feat: add Flux, LTX-Video, LTX-2/2.3, speculative decoding, and Windows CUDA support#45

Closed

icryo wants to merge 24 commits intoevilsocket:mainfrom

icryo:windows-cuda-support

Commits on Mar 7, 2026

feat: add LTX-2, Flux, LTX-Video model support + video generation pipeline

icryo
and
claude
committed

Commits on Mar 9, 2026

fix(ltx2): correct multiple LTX-2.3 transformer bugs

icryo
and
claude
committed
feat(ltx2): implement STG (Spatio-Temporal Guidance) for LTX-2.3

icryo
and
claude
committed
fix: use V2 per-token RMS normalization for text embeddings

icryo
and
claude
committed
fix(ltx2): guard cross-attention AdaLN on actual tensor dim

icryo
and
claude
committed
fix(ltx2): correct STG tensor unpacking order on worker

icryo
and
claude
committed
fix: skip generic HF download for image models (no root config.json)

icryo
and
claude
committed
fix: use default HF cache when model path is a repo ID

icryo
and
claude
committed
fix: only use model-local HF cache when model path is a real directory

icryo
and
claude
committed
fix: download sharded transformer weights from HF before loading

icryo
and
claude
committed

Commits on Mar 15, 2026

Commits on Apr 4, 2026

feat: Wan 14B video generation with GGUF quantized + F16 distributed support
icryo
committed

Commits on Apr 5, 2026