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

ArchConfigWithVisionCache

ArchConfigWithVisionCache

class max.pipelines.lib.interfaces.ArchConfigWithVisionCache(*args, **kwargs)

source

Bases: Protocol

Config for a vision-language architecture with a vision encoder cache.

Both hooks are architecture facts derived purely from the HuggingFace config — classmethods, so config construction and memory planning can consult them without building an arch config instance. Architectures without a vision cache simply do not implement this protocol; consumers gate on issubclass. Deliberately standalone (not an ArchConfig refinement) so that gate checks exactly these two hooks.

estimate_vision_cache_entry_bytes()

classmethod estimate_vision_cache_entry_bytes(huggingface_config)

source

Worst-case bytes for one vision encoder cache entry.

The memory a single max-resolution image (or video, for video models) occupies after the vision encoder’s spatial merge / patch merge step. 0 means the checkpoint has no vision cache.

Parameters:

huggingface_config (AutoConfig)

Return type:

int

get_vision_cache_row_spec()

classmethod get_vision_cache_row_spec(huggingface_config)

source

Describes one merged vision token’s embedding row in the cache.

Returns (hidden_size, dtype) for architectures that opt into the block-mode vision cache, or None when the checkpoint has no vision cache.

Parameters:

huggingface_config (AutoConfig)

Return type:

tuple[int, DType] | None