SolitaryThinkerandClaude Opus 5.5
d692d1cc3b
These ports were marked pending removal in #1820. Each has a newer, maintained counterpart in the repo, and their model-specific hooks leaked into shared stages, request fields, CLI args, and loaders. Remove each model's code, configs, presets, examples, converters, tests, cookbook pages, CI planner rules, and registry entries, along with the shared-code hooks only they used: - camera kwargs (GameCraft camera_states, DreamX y_camera) in the shared denoising loop - LongCat refine branches in latent preparation, the decoding crop, the --bsa-* CLI args, its train model, and third_party/longcat_video - Z-Image cfg_normalization/cfg_truncation and GameCraft/DreamX camera/action fields in SamplingParam, ForwardBatch, and the schema - the GLM-Image vision-language loader and the TurboDiffusion rCM scheduler and SLA auto-select in the train entrypoint - the Wan definition groups that only preserved DreamX detector order Removed env vars and their legacy aliases move to DEPRECATED_VARIABLES per the env-var policy. Matrix Game 2.0 stays: the new trainer's world model scenarios depend on it. The SLA and BSA attention backends stay. The TurboDiffusion int8 kernels in fastvideo-kernel are left for a separate kernel PR. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
2.7 KiB
2.7 KiB
scripts/checkpoint_conversion/ — Official → FastVideo Converters
Generated: 2026-05-02
Pre-commit excludes
scripts/. Format / lint by hand against neighboring converters before opening a PR.
What Lives Here
checkpoint_conversion/
├── convert_gen3c_to_fastvideo.py
├── convert_ltx2_weights.py
├── convert_minimax_h3_adaln_rank.py # Rank-reduces AdaLN in place; same family
├── stable_audio_to_diffusers.py
├── wan_to_diffusers.py
├── pt_to_safetensors.py # Generic format flip
└── create_hf_repo.py # Push to HF after conversion
Naming Convention
| Pattern | Use |
|---|---|
convert_<model>_*.py / <model>_to_<format>.py |
One-shot converter for a model family |
extract_<role>_*.py |
Pull a sub-component out of a multimodal repo |
validate_<model>_*.py |
Post-conversion shape / norm sanity checks |
<format>_to_<format>.py |
Generic format conversion (no model knowledge) |
Authoring a New Converter
- Prototype the FastVideo-native component first under
fastvideo/models/<role>/<model>.py. Itsstate_dict()is the target surface — never the other way around. - Mirror the official-checkpoint key pattern → FastVideo-native key pattern in
a
param_names_mapping(declared on the arch config infastvideo/configs/models/<role>/<model>.py). - The converter's only job: load the official checkpoint, apply the mapping
(with split/fuse for QKV / packed MLP), and write a safetensors directory
that the
ComponentLoader(fastvideo/models/loader/) can read. - Record intentionally skipped keys (training-only EMA, optimizer state, logvar, dynamic buffers) in a constant near the top of the converter — with a one-line reason each.
- Add a smoke test under
tests/local_tests/<model>/orfastvideo/tests/<role>/that loads the converted weights and runs a parity assertion against the official reference (1 forward pass).
Cross-Reference
fastvideo/layers/AGENTS.md— which native layer to target (fused QKV, parallel MLP, etc.). Determines the split/fuse logic in the converter.fastvideo/models/AGENTS.md— model-side discipline (lint excluded, native state-dict is the source of truth).
Anti-Patterns
- Hard-coding tensor renames inside the model class instead of the converter.
- Converting via a private fork of
transformers/diffusersweight loaders. Read tensors directly withsafetensorsortorch.load. - Pushing to HF before the validate / smoke test passes locally.
- Skipping the explicit "skipped keys" list — it documents intent and prevents silent precision loss across re-conversions.