transformer package

The transformer package provides a customizable and configurable implementation of the transformer model architecture. Each component of a transformer stack, from entire layers down to individual linear layers, can be customized by swapping in different PyTorch modules using the “spec” parameters (see here). The configuration of the transformer (hidden size, number of layers, number of attention heads, etc.) is provided via a TransformerConfig object.

This is the entire attention portion, either self or cross attention, of a transformer layer including the query, key, and value projections, a “core” attention calculation (e.g. dot product attention), and final output linear projection.

class core.transformer.attention.Attention(*args: Any, **kwargs: Any)

Bases: megatron.core.transformer.module.MegatronModule, abc.ABC

Attention layer abstract class.

This layer only contains common modules required for the “self attn” and “cross attn” specializations.

flash_decode(sequence_len_offset: torch.Tensor, query_layer: torch.Tensor, key_layer: torch.Tensor, value_layer: torch.Tensor, inference_key_memory: torch.Tensor, inference_value_memory: torch.Tensor, rotary_cos: torch.Tensor, rotary_sin: torch.Tensor, rotary_interleaved: bool = False)

The flash decoding kernel will do the following in a single execution: 1. Compute RoPE embedding with precomputed cos & sin tensors 2. Update the KV Cache 3. Performs the flash attention operation

flash_decode_and_prefill(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, max_seqlen_q, max_seqlen_k, cu_seqlens_q, cu_seqlens_k, seqlens_k, block_table) → torch.Tensor

Flash attention kernel for mixed decode and prefill samples.

Parameters

• q (Tensor) – Query tensor.

• k (Tensor) – Key tensor.

• v (Tensor) – Value tensor.

• max_seqlen_q (int) – Query total sequence length.

• max_seqlen_k (int) – Key total sequence length.

• cu_seqlens_q (Tensor) – Cumulative query sequence lengths.

• cu_seqlens_k (Tensor) – Cumulative key sequence lengths.

• seqlens_k (Tensor) – key sequence lengths.

• block_table (Tensor) – KV cache chunk ids for all samples.

Returns

(Tensor) Attention output.

forward(hidden_states: torch.Tensor, attention_mask: torch.Tensor, key_value_states: Optional[torch.Tensor] = None, inference_context: Optional[megatron.core.inference.contexts.BaseInferenceContext] = None, rotary_pos_emb: Optional[Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]] = None, rotary_pos_cos: Optional[torch.Tensor] = None, rotary_pos_sin: Optional[torch.Tensor] = None, attention_bias: Optional[torch.Tensor] = None, packed_seq_params: Optional[megatron.core.packed_seq_params.PackedSeqParams] = None, sequence_len_offset: Optional[int] = None, *, inference_params: Optional[megatron.core.inference.contexts.BaseInferenceContext] = None) → Tuple[torch.Tensor, torch.Tensor]

Perform a forward pass through the attention module.

Parameters

• hidden_states (Tensor) – Hidden states.

• attention_mask (Tensor) – Attention mask.

• key_value_states (Optional[Tensor]) – Key/value states (for cross attention).

• inference_context (Optional[BaseInferenceContext]) – Inference context that manages KV cache.

• rotary_pos_emb (Optional[Union[Tensor, Tuple[Tensor, Tensor]]]) – Rotary embedding tensor(s).

• rotary_pos_cos (Optional[Tensor]) – Rotary embedding cosine.

• rotary_pos_sin (Optional[Tensor]) – Rotary embedding sine.

• attention_bias (Optional[Tensor]) – Attention bias.

• packed_seq_params (Optional[PackedSeqparams]) – Parameters used for THD format.

• sequence_len_offset (Optional[int]) – Sequence length offset used for inference CUDA graphs.

