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

GraphRole

GraphRole

class max.pipelines.lib.graph_specs.GraphRole(value, names=<not given>, *values, module=None, qualname=None, type=None, start=1, boundary=None)

source

Bases: str, Enum

The roles of the graphs MAX’s pipelines run.

DECODER

DECODER = 'decoder'

source

The forward pass: tokens and KV cache to logits.

DRAFT

DRAFT = 'draft'

source

A speculative proposal model, composed with a target decoder.

ENCODER

ENCODER = 'encoder'

source

The single forward pass of a model without a KV cache, such as an embedding model: inputs to embeddings or hidden states.

EP_INIT

EP_INIT = 'ep_init'

source

The one-time expert-parallel setup graph run at engine startup.

SAMPLER

SAMPLER = 'sampler'

source

Samples the next token from the decoder’s logits.

SAMPLER_BITMASK

SAMPLER_BITMASK = 'sampler_bitmask'

source

Samples the next token, masked by a structured-output bitmask.

SPEC_DECODER

SPEC_DECODER = 'spec_decoder'

source

One speculative-decoding step, the target and draft models fused: tokens, the previous drafts and KV cache to the accepted tokens and the next drafts.

VISION

VISION = 'vision'

source

The vision encoder: image patches to soft-token embeddings.

VISION_PREPROCESS

VISION_PREPROCESS = 'vision_preprocess'

source

One image’s preprocessing, on the CPU: decoded pixels and the size to resize them to, to the vision encoder’s inputs for that image. Only max export builds it, for a serving engine; MAX serve preprocesses in its tokenizers.

VISION_PREPROCESS_VIDEO

VISION_PREPROCESS_VIDEO = 'vision_preprocess_video'

source

One encoder attention segment of a video clip’s preprocessing, on the CPU: the segment’s decoded frames and the sizes to resize them to, to the vision encoder’s inputs for that segment. Only max export builds it, for a model whose segment is not a one-frame image.