[misc]: simplify cosmos3 port (reuse embed, dedup prompt, hoist)

Behavior-preserving cleanup (95 tests green):
- reuse fastvideo.layers.visual_embedding.timestep_embedding in the DiT
- single-home COSMOS3_VIDEO_NEGATIVE_PROMPT in the lightweight presets
  module; drop the unused duplicate from cosmos3_pipeline
- drop the write-only Cosmos3VisionSpec.clean_latent field + its kwarg
- hoist device = next(transformer.parameters()).device out of the per-pass
  _run closure in cosmos3_get_cfg_velocity

fastvideo/models/dits/cosmos3.py

@@ -46,6 +46,7 @@ import torch.nn.functional as F
 from fastvideo.configs.models.dits.cosmos3 import Cosmos3VideoConfig
 from fastvideo.layers.layernorm import RMSNorm
 from fastvideo.layers.linear import ReplicatedLinear
+from fastvideo.layers.visual_embedding import timestep_embedding
 from fastvideo.models.dits.base import BaseDiT
 
 EntryClass = ["Cosmos3VFMTransformer"]
@@ -320,19 +321,8 @@ class Cosmos3TimestepEmbedder(nn.Module):
         self.act = nn.SiLU()
         self.linear_2 = ReplicatedLinear(hidden_size, hidden_size, bias=True)
 
-    @staticmethod
-    def timestep_embedding(t: torch.Tensor, dim: int, max_period: int = 10000) -> torch.Tensor:
-        half = dim // 2
-        freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
-            device=t.device)
-        args = t[:, None].float() * freqs[None]
-        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
-        if dim % 2:
-            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
-        return embedding
-
     def forward(self, t: torch.Tensor) -> torch.Tensor:
-        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        t_freq = timestep_embedding(t, self.frequency_embedding_size)
         # Sinusoidal runs in fp32 (timestep_scale makes inputs tiny); cast to the
         # MLP weight dtype (bf16 at inference; no-op in the fp32 parity tests).
         t_freq = t_freq.to(self.linear_1.weight.dtype)

fastvideo/pipelines/basic/cosmos3/cosmos3_pipeline.py

@@ -60,15 +60,6 @@ logger = init_logger(__name__)
 _SYSTEM_PROMPT_IMAGE = "You are a helpful assistant who will generate images from a give prompt."
 _SYSTEM_PROMPT_VIDEO = "You are a helpful assistant who will generate videos from a give prompt."
 
-# Default video negative prompt (framework / Cosmos quality prompt).
-COSMOS3_VIDEO_NEGATIVE_PROMPT = (
-    "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, "
-    "over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, "
-    "underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, "
-    "jerky movements, low frame rate, artifacting, color banding, unnatural transitions, outdated special effects, "
-    "fake elements, unconvincing visuals, poorly edited content, jump cuts, visual noise, and flickering. "
-    "Overall, the video is of poor quality.")
-
 
 # ===========================================================================
 # Special-token resolution (Qwen2 chat tokenizer)
@@ -175,15 +166,12 @@ class Cosmos3VisionSpec:
     """Geometry + conditioning for one vision item in a denoise run.
 
     Args:
-        clean_latent: VAE-encoded conditioning latent ``[C, T, H, W]`` used to
-            keep condition frames clean (I2V / T2I). May be ``None`` for T2V.
         condition_frame_indexes: Latent-frame indices kept clean.
         shape: ``(C, T, H, W)`` of the latent for this item.
     """
 
     shape: tuple[int, int, int, int]
     condition_frame_indexes: list[int]
-    clean_latent: torch.Tensor | None = None
 
     @property
     def numel(self) -> int:
@@ -244,6 +232,7 @@ def cosmos3_get_cfg_velocity(
     timestep_value = float(timestep.reshape(()).item())
 
     noise_x_vision = _split_flat_latent(flat_latent, specs)
+    device = next(transformer.parameters()).device
 
     def _run(token_ids: list[int]) -> torch.Tensor:
         samples = [
@@ -268,7 +257,6 @@ def cosmos3_get_cfg_velocity(
             base_fps=base_fps,
             temporal_compression_factor=temporal_compression_factor,
         )
-        device = next(transformer.parameters()).device
         out = transformer(**packed.to_dit_kwargs(device=device))
         preds = out.get("preds_vision")
         if preds is None:

fastvideo/pipelines/basic/cosmos3/presets.py

@@ -20,7 +20,7 @@ _DENOISE_STAGE = PresetStageSpec(
 )
 
 # Framework video negative prompt (Cosmos quality prompt).
-_COSMOS3_NEGATIVE_PROMPT = (
+COSMOS3_VIDEO_NEGATIVE_PROMPT = (
     "The video captures a series of frames showing ugly scenes, static with no motion, motion blur, "
     "over-saturation, shaky footage, low resolution, grainy texture, pixelated images, poorly lit areas, "
     "underexposed and overexposed scenes, poor color balance, washed out colors, choppy sequences, "
@@ -42,7 +42,7 @@ COSMOS3_NANO = InferencePreset(
         "fps": 24,
         "guidance_scale": 6.0,
         "num_inference_steps": 35,
-        "negative_prompt": _COSMOS3_NEGATIVE_PROMPT,
+        "negative_prompt": COSMOS3_VIDEO_NEGATIVE_PROMPT,
     },
 )
 
@@ -60,7 +60,7 @@ COSMOS3_NANO_I2V = InferencePreset(
         "fps": 24,
         "guidance_scale": 6.0,
         "num_inference_steps": 35,
-        "negative_prompt": _COSMOS3_NEGATIVE_PROMPT,
+        "negative_prompt": COSMOS3_VIDEO_NEGATIVE_PROMPT,
     },
 )
 

fastvideo/pipelines/stages/cosmos3_stages.py

@@ -30,7 +30,6 @@ from tqdm import tqdm
 from fastvideo.fastvideo_args import FastVideoArgs
 from fastvideo.logger import init_logger
 from fastvideo.pipelines.basic.cosmos3.cosmos3_pipeline import (
-    COSMOS3_VIDEO_NEGATIVE_PROMPT,
     Cosmos3DenoiseEngine,
     Cosmos3VisionSpec,
     _VaeNorm,
@@ -39,6 +38,8 @@ from fastvideo.pipelines.basic.cosmos3.cosmos3_pipeline import (
     cosmos3_vae_decode,
     cosmos3_vae_encode,
 )
+from fastvideo.pipelines.basic.cosmos3.presets import (
+    COSMOS3_VIDEO_NEGATIVE_PROMPT, )
 from fastvideo.pipelines.pipeline_batch_info import ForwardBatch
 from fastvideo.pipelines.stages.base import PipelineStage
 
@@ -141,7 +142,6 @@ class Cosmos3DenoisingStage(PipelineStage):
         spec = Cosmos3VisionSpec(
             shape=latent_shape,
             condition_frame_indexes=condition_frame_indexes,
-            clean_latent=clean_latent,
         )
 
         # ---- Scheduler timesteps ----