Returns

(Tuple[Tensor, Tensor]) Attention output and bias.

abstract get_query_key_value_tensors(hidden_states, key_value_states)

This method needs to be implemented based on whether the derived class is “self-attn” or “cross-attn”.

set_for_recompute_input_layernorm()

Set the attention layer for recompute input_layernorm. Only needed for fp8.

class core.transformer.attention.CrossAttention(*args: Any, **kwargs: Any)

Bases: core.transformer.attention.Attention

Cross-attention layer class

Cross-attention layer takes input with size [s, b, h] and context with size [s, b, h] and returns output of the same size.

get_query_key_value_tensors(hidden_states, key_value_states)

Derives query tensor from hidden_states, and key/value tensors from key_value_states.

class core.transformer.attention.CrossAttentionSubmodules(linear_q: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, linear_kv: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, core_attention: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, linear_proj: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None)

Bases: object

Configuration class for specifying the submodules of a cross-attention.

core_attention: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

linear_kv: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

linear_proj: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

linear_q: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

class core.transformer.attention.SelfAttention(*args: Any, **kwargs: Any)

Bases: core.transformer.attention.Attention

Self-attention layer class

Self-attention layer takes input with size [s, b, h] and returns output of the same size.

backward_dw() → NoReturn

Execute weight update operations

get_query_key_value_tensors(hidden_states, key_value_states=None)

Derives query, key and value tensors from hidden_states.

run_realtime_tests()

Performs a consistency check.

This function makes sure that tensors across devices are the same during an experiment. This is often not guaranteed to be so because of silent hardware failures (eg, memory corruption loading a checkpoint, network traffic corruption encountered during data transmission).

(TODO) In the future, more tensors should be checked across the training run and checked every X iterations. This is left for future work. Equality of tensors is probably not required; transmitting hashes is sufficient.

set_for_recompute_input_layernorm()

Set the attention layer for recompute input_layernorm. Only needed for fp8.

class core.transformer.attention.SelfAttentionSubmodules(linear_qkv: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, core_attention: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, linear_proj: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, q_layernorm: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, k_layernorm: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None)

Bases: object

Configuration class for specifying the submodules of a self-attention.

core_attention: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

k_layernorm: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

linear_proj: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

linear_qkv: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

q_layernorm: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

This is a PyTorch-only implementation of dot product attention. A more efficient implementation, like those provided by FlashAttention or CUDNN’s FusedAttention, are typically used when training speed is important.

class core.transformer.dot_product_attention.DotProductAttention(*args: Any, **kwargs: Any)

Bases: megatron.core.transformer.module.MegatronModule

Region where selective activation recomputation is applied. This region is memory intensive but less compute intensive which makes activation checkpointing more efficient for LLMs (20B+). See Reducing Activation Recomputation in Large Transformer Models: https://arxiv.org/abs/2205.05198 for more details.

We use the following notation:

h: hidden size n: number of attention heads p: number of tensor model parallel partitions b: batch size s: sequence length

forward(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, attention_mask: torch.Tensor, attn_mask_type: Optional[megatron.core.transformer.enums.AttnMaskType] = None, attention_bias: Optional[torch.Tensor] = None, packed_seq_params: Optional[megatron.core.packed_seq_params.PackedSeqParams] = None)

Forward.

sharded_state_dict(prefix: str = '', sharded_offsets: Tuple[Tuple[int, int, int]] = (), metadata: Optional[dict] = None) → megatron.core.dist_checkpointing.mapping.ShardedStateDict

Sharded state dict for the learnable softmax offset parameter

class core.transformer.enums.AttnBackend(value)

Bases: enum.Enum

Attention Backend

auto = 5

flash = 1

fused = 2

local = 4

unfused = 3

class core.transformer.enums.AttnMaskType(value)

Bases: enum.Enum

Attention Mask Type

arbitrary = 5

causal = 2

no_mask = 3

padding = 1

padding_causal = 4

class core.transformer.enums.AttnType(value)

Bases: enum.Enum

Attention type

cross_attn = 2

self_attn = 1

class core.transformer.enums.LayerType(value)

Bases: enum.Enum

Layer type embedding: embedding layer loss: loss layer encoder: encoder layer, not implemented yet, expect to be used in MLLM models decoder: decoder layer mtp: multi-token prediction layer, not implemented yet

decoder = 4

embedding = 1

encoder = 3

loss = 2

mtp = 5

class core.transformer.enums.ModelType(value)

Bases: enum.Enum

Model Type

encoder_or_decoder for bert, gpt etc encoder_and_decoder for multimodal , T5 etc

property encoder_and_decoder

Deprecated property - use encoder_or_decoder instead.

encoder_or_decoder = 1

This provides a pass-through module that can be used in specs to indicate that the operation should not be performed. For example, when using LayerNorm with the subsequent linear layer, an IdentityOp can be passed in as the LayerNorm module to use.

