trackwan-14b-720p-overfit-stageA

TrackWan 14B / 720p — Overfit Stage A checkpoint (step 2000)

Training-state checkpoint (PyTorch DCP) from Phase 0 / Stage A of the bidirectional TrackWan teacher recipe (point-track-conditioned I2V, 14B, 720p), built in FastVideo.

What this is

• Stage A overfit of the track pathway (track_encoder + patch-embed track slot), head trainable, on a small 50-clip 720p synthetic set. Fixed track-ID sampling (WANTRACK_FIXED_SAMPLE=1) so each clip presents an identical sparse conditioning pattern.
• Config: lr 1e-4 (constant), 2000 steps, global batch 8, flow_shift 6, sparse conditioning (WANTRACK_SPARSE=1, EXTRA_RANDOM=20), d64 track-ID embedding + bias track encoder.
• Base init: trackwan_14b_i2v_d64_bias_init (Wan2.1-I2V-14B-720P grafted to 52-ch TrackWan DiT).

Format & caveats

• This is a full training-state checkpoint: model + optimizer + scheduler + per-rank RNG, sharded as dcp/__<rank>_0.distcp (8 ranks, HSDP 2×4). It is not a from_pretrained model — to load weights, consolidate via the FastVideo export step (03_export.sh).
• Resume by placing checkpoint-2000/ under a training output_dir and setting resume_from_checkpoint: latest.
• The overfit base is discarded by the Phase-1 merge (only the track pathway is lifted into a pristine 14B base), so this checkpoint is primarily a reproducibility / seed artifact.

Provenance

Stage A → Stage B (random-ID) → export → merge → OpenVid stage-1 → synthetic stage-2. See the FastVideo data_pipeline/720_stage_1/ recipe.