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

ExpertParallelInputs

`ExpertParallelInputs`

class max.pipelines.lib.registration_runtime.ExpertParallelInputs

source

How a decoder’s graph takes the expert-parallel exchange.

The default takes EPBatchManager’s inputs: both buffer groups’ atomic counters on every device, then their send, receive and count pointer tables. A config whose decoder takes them in another layout sets ExpertParallelDeploymentConfig.expert_parallel_inputs to a subclass.

exchange(initializer)

source

Returns the buffers of initializer the graph takes.

input_tree(ep)

source

Returns the decoder’s ep input types for ep.