Minimax h3 lora training

[imprsnst]

Summary

MiniMax H3 LoRA training

Adds MiniMax H3 as a fourth trainable architecture. Trains on still images, and the adapter applies to video.

What's new

• H3 arch entry in arch.py, sharing Z-Image's flow convention
• Staged precache in h3.py, because the fp32 video VAE and the 32B conditioner cannot be resident together
• 4-bit base loader: factorise the AdaLN branch, then quantise, streaming block by block
• lora port on all four H3 nodes in runner.py, so the trained file is usable
• Trainer UI picker, plus the model download popup wired to the H3 requirements

Measured on an L40S

Phase	Peak
Latent caching	10.8GB
Caption caching	20.5GB
Training	11.7GB

• Run peaks at 20.6GB, set by the caption pass, so 512px and 768px cost the same
• 0.63s a step at 512px, 0.77s at 768px, about 7 minutes of startup
• 62GB checkpoint down to 11.7GB, host RAM stays near 1.1GB
• Full tables in TRAINING.md

Fixes two live bugs in the existing trainer

• _to_device cast every non-mask tensor to the compute dtype, which would destroy index tensors
• PEFT never saw is_loaded_in_4bit, so NF4 runs silently used the generic LoRA layer instead of the bitsandbytes one. Affects Krea 2 and FLUX.2 today

Verified

• 771 Python tests, 205 TS tests, typecheck and lint clean
• Trained a 400 step LoRA and rendered the same seed with and without it. Clips differ by 50.4 mean absolute out of 255, both coherent
• Fuse checked arithmetically: all 313 adapted modules, including 13 outside the block stack that a per-block fuse would have silently missed

Known, not fixed here

• A LoRA trained with alpha different from rank fuses at the wrong strength, on all four architectures. Hidden by the default of alpha equalling rank
• 15 pre-existing UP042 lint errors in unrelated files, deferred deliberately

Also bumps to 1.2.62.
Closes #

Checklist

• npm run typecheck && npm run lint && npm run test pass
• cd core && ruff check . && uv run pytest -q pass (if the engine changed)
• Commits use Conventional Commits and are signed off (git commit -s)
• Change is small and focused, and matches the surrounding style