class core.transformer.identity_op.IdentityFuncOp(*args: Any, **kwargs: Any)

Bases: core.transformer.identity_op.IdentityOp

This is a placeholder for IdentityFuncOp(…)(x) -> IdentityOp(x) -> x. Such a func is handy for ops like bias_dropout_fusion which themselves return a function at runtime based on passed arguments

forward(*args, **kwargs)

class core.transformer.identity_op.IdentityOp(*args: Any, **kwargs: Any)

Bases: torch.nn.Module

This is a placeholder for IdentityOp(x) -> x

forward(x, *args, **kwargs)

This is the entire MLP portion of the transformer layer with an input projection, non-linearity, and output projection.

class core.transformer.mlp.MLP(*args: Any, **kwargs: Any)

Bases: megatron.core.transformer.module.MegatronModule

MLP will take the input with h hidden state, project it to 4*h hidden dimension, perform nonlinear transformation, and project the state back into h hidden dimension.

Returns an output and a bias to be added to the output. If config.add_bias_linear is False, the bias returned is None.

We use the following notation:

h: hidden size p: number of tensor model parallel partitions b: batch size s: sequence length

backward_dw()

forward(hidden_states, per_token_scale=None)

Perform the forward pass through the MLP block.

sharded_state_dict(prefix: str = '', sharded_offsets: tuple = (), metadata: Optional[dict] = None) → megatron.core.dist_checkpointing.mapping.ShardedStateDict

Return the sharded state dictionary of the module.

class core.transformer.mlp.MLPSubmodules(linear_fc1: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, activation_func: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None, linear_fc2: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, type]] = None)

Bases: object

The dataclass for ModuleSpecs of MLP submodules including linear fc1, activation function, linear fc2.

activation_func: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

linear_fc1: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

linear_fc2: Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = None

core.transformer.mlp.apply_swiglu_sharded_factory(original_sh_ten, sharded_offsets, singleton_local_shards: bool = False)

This provides a common base class for all modules used in the transformer that contains some common functionality.

Megatron Module.

class core.transformer.module.Float16Module(*args: Any, **kwargs: Any)

Bases: core.transformer.module.MegatronModule

Float 16 Module.

config

Transformer config

Type

TransformerConfig

fp16

Specifies if the model runs in fp16 mode

Type

bool

bf16

Specifies if the model runs in bf16 mode

Type

bool

Parameters

config (TransformerConfig) – The transformer config used to initalize the model

forward(*inputs, fp32_output=True, **kwargs)

Execute the wrapped module in model precision and optionally upcast outputs to fp32.

On the first pipeline stage, positional/keyword tensor inputs are converted to the module precision (fp16 or bf16) before invoking the wrapped module. The wrapped module is called with the provided inputs and keyword arguments. On the last pipeline stage only, outputs are upcast to fp32 if fp32_output is True; otherwise, outputs are returned in the model precision (fp16/bf16).

Parameters

• *inputs – Positional inputs forwarded to the wrapped module (converted to fp16/bf16 on the pipeline first stage).

• fp32_output (bool, keyword-only) – If True (default), upcast outputs to fp32 on the pipeline last stage. Has no effect on non-last stages. Set to False to keep outputs in model precision when downstream consumers expect half precision or to avoid extra casts.

• **kwargs – Keyword arguments forwarded to the wrapped module.

Returns

The wrapped module’s outputs, potentially upcast to fp32 depending on pipeline stage and fp32_output.

load_state_dict(state_dict, strict=True)

set_input_tensor(input_tensor)

sharded_state_dict(prefix='', *args, **kwargs)

Retrieve sharded_state_dict from the module being wrapped.

state_dict(destination=None, prefix='', keep_vars=False)

state_dict_for_save_checkpoint(prefix='', keep_vars=False)

Retrieve state_dict from the module being wrapped.

class core.transformer.module.MegatronModule(*args: Any, **kwargs: Any)

Bases: torch.nn.Module

Base Megatron module inhertied by all Models.

Megatron specific extensions of torch Module with support for pipelining

Parameters

config (TransformerConfig) – Transformer config

set_is_first_microbatch()

Sets the is_first_microbatch flag if it exists and config.fp8==True. When this flag is set, TE modules will update their fp8 parameter cache. If kitchen is being used, kitchen controls quantization level.

set_symmetric_ar(set_to: Optional[str] = None) → None

Set symmetric all-reduce functionality across all eligible modules.

