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

ImageMetadata

ImageMetadata

class max.pipelines.context.ImageMetadata(*, start_idx, end_idx, pixel_values, image_hash=None, num_embedding_rows=None)

source

Bases: object

Metadata about an image in the prompt.

Each image corresponds to a range in the text token array [start_idx, end_idx).

Parameters:

• start_idx (int)
• end_idx (int)
• pixel_values (ndarray[tuple[Any, ...], dtype[Any]])
• image_hash (int | None)
• num_embedding_rows (int | None)

embedding_rows

property embedding_rows: int

source

Embedding rows for this entry (span width unless overridden).

end_idx

end_idx: int

source

One after the index of the last <vision_token_id> special token for the image

image_hash

image_hash: int | None = None

source

Hash of the image, for use in prefix caching

num_embedding_rows

num_embedding_rows: int | None = None

source

Embedding rows the encoder emits for this entry.

None means every token in [start_idx, end_idx) is a placeholder, so the row count equals the span width. Set by tokenizers whose spans interleave placeholder runs with other tokens (e.g. video timestamp text).

pixel_values

pixel_values: ndarray[tuple[Any, ...], dtype[Any]]

source

Pixel values for the image.

Can be various dtypes depending on the vision model:

• float32: Original precision
• uint16: BFloat16 bits stored as uint16 (workaround for NumPy’s lack of native bfloat16 support). Reinterpreted as bfloat16 on GPU.

start_idx

start_idx: int

source

Index of the first <vision_token_id> special token for the image