utils.schemas.model
utils.schemas.model
Pydantic models for model input / output, etc. configuration
Classes
| Name | Description |
|---|---|
| BnbBaseQuantConfig | bitsandbytes base-weight quantization for LoRA training. |
| FineGrainedFp8BaseQuantConfig | Structured form of model_quantization_config: FineGrainedFP8Config. |
| ModelInputConfig | Model configuration subset |
| ModelOutputConfig | model save configuration subset |
| ModelQuantizationConfig | Structured discriminator for the base model’s quantization scheme. |
| Mxfp4BaseQuantConfig | Structured form of model_quantization_config: Mxfp4Config. |
| SpecialTokensConfig | Special tokens configuration subset |
| TorchAoBaseQuantConfig | torchao base-weight quantization for LoRA training. |
BnbBaseQuantConfig
utils.schemas.model.BnbBaseQuantConfig()bitsandbytes base-weight quantization for LoRA training.
Replaces the older adapter: qlora + load_in_4bit: true boilerplate
for the common case. nf4 implies 4-bit QLoRA (auto-promotes
adapter: lora to qlora and sets load_in_4bit); int8 stays
as 8-bit LoRA (sets load_in_8bit).
FineGrainedFp8BaseQuantConfig
utils.schemas.model.FineGrainedFp8BaseQuantConfig()Structured form of model_quantization_config: FineGrainedFP8Config.
ModelInputConfig
utils.schemas.model.ModelInputConfig()Model configuration subset
ModelOutputConfig
utils.schemas.model.ModelOutputConfig()model save configuration subset
ModelQuantizationConfig
utils.schemas.model.ModelQuantizationConfig()Structured discriminator for the base model’s quantization scheme.
Exactly one of bnb / torchao / mxfp4 / fp8 must be set.
The legacy string form (model_quantization_config: Mxfp4Config) keeps
working via a normalizer in the top-level validator.
Attributes
| Name | Description |
|---|---|
| backend | Name of the selected discriminator (one of bnb/torchao/mxfp4/fp8). |
Mxfp4BaseQuantConfig
utils.schemas.model.Mxfp4BaseQuantConfig()Structured form of model_quantization_config: Mxfp4Config.
Pass-through config_kwargs go straight to transformers.Mxfp4Config.
SpecialTokensConfig
utils.schemas.model.SpecialTokensConfig()Special tokens configuration subset
TorchAoBaseQuantConfig
utils.schemas.model.TorchAoBaseQuantConfig()torchao base-weight quantization for LoRA training.
Compile- and FSDP2-friendly alternative to bitsandbytes. 4-bit dtypes
(int4 / nf4 / nvfp4) auto-promote the adapter to qlora;
int8 / fp8 stay as weight-only LoRA. mxfp4 is rejected here
because torchao has no weight-only flavour for arbitrary linears — for
MoE experts use quantize_moe_experts: true instead.