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.interfaces

Interfaces for MAX pipelines.

Architecture config

ArchConfig	Config for a model architecture.
ArchConfigWithAttentionKVCache	Predefined configuration for architectures that use attention KV cache blocks.
ArchConfigWithBoundedMaxSeqLen	Mixin for configs that store the received max_seq_len.
ArchConfigWithKVCache	Config for a model architecture that uses a KV cache.
ArchConfigWithPermissiveMaxSeqLen	Mixin for configs that honor max_length without bounding.
ArchConfigWithStoredKVParams	Mixin that implements get_kv_params() as the kv_params field.
ArchConfigWithVisionCache	Config for a vision-language architecture with a vision encoder cache.
ArchVLConfigWithTextSubconfig	Mixin for VLMs that embed a language-model arch config.

Batch processing

BatchProcessor	Batches pipeline contexts into model inputs and parses execution outputs.
BatchProcessorRuntime	Runtime dependencies shared by batch processors.
ModuleV3SingleReplicaBatchProcessor	Single-replica ragged KV batching for ModuleV3 compile-path models.
RaggedBatchProcessor	Base for ragged KV text batching.
SingleReplicaRaggedBatchProcessor	Single-replica ragged KV batching for Graph-path models (no DP / LoRA).

Functions

arch_has_vision_tower	Whether this architecture encodes images at all.
build_single_replica_ragged_token_arrays	Builds concatenated token and row-offset arrays for a ragged batch.
modulev3_ragged_kv_symbolic_input_tree	Returns the named inputs of a ModuleV3 ragged KV model.
ragged_batch_inputs	Returns the named per-batch graph inputs, in the order fields gives.
ragged_kv_symbolic_input_tree	Returns the named graph input types of a standard ragged KV text model.
signal_inputs	Returns the named per-device allreduce signal buffers.
single_replica_context_batch	Returns the sole replica batch or raises when DP is unsupported.
vision_merge_inputs	Returns the per-device vision embeddings and where they merge.