For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python function

resolve_hf_quant_config

resolve_hf_quant_config()

max.pipelines.weights.resolve_hf_quant_config(huggingface_config, state_dict)

source

Resolves a checkpoint’s Hugging Face quantization config.

Tries, in order: the config’s own quantization_config field, a standalone hf_quant_config.json beside the weights, and finally a sniff of the state dict for modelopt NVFP4’s weight_scale_2 tensors.

Parameters:

• huggingface_config (AutoConfig) – The config to read quantization_config from. For a nested multimodal config this must be the top-level one – the sub-config does not carry the field.
• state_dict (Mapping[str, WeightData]) – The checkpoint weights, used only by the final sniff.

Returns:

The resolved quantization config, or None when the checkpoint declares no quantization.

Return type:

dict[str, Any] | None