This method traverses the model’s module hierarchy to find all modules with the ‘symmetric_ar_type’ attribute, caches them, and then sets their ‘_symmetric_ar_cache’ attribute to the specified value to enable or disable symmetric all-reduce operations.

Parameters

• set_to (Any, optional) – Value to set for the ‘symmetric_ar_type’ to.

• ['two_shot' (Allowed choices) –

• "one_shot" –

• "multimem_all_reduce" –

• None] –

sharded_state_dict(prefix: str = '', sharded_offsets: Tuple[Tuple[int, int, int]] = (), metadata: Optional[dict] = None) → megatron.core.dist_checkpointing.mapping.ShardedStateDict

Default implementation for sharded state dict for distributed checkpointing.

General definition of sharded_state_dict simply calls sharded_state_dict_default (which call sharded_state_dict method if possible or a default implementation otherwise) recursively on all submodules.

Parameters

• prefix (str) – prefix for the state dict keys

• sharded_offsets (Tuple[Tuple[int, int, int]], optional) – sharding already applied (e.g. PP related) by sup-modules. Passed along to ShardedTensor

• metadata (dict, optional) – metadata passed recursively to sharded_state_dict methods

Returns

dictionary of state dict keys mapped to ShardedTensors

Return type

dict

state_dict_for_save_checkpoint(prefix: str = '', keep_vars: bool = False)

Override state dict for saving checkpoints Use this function to override the state dict for saving checkpoints.

Parameters

• prefix (str, optional) – _description_. Defaults to ‘’.

• keep_vars (bool, optional) – _description_. Defaults to False.

Returns

_description_

Return type

_type_

core.transformer.module.conversion_helper(val, conversion)

Recursively applies a conversion function to values in nested data structures.

Parameters

• val – A single value or a nested structure (tuple/list) of values to convert

• conversion (callable) – A function that performs the desired conversion on a single value

Returns

The converted value, maintaining the same nested structure as the input. If input is a single value, returns the converted value. If input is a tuple/list, returns a tuple/list with all elements converted.

core.transformer.module.float16_to_fp32(val)

Converts floating-point values from fp16 to fp32.

Parameters

val – The value to convert. Can be a single number, a tuple, or a list.

core.transformer.module.fp32_to_float16(val, float16_convertor)

Converts floating-point values from fp32 to fp16.

Parameters

• val – The value to convert. Can be a single number, a tuple, or a list.

• float16_convertor – A function that converts a single fp32 value to fp16

core.transformer.module.param_is_not_shared(param)

A block, or stack, of several transformer layers. The layers can all be the same or each can be unique.

class core.transformer.transformer_block.TransformerBlock(*args: Any, **kwargs: Any)

Bases: megatron.core.transformer.module.MegatronModule

Transformer class.

forward(hidden_states: Union[torch.Tensor, megatron.core.utils.WrappedTensor], attention_mask: Optional[torch.Tensor], context: Optional[torch.Tensor] = None, context_mask: Optional[torch.Tensor] = None, rotary_pos_emb: Optional[torch.Tensor] = None, rotary_pos_cos: Optional[torch.Tensor] = None, rotary_pos_sin: Optional[torch.Tensor] = None, attention_bias: Optional[torch.Tensor] = None, inference_context: Optional[megatron.core.inference.contexts.BaseInferenceContext] = None, packed_seq_params: Optional[megatron.core.packed_seq_params.PackedSeqParams] = None, sequence_len_offset: Optional[torch.Tensor] = None, *, inference_params: Optional[megatron.core.inference.contexts.BaseInferenceContext] = None)

Perform the forward pass through the transformer block.

This method handles the core computation of the transformer, including self-attention, optional cross-attention, and feed-forward operations.

Parameters

• hidden_states (Union[Tensor, WrappedTensor]) – Input tensor of shape [s, b, h] where s is the sequence length, b is the batch size, and h is the hidden size. Can be passed as a WrappedTensor during inference to avoid an obsolete reference in the calling function.

• attention_mask (Tensor) – Boolean tensor of shape [1, 1, s, s] for masking self-attention.

• context (Tensor, optional) – Context tensor for cross-attention.

