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

Interfaces for running models end to end: registry, configuration, and pipeline classes for text, embeddings, and pixel generation.

A pipeline wires a model family into an executable workflow. A SupportedArchitecture registers the model in the pipeline registry, PipelineConfig sets how it runs, and pipeline classes such as TextGenerationPipeline, EmbeddingsPipeline, and PixelGenerationPipeline drive tokenization, weight loading, and generation.

Submodules

max.pipelines.architectures	Built-in model architectures for MAX pipelines.
max.pipelines.audio	Audio generation pipelines.
max.pipelines.context	Request contexts and supporting types for MAX pipelines.
max.pipelines.diffusion	Diffusion-specific pipeline components.
max.pipelines.kv_cache	KV cache management for MAX pipelines.
max.pipelines.lib	Types to interface with ML pipelines such as text/token generation.
max.pipelines.lib.arch_lookup	Architecture registration and lookup tables.
max.pipelines.lib.interfaces	Interfaces for MAX pipelines.
max.pipelines.lib.log_probabilities	Builds computation graphs for log probabilities over batched input sequences.
max.pipelines.lib.registry	Model registry, for tracking various model variants.
max.pipelines.modeling.base	Base classes for MAX pipeline model definitions.
max.pipelines.modeling.dataprocessing	Token-batch preparation helpers: collation, padding, and attention masks.
max.pipelines.modeling.types	Pipeline modeling types: request/input/pipeline interfaces.
max.pipelines.weights	Weight loading utilities for MAX pipelines.
max.pipelines.lora	LoRA adapter management for MAX pipelines.
max.pipelines.request	Request types for MAX API.
max.pipelines.sampling	Token sampling for MAX pipelines: samplers, logits processors, and the acceptance runners used in speculative decoding.
max.pipelines.speculative	Speculative decoding pipelines and configuration for MAX.

Configuration

KVCacheConfig	Configuration for the paged KV cache.
MAXModelConfig	Configuration for a pipeline model.
PipelineArgs	User-settable input arguments for a pipeline.
PipelineConfig	Configuration for a pipeline.
PipelineRuntimeConfig	Model-agnostic runtime settings for pipeline execution.
ProfilingConfig	Configuration for the GPU (NVTX/Nsight) profiler and kernel tracing.
SamplingConfig	Configuration for the sampling stage of token generation.
SpeculativeConfig	Configures speculative decoding for a pipeline.

In-process generation

AutoModel	Loads a text generation model in-process.
AutoPipeline	Generates text from text with an AutoModel.
AutoTokenizer	Loads the Hugging Face tokenizer serve uses to encode a model's text.

Pipelines

EmbeddingsPipeline	Generalized token generator pipeline.
PixelGenerationPipeline	Pixel generation pipeline for diffusion models.
TextGenerationPipeline	Generalized token generator pipeline.
TextGenerationPipelineInterface	Interface for text generation pipelines.

Model interface

MemoryEstimator	Plans device memory for a pipeline and estimates what it will use.
ModelInputs	Base class for model inputs.
ModelOutputs	Pipeline model outputs.
PipelineModel	A pipeline model with setup, input preparation and execution methods.

Eager usage validation

The eager usage validator adds guard rails around eager code running inside a compiled pipeline. Configure it with the runtime.eager_usage_validator MAX config setting.

EagerUsageValidator	A scope that tracks production-unsafe API usages.
EagerValidatorMode	Configures when the eager usage validator runs.
Phase	Which part of a pipeline's life a validation scope covers.

eager_usage_validator	Builds the validator guarding phase, per the configured mode.

eager_validator	Use to opt in to the escape hatches for EagerUsageValidator.

Tokenizers

IdentityPipelineTokenizer	A pass-through tokenizer that returns prompts unchanged.
TextAndVisionTokenizer	Encapsulates creation of TextAndVisionContext and specific token encode/decode logic.
TextTokenizer	Encapsulates creation of TextContext and specific token encode/decode logic.

Enums

PipelineRole	Indicates whether the pipeline should do prefill and/or decode.
PrometheusMetricsMode	Controls the Prometheus metrics mode.
RepoType	Specifies the source location type of a model repository.
RopeType	The rotary position embedding (RoPE) implementation a model configures.
SupportedEncoding	All possible encodings which may be supported by a particular model.

Utilities

download_weight_files	Downloads weight files for a Hugging Face model and returns local paths.
is_float4_encoding	Returns whether the given encoding is a float4 type.
parse_supported_encoding_from_file_name	Infers a SupportedEncoding from a file name string.
supported_encoding_dtype	Returns the underlying model dtype for the given encoding.
supported_encoding_quantization	Returns the QuantizationEncoding for the given encoding.
supported_encoding_supported_devices	Returns the devices that the given encoding is supported on.
supported_encoding_supported_on	Returns whether the given encoding is supported on a device.
upper_bounded_default	Returns a value not exceeding the upper bound.

ADAPTER_CONFIG_FILE	File name of the LoRA adapter's config file inside the adapter's local directory, read when the adapter is loaded.