Add LTX2Model plugin to fastvideo/train (FineTuneMethod-compatible),
matching the official Lightricks LTX-2 trainer semantics:
- stretched shifted-logit-normal sigma sampling with token-count-
dependent shift (0.95@1024 -> 2.05@4096 tokens) and 10% uniform mix
- per-token sigma timesteps; DiT x0 output converted to velocity so
the default noise-minus-clean MSE target applies
- fps-carrying forward context so training RoPE matches validation
- activation checkpointing applied to the nested transformer.model
- audio/cross-modal params (5.8B) frozen by default for video-only FT
Also add a standalone overfit preprocess script (LTX-2 VAE latents +
post-connector Gemma embeddings -> t2v parquet), a 4-GPU overfit
config, single-step GPU smoke tests with a seeded GB200 grad-norm
reference, and a nightly e2e overfit test on the single-cat sample.
Verified on 4x GB200 (FSDP full shard, sp_size=1): 300 steps, train
loss 0.51 -> 0.012, validation mean MS-SSIM vs the training clip
0.15 (step 0) -> 0.96 (step 300); e2e test passes.
Will Lin
f96b8e2ad2