models
Modules
Snowflake Arctic specs (trust-remote-code model type |
|
DBRX specs (HF model type |
|
DeepSeek-MoE specs (trust-remote-code model type |
|
Gemma 1/2/3 specs (HF model types |
|
Gemma4 specs (HF model type |
|
GPT-OSS specs (HF model type |
|
Llama specs (HF model type |
|
Mixtral specs (HF model type |
|
Nemotron specs (HF model type |
|
Nemotron-H specs (HF model type |
|
Qwen2-MoE specs (HF model type |
|
Qwen3 (dense) specs (HF model type |
|
Qwen3.5-MoE specs (HF model type |
|
Qwen3-MoE specs (HF model type |
|
Qwen3-Next specs (HF model type |
|
Registry indexing the per-model |
|
Per-model descriptor classes. |
Per-model descriptors, one module per HF model type.
Module names mirror the transformers.models layout
(https://github.com/huggingface/transformers/tree/main/src/transformers/models);
trust-remote-code models (arctic, deepseek) use their config model_type.
Each model module registers one global ModelSpec (see specs.py); importing
this package registers them all. Consumers resolve a spec via the registry lookups
(get_spec / match_moe_block) and read its fields; an unmatched lookup
returns None so callers fail loudly or fall back per their own policy.