• context_mask (Tensor, optional) – Mask for cross-attention context

• rotary_pos_emb (Tensor, optional) – Rotary positional embeddings.

• attention_bias (Tensor) – Bias tensor for Q * K.T of shape in shape broadcastable to [b, num_head, sq, skv], e.g. [1, 1, sq, skv]. Used as an alternative to apply attention mask for TE cuDNN attention.

• inference_context (BaseInferenceContext, optional) – Parameters for inference-time optimizations.

• packed_seq_params (PackedSeqParams, optional) – Parameters for packed sequence processing.

Returns

The output hidden states tensor of shape [s, b, h], and optionally the updated context tensor if cross-attention is used.

Return type

Union[Tensor, Tuple[Tensor, Tensor]]

set_input_tensor(input_tensor: torch.Tensor)

Set input tensor to be used instead of forward()’s input.

When doing pipeline parallelism the input from the previous stage comes from communication, not from the input, so the model’s forward_step_func won’t have it. This function is thus used by internal code to bypass the input provided by the forward_step_func

sharded_state_dict(prefix: str = '', sharded_offsets: tuple = (), metadata: Optional[dict] = None) → megatron.core.dist_checkpointing.mapping.ShardedStateDict

Generate a sharded state dictionary for the transformer block.

Parameters

• prefix (str, optional) – Prefix to be added to all keys in the state dict. Defaults to an empty string.

• sharded_offsets (tuple, optional) – Tuple of sharding offsets.

• metadata (dict, optional) – Additional metadata for sharding. Can specify if layers are non-homogeneous. Defaults to None.

Returns

A dictionary containing the sharded state of the model.

Return type

ShardedStateDict

class core.transformer.transformer_block.TransformerBlockSubmodules(layer_specs: Optional[List[megatron.core.transformer.spec_utils.ModuleSpec]] = None, layer_norm: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, torch.nn.Module]] = None)

Bases: object

Dataclass for specifying the submodules of a transformer block.

This class defines the structure for configuring the layers and normalization within a transformer block, allowing for flexible and customizable architecture designs.

Parameters

• layer_specs (List[ModuleSpec], optional) – A list of module specifications for the layers within the transformer block. Each specification typically defines a complete transformer layer (e.g., self-attention, feed-forward network).

• layer_norm (Optional[Union[ModuleSpec, torch.nn.Module]], optional) – Specification or instance of the layer normalization to be applied.

layer_norm: Optional[Union[megatron.core.transformer.spec_utils.ModuleSpec, torch.nn.Module]] = None

layer_specs: List[megatron.core.transformer.spec_utils.ModuleSpec] = None

core.transformer.transformer_block.get_num_layers_to_build(config: megatron.core.transformer.transformer_config.TransformerConfig, vp_stage: Optional[int] = None, pp_rank: Optional[int] = None) → int

Determine the number of transformer layers to build for the current pipeline stage. :param config: Configuration object containing transformer model parameters. :type config: TransformerConfig :param vp_stage: Virtual pipeline stage number. :type vp_stage: Optional[int] :param pp_rank: Pipeline parallel rank. :type pp_rank: Optional[int]

Returns

The number of layers to be built for the current pipeline stage.

Return type

int

This contains all of the configuration options for the transformer. Using a dataclass reduces code bloat by keeping all arguments together in a dataclass instead of passing several arguments through multiple layers of function calls.

