IMPORTANT: To view this page as Markdown, append `.md` to the URL (e.g. /get-started.md). For the complete documentation index, see llms.txt.
Skip to main content
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.architectures.llama3

Llama 3 transformer architecture for text generation.

Llama3BatchProcessor​

class max.pipelines.architectures.llama3.Llama3BatchProcessor(config, runtime)

source

Bases: RaggedBatchProcessor[TextContext, Llama3Inputs]

Ragged batching with pinned host buffers and optional DP / LoRA.

Parameters:

get_symbolic_inputs()​

get_symbolic_inputs(*, kv_params, device_refs)

source

Returns non-KV graph input types in execution order.

Parameters:

Return type:

list[TensorType | BufferType]

prepare_initial_token_inputs()​

prepare_initial_token_inputs(replica_batches, kv_cache_inputs=None, return_n_logits=1)

source

Prepares inputs for the first execution step of a batch.

Parameters:

  • replica_batches (Sequence[Sequence[TextContext]])
  • kv_cache_inputs (KVCacheInputsInterface[Buffer, Buffer] | None)
  • return_n_logits (int)

Return type:

Llama3Inputs

process_outputs()​

process_outputs(outputs)

source

Maps raw Model.execute buffers to ModelOutputs.

Parameters:

outputs (Sequence[Buffer | object])

Return type:

ModelOutputs

Llama3Config​

class max.pipelines.architectures.llama3.Llama3Config(*, hidden_size, num_attention_heads, num_key_value_heads, num_hidden_layers, rope_theta, rope_scaling_params, max_seq_len, intermediate_size, interleaved_rope_weights, vocab_size, dtype, model_quantization_encoding, quantization_config, kv_params, return_logits=ReturnLogits.LAST_TOKEN, norm_method='rms_norm', norm_dtype=None, attention_bias=False, rms_norm_eps=None, tie_word_embeddings=False, stacked_mlp=False, stacked_qkv=False, attention_multiplier, embedding_multiplier, residual_multiplier, devices, clip_qkv, quant_config=None, lora_config=None, longrope_scaling_params=None, logits_scaling=1.0, return_hidden_states=ReturnHiddenStates.NONE, target_layer_ids=None, use_subgraphs=True, data_parallel_degree=1, sliding_window=None, quantization_encoding=None)

source

Bases: ArchConfigWithStoredKVParams, ArchConfigWithKVCache

Model configuration for Llama3 graph construction/execution.

Parameters:

DEFAULT_ENCODING​

DEFAULT_ENCODING: ClassVar[SupportedEncoding] = 'q4_k'

source

SUPPORTED_ENCODINGS​

SUPPORTED_ENCODINGS: ClassVar[set[SupportedEncoding]] = {'bfloat16', 'float32', 'float4_e2m1fnx2', 'float8_e4m3fn', 'gptq', 'q4_0', 'q4_k', 'q6_k'}

source

attention_bias​

attention_bias: bool = False

source

attention_multiplier​

attention_multiplier: float

source

calculate_attention_multiplier()​

static calculate_attention_multiplier(huggingface_config)

source

The attention multiplier is a scalar that scales the attention scores. It is used to control the variance of the attention scores.

This function is used to get the attention multiplier from the huggingface config. If the attention multiplier is not set, it will be calculated as the square root of 1.0 divided by the head dimension.

Parameters:

huggingface_config (AutoConfig)

Return type:

float

clip_qkv​

clip_qkv: float | None

source

construct_kv_params()​

classmethod construct_kv_params(huggingface_config, pipeline_config, devices, kv_cache_config, cache_dtype)

source

Grouped-attention KV with EAGLE draft-token count when speculative is on.

Parameters:

Return type:

KVCacheParams

data_parallel_degree​

data_parallel_degree: int = 1

source

devices​

devices: list[DeviceRef]

source

dtype​

dtype: DType

source

embedding_multiplier​

embedding_multiplier: float

source

finalize()​

finalize(huggingface_config, state_dict, return_logits, return_hidden_states=ReturnHiddenStates.NONE, norm_method='rms_norm', attention_bias=False)

source

Define parameters that can’t be determined just from the pipeline config.

Parameters:

Return type:

None

get_num_layers()​

static get_num_layers(huggingface_config)

source

Layer count for the decoder stack (override when HF uses a different field).

Parameters:

huggingface_config (AutoConfig)

Return type:

int

hidden_size​

hidden_size: int

source

initialize()​

classmethod initialize(pipeline_config, model_config=None)

source

Initialize the config from a PipelineConfig.

Parameters:

  • pipeline_config (PipelineConfig) – The pipeline configuration.
  • model_config (MAXModelConfig | None) – The model configuration to read from. When None (the default), pipeline_config.model is used. Pass an explicit config (e.g. pipeline_config.draft_model) to initialize the arch config for a different model.

