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

PreprocessedImageProbe

PreprocessedImageProbe

class max.pipelines.modeling.types.PreprocessedImageProbe(*args, **kwargs)

source

Bases: Protocol

Optional tokenizer capability: report already-preprocessed images.

A tokenizer that caches preprocessed image tensors implements this so the API server can skip the pixel decode for an image whose tensor it already holds – nothing downstream reads those pixels. A tokenizer without such a cache simply does not implement it, and every image is decoded as before.

Declared as a protocol rather than read off the tokenizer with getattr, because unlike the model-specific media limits the route also reads (plain data attributes with safe defaults) this is a callable carrying a positional-argument contract and a length invariant. The protocol is what states them.

preprocessed_image_mask()

preprocessed_image_mask(images, messages)

source

Which of images this tokenizer can already serve preprocessed.

Must not decode, mutate or cache anything: the API server calls it on the event loop before deciding what to decode, and the answer is only a hint. An entry may be evicted between this call and tokenization, so a True that later misses must remain correct (the image is simply decoded then).

Parameters:

• images (list[bytes]) – Raw encoded bytes per image, in request order.
• messages (list[TextGenerationRequestMessage]) – The request’s messages, carrying any per-image sizing hints that the cache key folds in.

Returns:

One flag per entry in images, in the same order. Returning a different length is a contract violation.

Return type:

Sequence[bool]