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.gemma3

Gemma 3 transformer architecture for text generation.

Gemma3Config​

class max.pipelines.architectures.gemma3.Gemma3Config(*, vocab_size, hidden_size, intermediate_size, num_hidden_layers, num_attention_heads, num_key_value_heads, head_dim, hidden_activation, max_position_embeddings, rms_norm_eps, rope_theta, attention_bias, query_pre_attn_scalar, sliding_window, final_logit_softcapping, attn_logit_softcapping, rope_scaling, rope_local_base_freq, sliding_window_pattern, dtype, devices, interleaved_rope_weights, return_logits=ReturnLogits.LAST_TOKEN, kv_params, tie_word_embeddings=False, quant_config=None, quantization_encoding=None)

source

Bases: ArchConfigWithPermissiveMaxSeqLen, ArchConfigWithStoredKVParams, ArchConfigWithKVCache

Represents the MAX Engine configuration for Gemma 3 models.

Contains parameters specific to the Gemma 3 architecture (typically extracted from HuggingFace configs), plus MAX-specific runtime settings and helpers.

Parameters:

DEFAULT_ENCODING​

DEFAULT_ENCODING: ClassVar[SupportedEncoding] = 'bfloat16'

source

SUPPORTED_ENCODINGS​

SUPPORTED_ENCODINGS: ClassVar[set[SupportedEncoding]] = {'bfloat16'}

source

attention_bias​

attention_bias: bool

source

Whether to use a bias in the query, key, value and output projection layers during self-attention.

attn_logit_softcapping​

attn_logit_softcapping: int | None

source

Scaling factor when applying tanh softcapping on the attention scores.

devices​

devices: list[DeviceRef]

source

Devices to run the model with.

dtype​

dtype: DType

source

DType of the model weights and input.

final_logit_softcapping​

final_logit_softcapping: float | None

source

Scaling factor when applying tanh softcapping on the logits.

finalize()​

finalize(huggingface_config, state_dict, return_logits, quant_config=None)

source

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

This method sets fields that require introspection of the model weights (state_dict), such as tie_word_embeddings and quant_config.

Parameters:

  • huggingface_config (AutoConfig) – The HuggingFace model configuration object.
  • state_dict (dict[str, WeightData]) – The model’s state dictionary containing weights.
  • return_logits (ReturnLogits) – Whether to return the last token, all tokens or a variable number of logits.
  • quant_config (QuantConfig | None) – Scaled quantization configuration (optional).

Return type:

None

get_num_layers()​

static get_num_layers(huggingface_config)

source

Retrieves the number of hidden layers from the HuggingFace configuration.

Parameters:

huggingface_config (AutoConfig) – The HuggingFace model configuration object (transformers.AutoConfig).

Returns:

The number of hidden layers specified in the configuration’s text config.

Return type:

int

head_dim​

head_dim: int

source

The attention head dimension.

hidden_activation​

hidden_activation: str

source

The non-linear activation function (function or string) in the decoder. Will default to “gelu_tanh” if not specified. “gelu_tanh” uses an approximation of the “gelu” activation function.

hidden_size​

hidden_size: int

source

Dimension of the hidden representations.

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)

source

Initializes a Gemma3Config instance from pipeline and HuggingFace configuration.

This method creates a config instance with all fields that can be determined from the pipeline and HuggingFace configuration, without needing the state_dict. Fields that depend on the state_dict (like tie_word_embeddings, quant_config) should be set via the finalize() method.

Parameters:

  • pipeline_config (PipelineConfig) – The MAX Engine pipeline configuration.
  • huggingface_config (AutoConfig) – The HuggingFace model configuration object.

Returns:

An initialized Gemma3Config instance.

Return type:

Self

interleaved_rope_weights​

interleaved_rope_weights: bool

source

True if the rope weights are in interleaved complex format.

intermediate_size​

intermediate_size: int

source

Dimension of the MLP representations.

kv_params​

kv_params: KVCacheParams

source

KV cache parameters.

max_position_embeddings​

max_position_embeddings: int

source

The maximum sequence length that this model might ever be used with.

num_attention_heads​

num_attention_heads: int

source

Number of attention heads for each attention layer in the Transformer decoder.

num_hidden_layers​

num_hidden_layers: int

source

Number of hidden layers in the Transformer decoder.

num_key_value_heads​

num_key_value_heads: int

source

Number of key_value heads that should be used to implement Grouped Query Attention.

quant_config​

quant_config: QuantConfig | None = None

source

Scaled quantization configuration.

quantization_encoding​

quantization_encoding: SupportedEncoding | None = None

source

The resolved quantization encoding the model runs with.

query_pre_attn_scalar​

query_pre_attn_scalar: float | None

source

Scaling factor used on the attention scores.

return_logits​

return_logits: ReturnLogits = 'last_token'

source

Whether to return the last token, all logits, or a variable number of logits.

rms_norm_eps​

rms_norm_eps: float

source

The epsilon used by the rms normalization layers.

rope_local_base_freq​

rope_local_base_freq: float

source

The base period of the RoPE embeddings for local attention.

rope_scaling​

rope_scaling: LinearScalingParams | None

source

Scaling configuration for the RoPE embeddings used in global attention.

rope_theta​

rope_theta: float

source

The base period of the RoPE embeddings.

sliding_window​

sliding_window: int

source

In the Gemma3 language model, every other layer uses sliding window attention. This is the size of the sliding window.

sliding_window_pattern​

sliding_window_pattern: int

source

Pattern for the sliding window attention.

tie_word_embeddings​

tie_word_embeddings: bool = False

source

Whether to tie weight embeddings. When true, the output linear layer uses the same weight as the embedding layer.

vocab_size​

vocab_size: int

source

Vocabulary size of the Gemma3Text model.

Gemma3Inputs​

class max.pipelines.architectures.gemma3.Gemma3Inputs(tokens, input_row_offsets, signal_buffers, return_n_logits, *, 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 Gemma3 model.

This class encapsulates the input tensors required for the Gemma3 model execution.

Parameters:

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

Number of logits to return.

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.

Gemma3Model​

class max.pipelines.architectures.gemma3.Gemma3Model(pipeline_config, session, devices, kv_cache_config, weights, adapter=None, return_logits=ReturnLogits.LAST_TOKEN, max_batch_size=1)

source

Bases: LogProbabilitiesMixin, AlwaysSignalBuffersMixin, GraphPipelineModelWithKVCache[TextContext]

A Gemma 3 pipeline model for text generation.

This class integrates the Gemma 3 architecture with the MAX Engine pipeline infrastructure, handling model loading, KV cache management, and input preparation for inference.

Parameters:

batch_processor_cls​

batch_processor_cls

source

alias of Gemma3BatchProcessor

execute()​

execute(model_inputs)

source

Executes the Gemma 3 model with the prepared inputs.

Parameters:

model_inputs (ModelInputs) – The prepared inputs for the model execution, typically including token IDs, attention masks/offsets, and KV cache inputs.

Returns:

An object containing the output logits from the model execution.

Return type:

ModelOutputs

get_num_layers()​

classmethod get_num_layers(huggingface_config)

source

Gets the number of hidden layers from the HuggingFace configuration.

Delegates to the Gemma3Config.get_num_layers static method.

Parameters:

huggingface_config (AutoConfig) – The HuggingFace model configuration object (transformers.AutoConfig).

Returns:

The number of hidden layers.

Return type:

int

model​

model: Model

source

The compiled and initialized MAX Engine model ready for inference.

model_config_cls​

model_config_cls

source

alias of Gemma3Config

Models

Cookbook

Puzzles

Blog

Community

Contact