Delete files lance_mlx/__pycache__/__init__.cpython-314.pyc lance_mlx/__pycache__/lance.cpython-314.pyc lance_mlx/__pycache__/modeling_utils.cpython-314.pyc lance_mlx/__pycache__/qwen2_navit_mlx.cpython-314.pyc lance_mlx/__pycache__/vae_wan22.cpython-314.pyc with huggingface_hub

Bundle lance_mlx Python package

Update README: remove broken github link, point to bundled code

Add requirements.txt

Add self-contained inference.py + bundled lance_mlx package

docs: full Phase 2 sync — TI2I + X→T statuses, sequence layouts, perf tables, file descriptions

TI2I edit: Lance chat template + 3-component CFG verified. Semantic edits working (bow tie, panther).

Add TI2I edit attempt sample

TI2I edit pipeline: architecture working end-to-end with ViT + VAE conditioning + CFG. Output calibration pending.

Phase 2: X-T (understanding) verified — accurate image captioning at ~30 tok/s with KV cache. T2V cross-linked.

Initial release: Lance-3B-MLX first MLX port, T2I verified end-to-end

README: honest status — weight conversion done, inference wrapper partial; document MoE-gen + QK-norm gaps

Add ViT config

Add ViT (Qwen2.5-VL visual encoder) — 390 tensors / 668M params

Initial MLX port of Lance 3B (image variant) + Wan 2.2 VAE

initial commit