Extend the LTX-2 training plugin to LTX-2.3 checkpoints (gated
attention, cross-attention AdaLN, 4096-d post-connector text
embeddings, no in-DiT caption projection):
- preprocess captions via the text encoder's forward() (the inference
path) instead of preprocess_text_embeddings + run_connectors, which
crashes on 2.3 (4096-d video features fed into the 2048-d audio
connector); output verified byte-identical for 2.0
- fail fast on parquet/checkpoint text-width mismatches; the check
reads the loaded transformer's config since the transformer loader
deepcopies dit_config before applying checkpoint arch flags
- clear the map-style dataset cache when regenerating overfit parquet
- add overfit_ltx2_3_t2v.yaml and parametrize the smoke, load, and
nightly e2e tests over ltx2/ltx2_3; seed the GB200 grad-norm
reference for 2.3
Verified on 4x GB200: LTX-2.3 overfit 300 steps, train loss -> 0.0085,
validation mean MS-SSIM vs the training clip 0.127 (step 0) -> 0.957
(step 300); parametrized nightly e2e passes for both versions
(0.964 / 0.957).
Will Lin
86a360ebc7