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

DeploymentConfig

DeploymentConfig

class max.pipelines.lib.registration_runtime.DeploymentConfig

source

Bases: object

A registered config that the deployment completes.

A registered config that subclasses it gets apply_runtime() called before any module is constructed, and declares through its class fields what else the framework does for it. A config that doesn’t subclass it takes every default.

apply_model_load_context()

apply_model_load_context(context)

source

Returns the config completed with composition-only settings.

Parameters:

context (ModelLoadContext)

Return type:

Self

apply_runtime()

apply_runtime(runtime)

source

Returns the config completed with runtime.

The default needs nothing from the deployment.

Parameters:

runtime (ModuleRuntime)

Return type:

Self

expert_parallel_communication

expert_parallel_communication: ClassVar[bool] = True

source

Whether expert parallelism uses a runtime exchange. A config that disables it validates local expert sharding in apply_runtime().

export_ep_init

export_ep_init: ClassVar[bool] = False

source

Whether artifacts include the exchange initialization graph.

export_records()

export_records(runtime)

source

Returns what an exporter records of the deployment, by name.

The serving pipeline model keeps them as its export_records, where an exporter reads them. The default records nothing.

Parameters:

runtime (ModuleRuntime)

Return type:

Mapping[str, object]

prepare_weights()

prepare_weights(state_dict)

source

Returns checkpoint weights prepared for the registered modules.

Serving and composed decoders apply this before role prefixes and parameter dtype conversion. The default preserves every weight.

Parameters:

state_dict (Mapping[str, WeightData])

Return type:

dict[str, WeightData]

preserve_checkpoint_keys

preserve_checkpoint_keys: ClassVar[bool] = False

source

Whether finalization reads raw keys before prepare_weights() performs the architecture’s declared weight transforms.

takes_batch_context_lengths

takes_batch_context_lengths: ClassVar[bool] = False

source

Whether the decoder’s graph takes batch_context_lengths, as the MLA family does to size its prefill buffers without a device-to-host copy.

weight_prefixes

weight_prefixes: ClassVar[Mapping[str, str]] = mappingproxy({})

source

The checkpoint prefix each role’s weights sit under, keyed by GraphRole, when one checkpoint holds the weights of several roles. A role without one takes every weight (role_weights()).