Return type:

Self

initialize_from_config()​

classmethod initialize_from_config(pipeline_config, huggingface_config, model_config=None)

source

Parameters:

Return type:

Self

interleaved_rope_weights​

interleaved_rope_weights: bool

source

intermediate_size​

intermediate_size: int

source

kv_params​

kv_params: KVCacheParams

source

logits_scaling​

logits_scaling: float = 1.0

source

longrope_scaling_params​

longrope_scaling_params: LongRoPEScalingParams | None = None

source

lora_config​

lora_config: LoRAConfig | None = None

source

max_seq_len​

max_seq_len: int

source

model_quantization_encoding​

model_quantization_encoding: QuantizationEncoding | None

source

norm_dtype​

norm_dtype: DType | None = None

source

norm_method​

norm_method: Literal['rms_norm'] | Literal['layer_norm'] = 'rms_norm'

source

num_attention_heads​

num_attention_heads: int

source

num_hidden_layers​

num_hidden_layers: int

source

num_key_value_heads​

num_key_value_heads: int

source

quant_config​

quant_config: QuantConfig | None = None

source

quantization_config​

quantization_config: QuantizationConfig | None

source

quantization_encoding​

quantization_encoding: SupportedEncoding | None = None

source

residual_multiplier​

residual_multiplier: float

source

return_hidden_states​

return_hidden_states: ReturnHiddenStates = 'none'

source

return_logits​

return_logits: ReturnLogits = 'last_token'

source

rms_norm_eps​

rms_norm_eps: float | None = None

source

rope_scaling_params​

rope_scaling_params: Llama3RopeScalingParams | None

source

rope_theta​

rope_theta: float

source

sliding_window​

sliding_window: int | None = None

source

stacked_mlp​

stacked_mlp: bool = False

source

stacked_qkv​

stacked_qkv: bool = False

source

target_layer_ids​

target_layer_ids: list[int] | None = None

source

tie_word_embeddings​

tie_word_embeddings: bool = False

source

use_subgraphs​

use_subgraphs: bool = True

source

vocab_size​

vocab_size: int

source

Llama3Inputs​

class max.pipelines.architectures.llama3.Llama3Inputs(tokens, input_row_offsets, signal_buffers, return_n_logits, data_parallel_splits=None, *, kv_cache_inputs=None, lora=None, lora_buffers=(), vision_embeddings=<factory>, vision_scatter_indices=<factory>, hidden_states=None)

source

Bases: ModelInputs

A class representing inputs for the Llama3 model.

This class encapsulates the input tensors required for the Llama3 model execution.

Parameters:

buffers​

property buffers: tuple[Buffer, ...]

source

Returns positional Buffer inputs for model ABI calls.

data_parallel_splits​

data_parallel_splits: Buffer | Sequence[Sequence[int]] | None = None

source

Tensor containing the data parallel splits.

input_row_offsets​

input_row_offsets: Buffer

source

Tensor containing the offsets for each row in the ragged input sequence.

return_n_logits​

return_n_logits: Buffer

source

signal_buffers​

signal_buffers: list[Buffer]

source

Device buffers used for synchronization in communication collectives.

tokens​

tokens: Buffer

source

Tensor containing the input token IDs.

Llama3Model​

class max.pipelines.architectures.llama3.Llama3Model(pipeline_config, session, devices, kv_cache_config, weights, adapter=None, return_logits=ReturnLogits.LAST_TOKEN, return_hidden_states=ReturnHiddenStates.NONE, max_batch_size=1)

source

Bases: LlamaModelBase

Llama 3 pipeline model implementation.

Parameters:

config_class​

config_class

source

alias of Llama3Config

norm_method​

norm_method: Literal['rms_norm'] | Literal['layer_norm'] = 'rms_norm'

source

Normalization layer.

LlamaModelBase​

class max.pipelines.architectures.llama3.LlamaModelBase(pipeline_config, session, devices, kv_cache_config, weights, adapter=None, return_logits=ReturnLogits.LAST_TOKEN, return_hidden_states=ReturnHiddenStates.NONE, max_batch_size=1)

source

Bases: LogProbabilitiesMixin, GraphPipelineModelWithKVCache[TextContext]

Base Llama pipeline model implementation.

Parameters:

attention_bias​

attention_bias: bool = False

source

Whether to use attention bias.

batch_processor_cls​

batch_processor_cls

source

alias of Llama3BatchProcessor

execute()​

execute(model_inputs)

source

Executes the graph with the given inputs.

Parameters:

model_inputs (ModelInputs) – The model inputs to execute, containing tensors and any other required data for model execution.

Returns:

ModelOutputs containing the pipeline’s output tensors.

Return type:

ModelOutputs

This is an abstract method that must be implemented by concrete PipelineModels to define their specific execution logic.

model​

model: Model

source

Compiled and initialized model ready for inference.

model_config_cls​

model_config_cls

source

alias of Llama3Config

norm_method​

norm_method: Literal['rms_norm'] | Literal['layer_norm']

source

Normalization layer.

prepare_initial_token_inputs()​

prepare_initial_token_inputs(replica_batches, kv_cache_inputs=None, return_n_logits=1)

source

Delegates to the batch processor and narrows to Llama3Inputs.

Parameters:

Return type:

Llama3Inputs

state_dict​

state_dict: dict[str, Any]

source

Weights to load into the model.

Models

Cookbook

Puzzles

Blog

Community

Contact