SolitaryThinker 4dcdb0d42c [docs]: PORT_NOTES — RL roles are FSDP2 mp bf16-compute over fp32 storage

2026-07-23 12:53:57 -07:00

..

gates

[misc]: probe v3 — weight dtype/hash + op-input hashes for to_q/norm_q

2026-07-23 12:49:49 -07:00

__init__.py

[fix]: sfwan golden dir name in anchor; [docs] training port authority map

2026-07-23 04:07:49 -07:00

data.py

[feat]: v2.1 parquet dataloader port + order gate; RL capture PickScore-only CLIP unwrap

2026-07-23 11:55:22 -07:00

diffusion_nft.py

[fix]: NFTStep — modular stack IS fp32-master/bf16-compute (FSDP2 mp param_dtype=bf16); restore _MasterOpt chain + master-space old-EMA

2026-07-23 12:52:41 -07:00

dmd2.py

[feat]: self-forcing training gate — sequential draw replay through the causal serving model

2026-07-23 06:13:42 -07:00

finetune.py

[feat]: VSA training gate — capture/anchor vsa mode, per-step ramped metadata through FinetuneStep

2026-07-23 04:50:49 -07:00

PORT_NOTES.md

[docs]: PORT_NOTES — RL roles are FSDP2 mp bf16-compute over fp32 storage

2026-07-23 12:53:57 -07:00

self_forcing.py

[fix]: SF rollout forwards under autocast (plain-LN causal blocks are autocast-sensitive)

2026-07-23 10:51:08 -07:00

videoalign.py

[feat]: vendor VideoAlign reward stack from PR #1476 (byte-identical) + MQ/VQ/TA scorers + gates

2026-07-23 12:15:18 -07:00