---

name: model-zoo description: "Guide selection, import, and compatibility of Fengshen model/config/tokenizer families and top-level exports." metadata: disco-role: operating disable-model-invocation: true license: Apache 2.0

Fengshenbang-LM Model Zoo

Use this sub-skill when a task is about choosing, importing, or troubleshooting Fengshen model, config, tokenizer, and generation-helper families. Keep this file as a router; use the bundled references for tables and details.

Load order

1. Read references/model-overview.md to choose the model family and decide whether the package surface, a pipeline surface, or an example/conversion recipe is the right owner.
2. Read references/api-reference.md for top-level exports, direct import paths, generation-helper routes, and the fengshen_model_type pipeline mapping.
3. Read references/auto-and-tokenizers.md before using Fengshen auto factories, custom config keys, AutoTokenizer, or tokenizer-specific files.
4. Read references/compatibility.md before installing or repairing dependencies, especially Transformers compatibility.
5. Read references/troubleshooting.md when imports, tokenizers, offline caches, CUDA size limits, or optional DeepSpeed/Megatron paths fail.

Choose this sub-skill for

• Top-level imports from fengshen, especially LongformerConfig, LongformerModel, RoFormerConfig, RoFormerModel, T5Config, T5EncoderModel, UbertPipelines, and UbertModel.
• Custom model families: Longformer, RoFormer, Megatron-T5, ZEN1/ZEN2, DeBERTa-v2, DeltaLM, LLaMA/Ziya, Taiyi CLIP, Transformer-XL denoise/paraphrase/reasoning, BART, ALBERT, UniMC/UniEX/TCBert/Ubert, and VAE variants.
• Fengshen auto factories (fengshen.models.auto), custom model/tokenizer keys (fengshen_model_type, fengshen_tokenizer_type), and deciding whether direct imports are safer than auto factories.
• Safe installed-package import checks that do not download models or mutate checkpoints: run python scripts/check_model_imports.py --help from this sub-skill directory.
• Diagnosing errors such as missing cached_path, missing softmax_backward_data, absent SentencePiece/tokenizer files, unsupported custom config keys, model downloads in offline mode, model size on CPU, or optional DeepSpeed imports.

Route elsewhere

• Training loops, dataloaders, metrics, optimizer/scheduler arguments, checkpoints, and distributed/Megatron training go to ../data-training/SKILL.md.
• Example recipes, model-family command conversion, Ziya conversion, Taiyi diffusion recipes, and checkpoint mutation/conversion planning go to ../examples-conversion/SKILL.md.
• fengshen-pipeline CLI invocation, public pipeline data schemas, and CLI parser behavior go to ../pipelines-cli/SKILL.md.

Safety rules

• Do not instantiate from_pretrained on remote model IDs unless the task explicitly authorizes downloads and cache use.
• Prefer import-only, config-only, or local_files_only=True checks for diagnosis.
• Treat CUDA, DeepSpeed, Megatron fused kernels, Taiyi diffusion, Ziya large-model inference, and full checkpoint conversion as optional/heavy paths unless a separate verification plan requires them.
• Do not depend on the original repository checkout. Use only this skill tree, the installed fengshen package, user-provided model directories, and user-authorized caches.