeypatch)	The captured tail zeroes the clean frames' velocity, re-pins masked frames, and closes a chained window at its last step.
test_video_capture_buckets_and_keys(monkeypatch)
test_windowed_captured_inputs_carry_the_window_layout(...)	A chained window past the first stages the graph's inputs: timesteps from the within-window step over every token, and the token mask zero on the re-pinned overlap frames.

mstar.model.cosmos3.tests.test_video_capture.test_masked_capture_matches_eager_layout()

On a tiny transformer: the all-frames-noisy graph layout with the clean/noisy mask reproduces the eager i2v layout’s velocity on the noisy frames exactly (the clean frame’s output is what postprocess zeroes).

Return type:

None

mstar.model.cosmos3.tests.test_video_capture.test_postprocess_captured_video_step(monkeypatch)

The captured tail zeroes the clean frames’ velocity, re-pins masked frames, and closes a chained window at its last step.

Return type:

None

mstar.model.cosmos3.tests.test_video_capture.test_video_capture_buckets_and_keys(monkeypatch)

Return type:

None

mstar.model.cosmos3.tests.test_video_capture.test_windowed_captured_inputs_carry_the_window_layout(monkeypatch)

A chained window past the first stages the graph’s inputs: timesteps from the within-window step over every token, and the token mask zero on the re-pinned overlap frames.

Return type:

None