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 class

AutoProcessor

AutoProcessor

class max.pipelines.AutoProcessor(model_path, *, revision=None, trust_remote_code=False)

source

Bases: object

Prepares chat conversations with images for AutoModel.

Calling it returns the messages and images that AutoModel.generate() takes, not Hugging Face’s input_ids and pixel_values. The model’s own tokenizer applies the chat template and expands each image at generate time, as serve does. Decoding uses the model’s Hugging Face tokenizer.

from PIL import Image

processor = AutoProcessor("google/gemma-3-4b-it")
model = AutoModel("google/gemma-3-4b-it", max_batch_size=2)
conversations = [
    [{"role": "user", "content": [
        {"type": "image", "image": Image.open("cat.png")},
        {"type": "text", "text": "Describe this image."},
    ]}],
    [{"role": "user", "content": "What is 2 + 2?"}],
]
outputs = model.generate(**processor(conversations), max_new_tokens=32)
new_tokens = [
    sequence[length:]
    for sequence, length in zip(
        outputs.sequences, outputs.prompt_lengths
    )
]
print(processor.batch_decode(new_tokens, skip_special_tokens=True))

Parameters:

• model_path (str) – A Hugging Face repo ID or a local model directory.
• revision (str | None) – The repo revision to load.
• trust_remote_code (bool) – Whether to run tokenizer code from the repo.

batch_decode()

batch_decode(sequences, *, skip_special_tokens=False)

source

Returns the text for each token-id list in sequences.

Parameters:

• sequences (Iterable[Sequence[int]])
• skip_special_tokens (bool)

Return type:

list[str]

decode()

decode(token_ids, *, skip_special_tokens=False)

source

Returns the text for token_ids.

Parameters:

• token_ids (Sequence[int])
• skip_special_tokens (bool)

Return type:

str