nemo_automodel.components.models.minimax_m3_vl.processing

Module Contents

Functions

Name	Description
`build_minimax_m3_vl_processor`	Load the MiniMax M3 VL processor with its chat template attached.

API

nemo_automodel.components.models.minimax_m3_vl.processing.build_minimax_m3_vl_processor(
    pretrained_model_name_or_path: str,
    kwargs = {}
) -> transformers.processing_utils.ProcessorMixin

Load the MiniMax M3 VL processor with its chat template attached.

The checkpoint ships a chat_template.jinja, but the remote-code MiniMaxVLProcessor.__init__ declares only (image_processor, tokenizer, video_processor, **kwargs) and does not forward chat_template to ProcessorMixin. Transformers’ from_args_and_dict only passes the constructor’s explicit named arguments, so the loaded template is classified as an unused kwarg and dropped, leaving processor.chat_template as None and apply_chat_template raising “does not have a chat template”. Reattach it from the checkpoint’s chat_template.jinja.

Parameters:

pretrained_model_name_or_path

str

Local path (or hub id) of the checkpoint.

**kwargs

Defaults to {}

Forwarded to AutoProcessor.from_pretrained (e.g. trust_remote_code=True).

Returns: ProcessorMixin

The processor with a non-None chat_template when one is available.