Manage video model components for production

Download and validate LTX-2.3 video model components sequentially

Idempotent setup of large LTX video generation model components that respects HF rate limits and resumes from failures.

• · LTX_COMPONENTS spec (5 components: transformer, lora, text_encoder, vae, audio)
• · HuggingFace token (from ~/.huggingface/token or env)
• · project root directory

• · downloaded model files in component subdirs
• · manifest.json with updated size_bytes per component
• · assembly symlink verification passed
• · ltx_pipelines_mlx import test passed

• · HuggingFace hub (huggingface-hub CLI / Python)
• · Python 3 + mlx_arsenal + mlx_lm + huggingface_hub
• · bash shell for symlink setup

• · HuggingFace token available (user has access to dgrauet/ltx-2.3-mlx-q8 repo)
• · Sufficient disk space (~34GB for full set, ~33.5GB without audio)
• · Python env with deps installed (mlx_arsenal, mlx_lm, huggingface_hub)
• · Project structure initialized (component subdirs exist or created)

• · HF rate limit hit during sequential downloads → retry with delay (idempotent by design)
• · Network timeout mid-file → resume from last byte (HF supports range requests)
• · Disk full → operation halts, already-downloaded files retained
• · Manifest sync fails → assembly check fails, recommends re-run
• · ltx_pipelines_mlx import fails → detailed error message, debug steps listed

• · Sequential downloads (not parallel) — respects HF rate limits by design
• · Idempotent: re-run after partial failure resumes from where it stopped
• · Assembly symlink verified before import test (catches setup errors early)
• · Manifest size_bytes tracked for quota awareness
• · 5 phases with clear separation (resolve, download, conversion, validate, report)