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 module

max.pipelines.lib.registration_runtime

Completes a registered module’s config with its deployment.

A config built from the checkpoint doesn’t know where it runs. After the config is finalized, and before any module is constructed, the framework hands it a ModuleRuntime: the device mesh, the batch limits the memory planner chose and, under expert parallelism, the initialized EPConfig.

A registered config takes part by subclassing DeploymentConfig, or ExpertParallelDeploymentConfig when it serves its experts across devices. Their fields declare what the deployment adds to the decoder’s inputs and how the checkpoint splits between roles.

Deployment

DeploymentConfig	A registered config that the deployment completes.
ExpertParallelDeploymentConfig	A registered config whose decoder serves its experts across devices.
ModuleRuntime	The deployment a registered module's config is completed with.

Expert parallelism

ExpertParallelExchange	The initialized exchange's buffers, as each batch passes them.
ExpertParallelInputs	How a decoder's graph takes the expert-parallel exchange.

ep_input_tree	Returns the decoder's ep input types for an initialized config.
initialize_ep_comm_buffers	Sizes and initializes the expert-parallel communication buffers.
uses_expert_parallel_exchange	Returns whether expert sharding requires exchange initialization.

Weights

prepare_weights	Prepares a checkpoint before selecting registered role weights.
role_weights	Returns the weights of one role, named as its module names them.

Graph constants

ConstantInputs	The declared graph types and lifetime-owned deployment buffers.

Batch context lengths

BatchContextLengths	Each replica's page-aligned KV context length, refreshed per batch.

batch_context_length_types	Returns the batch_context_lengths input types.