[feat] cosmos3 PR2: t2vs pipeline (joint denoise + AVAE decode + AV mux)

Wire text-to-video+sound end-to-end, completing the audio PR. - CFG velocity (`cosmos3_get_cfg_velocity`) + `Cosmos3DenoiseEngine` take an optional `sound_specs`: the denoise latent becomes a combined [vision | sound] flat vector; each step packs vision+sound, forwards the DiT, zeros the prediction on conditioning frames for both, and returns the concatenated velocity. The UniPC scheduler steps the joint vector. `Cosmos3SoundSpec` + `_split_sound_latent` added. - `Cosmos3DenoisingStage` t2vs path (gated on `COSMOS3_T2VS`): size the sound latent from the video duration (framework `create_placeholder_audio` + `get_latent_num_samples`), append sound noise, joint-denoise, split, AVAE-decode the sound latent, and set `batch.extra["audio"]` / `["audio_sample_rate"]` so the generator muxes a stereo 48 kHz AAC track into the mp4. Sound AVAE is lazy-loaded from `<model_path>/sound_tokenizer` (`_get_sound_vae`). - Parity (`test_cosmos3_sound_parity::test_t2vs_cfg_velocity_matches_framework`): the combined [vision|sound] sequential-CFG velocity vs a framework-DiT oracle — max abs diff = 0.0, mean abs diff = 0.0 across t2vs/i2vs cases. - Example `basic_cosmos3_t2vs_new_api.py`. Verified on B200 (real weights, 1280x704, 29f, 35 steps): coherent ocean/sunset video WITH a real stereo 48 kHz audio track (mean -10.2 dB, peak -0.4 dB; not silence). Full cosmos3 suite: 130 passed, 0 skipped.
2026-07-20 12:19:46 -07:00
parent d6affff3a1
commit eb9d439bab
4 changed files with 261 additions and 17 deletions
@@ -0,0 +1,67 @@
+import os
+import time
+
+# t2vs (text -> video + sound). The Cosmos3 denoise stage generates a joint
+# [vision | sound] latent and AVAE-decodes the sound to a waveform muxed into the
+# mp4. The joint-sound path is gated on COSMOS3_T2VS (set here for the example).
+os.environ.setdefault("COSMOS3_T2VS", "1")
+
+from fastvideo import VideoGenerator  # noqa: E402
+from fastvideo.api import (  # noqa: E402
+    EngineConfig,
+    GenerationRequest,
+    GeneratorConfig,
+    OffloadConfig,
+    OutputConfig,
+    SamplingConfig,
+)
+
+OUTPUT_PATH = "video_samples_cosmos3_t2vs"
+
+
+def main():
+    model_name = os.environ.get("COSMOS3_MODEL_PATH", "nvidia/Cosmos3-Nano")
+
+    generator_config = GeneratorConfig(
+        model_path=model_name,
+        engine=EngineConfig(
+            num_gpus=1,
+            use_fsdp_inference=False,
+            offload=OffloadConfig(text_encoder=True, pin_cpu_memory=True, dit=False, vae=False),
+        ),
+    )
+
+    load_start = time.perf_counter()
+    generator = VideoGenerator.from_config(generator_config)
+    load_time = time.perf_counter() - load_start
+
+    prompt = (
+        "Ocean waves crash against a rocky shore at sunset, white foam spraying "
+        "into the air as seagulls wheel overhead. Golden light, cinematic wide "
+        "shot, the rhythmic roar of the surf."
+    )
+    request = GenerationRequest(
+        prompt=prompt,
+        sampling=SamplingConfig(
+            num_frames=int(os.environ.get("COSMOS3_NUM_FRAMES", "189")),
+            height=int(os.environ.get("COSMOS3_HEIGHT", "704")),
+            width=int(os.environ.get("COSMOS3_WIDTH", "1280")),
+            num_inference_steps=int(os.environ.get("COSMOS3_STEPS", "35")),
+            guidance_scale=6.0,
+            fps=24,
+            seed=1024,
+        ),
+        output=OutputConfig(output_path=OUTPUT_PATH, save_video=True, return_frames=False),
+    )
+
+    start = time.perf_counter()
+    result = generator.generate(request)
+    gen_time = time.perf_counter() - start
+
+    print(f"Time taken to load model: {load_time} seconds")
+    print(f"Time taken to generate video+sound: {gen_time} seconds")
+    print(f"Output written to: {result.video_path}")
+
+
+if __name__ == "__main__":
+    main()