henry-y1/wf-ckpt

WaveForcing 14B — S3 step-1500 inference weights (bf16)

model.pt is the generator_ema state dict of the final S3 checkpoint, cast to bf16. It is the only part of the training checkpoint inference reads.

Source	waveforcing_assets/runs/wf14b-hsdp-fast-s3-full-002/s3/checkpoint_model_001500/model.pt
Recipe	14b-hsdp-fast, stage S3, 64 GPUs (HSDP 8×8, SP8, accum 1), max_steps: 1500, status completed
Contents	{"generator_ema": {...}}, 1095 tensors, keys prefixed model. (e.g. model.patch_embedding.weight) — exactly what pipeline.generator.load_state_dict expects
Dropped	generator (raw), critic, the 64 trainer_state_rank*.pt files — 171 GiB model.pt + 286 GiB rank state. Cannot resume training from this file.
Precision	fp32 → bf16. Lossless for inference: infer.py does pipeline.to(dtype=torch.bfloat16) right after load_state_dict.
Denoising schedule	1000,750,500,250, timestep_shift: 5.0, warp_denoising_step: true

Provenance and SHA-256 are in model.source.json.

What else inference needs (not in this bundle)

A native Wan2.1-T2V-14B model root — config.json, the six DiT safetensors shards, models_t5_umt5-xxl-enc-bf16.pth, google/ tokenizer, Wan2.1_VAE.pth (~65 GB, public download). Only the text encoder and VAE are actually used at inference; the base DiT shards are replaced by model.pt.

Running it (on an 8-GPU host, per training/README + tools/standard_infer)

RollingForcing pipeline, one prompt per GPU:

tools/standard_infer/infer.py --mode rf \
  --checkpoint /path/to/model.pt \
  --prompt-dir /path/to/prompts --num-prompts 10 \
  --preset standard --steps 1000,750,500,250 --seed 42 \
  --model-name Wan2.1-T2V-14B --asset-root /path/to/rf_assets \
  --gpus 0,1,2,3,4,5,6,7 --output /path/to/out

WaveRT wave-parallel runtime (4-step → wp_size 5, rf_step 4, vae_stages 3):

tools/standard_infer/infer.py --mode wf \
  --checkpoint /path/to/model.pt ... # same flags

--mode wf launches python -m wave_rt serve --model-path <diffusers view> --gen-ckpt <model.pt>. Protocol is fixed at seed 42, 480×832, 16 fps.

Note: this is an S3 endpoint checkpoint, not a quality-validated release.