class core.transformer.transformer_config.MLATransformerConfig(tensor_model_parallel_size: int = 1, pipeline_model_parallel_comm_backend: Optional[str] = None, pipeline_model_parallel_size: int = 1, virtual_pipeline_model_parallel_size: Optional[int] = None, sequence_parallel: bool = False, context_parallel_size: int = 1, hierarchical_context_parallel_sizes: Optional[list[int]] = None, expert_model_parallel_size: int = 1, expert_tensor_parallel_size: Optional[int] = None, moe_extended_tp: bool = False, perform_initialization: bool = True, use_cpu_initialization: bool = False, fp16: bool = False, bf16: bool = False, params_dtype: torch.dtype = torch.float32, timers: Optional[Callable] = None, finalize_model_grads_func: Optional[Callable] = None, grad_scale_func: Optional[Callable] = None, no_sync_func: Optional[Callable] = None, grad_sync_func: Optional[Callable] = None, param_sync_func: Optional[Callable] = None, deterministic_mode: bool = False, enable_autocast: bool = False, autocast_dtype: Optional[torch.dtype] = None, num_microbatches_with_partial_activation_checkpoints: Optional[int] = None, gradient_accumulation_fusion: bool = False, async_tensor_model_parallel_allreduce: bool = False, use_te_rng_tracker: bool = False, tp_comm_overlap: bool = False, tp_comm_bulk_wgrad: bool = True, tp_comm_bulk_dgrad: bool = True, tp_comm_overlap_ag: bool = True, tp_comm_overlap_rs: bool = True, tp_comm_overlap_rs_dgrad: bool = False, tp_comm_split_ag: bool = True, tp_comm_atomic_ag: bool = False, tp_comm_split_rs: bool = True, tp_comm_atomic_rs: bool = False, cross_entropy_loss_fusion: bool = False, cross_entropy_fusion_impl: str = 'native', tp_comm_overlap_disable_qkv: bool = False, tp_comm_overlap_disable_fc1: bool = False, tp_comm_bootstrap_backend: str = 'nccl', overlap_moe_expert_parallel_comm: bool = False, delay_wgrad_compute: bool = False, pipeline_dtype: Optional[torch.dtype] = None, variable_seq_lengths: bool = False, overlap_p2p_comm: bool = False, batch_p2p_comm: bool = True, batch_p2p_sync: bool = True, use_ring_exchange_p2p: bool = False, deallocate_pipeline_outputs: bool = False, defer_embedding_wgrad_compute: bool = False, wgrad_deferral_limit: int = 0, overlap_p2p_comm_warmup_flush: bool = False, microbatch_group_size_per_vp_stage: Optional[int] = None, cpu_offloading: bool = False, cpu_offloading_num_layers: int = 0, _cpu_offloading_context: Optional[ContextManager] = None, cpu_offloading_activations: bool = True, cpu_offloading_weights: bool = False, cpu_offloading_double_buffering: bool = False, barrier_with_L1_time: bool = True, num_layers: int = 0, mtp_num_layers: Optional[int] = None, mtp_loss_scaling_factor: Optional[float] = None, num_layers_in_first_pipeline_stage: Optional[int] = None, num_layers_in_last_pipeline_stage: Optional[int] = None, pipeline_model_parallel_layout: Optional[Union[str, list, megatron.core.transformer.pipeline_parallel_layer_layout.PipelineParallelLayerLayout]] = None, account_for_embedding_in_pipeline_split: bool = False, account_for_loss_in_pipeline_split: bool = False, hidden_size: int = 0, num_attention_heads: int = 0, attention_backend: megatron.core.transformer.enums.AttnBackend = megatron.core.transformer.enums.AttnBackend.auto, softmax_scale: Optional[float] = None, softmax_type: Literal['vanilla', 'off-by-one', 'learnable'] = 'vanilla', num_query_groups: Optional[int] = None, ffn_hidden_size: Optional[int] = None, kv_channels: Optional[int] = None, hidden_dropout: float = 0.1, attention_dropout: float = 0.1, fp32_residual_connection: bool = False, apply_residual_connection_post_layernorm: bool = False, layernorm_epsilon: float = 1e-05, layernorm_zero_centered_gamma: bool = False, add_bias_linear: bool = True, add_qkv_bias: bool = False, gated_linear_unit: bool = False, activation_func: Callable = torch.nn.functional.gelu, activation_func_fp8_input_store: bool = False, glu_linear_offset: float = 0.0, activation_func_clamp_value: Optional[float] = None, num_moe_experts: Optional[int] = None, rotary_interleaved: bool = False, window_size: Optional[Tuple[int, int]] = None, normalization: str = 'RMSNorm', qk_layernorm: bool = False, test_mode: bool = False, calculate_per_token_loss: bool = False, multi_latent_attention: bool = True, no_rope_freq: Optional[Union[int, List[int]]] = None, moe_deepep_num_sms: int = 20, init_method: Optional[Callable] = None, output_layer_init_method: Optional[Callable] = None, init_method_std: float = 0.02, embedding_init_method: Optional[Callable] = None, embedding_init_method_std: Optional[float] = None, init_model_with_meta_device: bool = False, apply_query_key_layer_scaling: bool = False, attention_softmax_in_fp32: bool = True, disable_bf16_reduced_precision_matmul: bool = False, bias_activation_fusion: bool = False, masked_softmax_fusion: bool = False, persist_layer_norm: bool = False, memory_efficient_layer_norm: bool = False, bias_dropout_fusion: bool = False, apply_rope_fusion: bool = False, use_fused_weighted_squared_relu: bool = False, recompute_granularity: Optional[str] = None, recompute_method: Optional[str] = None, recompute_num_layers: Optional[int] = None, distribute_saved_activations: Optional[bool] = None, recompute_modules: Optional[List[str]] = None, fp8: Optional[str] = None, fp8_recipe: Optional[str] = 'delayed', fp8_param: bool = False, fp8_margin: int = 0, fp8_interval: int = 1, fp8_amax_history_len: int = 1, fp8_amax_compute_algo: str = 'most_recent', fp8_wgrad: bool = True, fp8_dot_product_attention: bool = False, fp8_multi_head_attention: bool = False, tp_only_amax_red: bool = False, first_last_layers_bf16: bool = False, num_layers_at_start_in_bf16: int = 1, num_layers_at_end_in_bf16: int = 1, use_kitchen: bool = False, fp4: Optional[str] = None, fp4_recipe: Optional[str] = 'nvfp4', fp4_param: bool = False, moe_shared_expert_intermediate_size: Optional[int] = None, moe_shared_expert_overlap: bool = False, moe_layer_freq: Union[int, List[int]] = 1, moe_ffn_hidden_size: Optional[int] = None, moe_router_load_balancing_type: Union[str, List[str]] = 'aux_loss', moe_router_topk: int = 2, moe_router_topk_limited_devices: Optional[int] = None, moe_router_padding_for_fp8: Optional[bool] = False, moe_router_num_groups: Optional[int] = None, moe_router_group_topk: Optional[int] = None, moe_router_pre_softmax: bool = False, moe_router_topk_scaling_factor: Optional[float] = None, moe_router_score_function: str = 'softmax', moe_router_dtype: Optional[str] = None, moe_router_enable_expert_bias: bool = False, moe_router_bias_update_rate: float = 0.001, moe_router_force_load_balancing: bool = False, moe_grouped_gemm: bool = False, moe_use_legacy_grouped_gemm: bool = False, moe_aux_loss_coeff: Union[float, List[float]] = 0.0, moe_z_loss_coeff: Optional[float] = None, moe_input_jitter_eps: Optional[float] = None, moe_token_dropping: bool = False, moe_token_dispatcher_type: str = 'allgather', moe_enable_deepep: bool = False, moe_per_layer_logging: bool = False, moe_expert_capacity_factor: Optional[float] = None, moe_pad_expert_input_to_capacity: bool = False, moe_token_drop_policy: str = 'probs', moe_layer_recompute: bool = False, moe_permute_fusion: bool = False, moe_router_fusion: bool = False, moe_apply_probs_on_input: bool = False, cp_comm_type: Optional[Union[str, List[str]]] = None, enable_cuda_graph: bool = False, cuda_graph_use_single_mempool: bool = False, cuda_graph_retain_backward_graph: bool = False, cuda_graph_warmup_steps: int = 3, external_cuda_graph: bool = False, cuda_graph_scope: str = 'full', clone_scatter_output_in_embedding: bool = True, disable_parameter_transpose_cache: bool = False, config_logger_dir: str = '', flash_decode: bool = False, use_te_activation_func: bool = False, inference_rng_tracker: bool = False, inference_sampling_seed: int = 42, symmetric_ar_type: Optional[str] = None, mrope_section: Optional[List[int]] = None, is_hybrid_model: bool = False, mamba_state_dim: int = 128, mamba_head_dim: int = 64, mamba_num_groups: int = 8, mamba_num_heads: Optional[int] = None, use_mamba_mem_eff_path: bool = True, mlp_chunks_for_prefill: int = 1, heterogeneous_block_specs: bool = False, hetereogenous_dist_checkpoint: bool = False, quant_recipe: Optional[megatron.core.quantization.quant_config.RecipeConfig] = None, transformer_impl: str = 'transformer_engine', q_lora_rank: int = 512, kv_lora_rank: int = 512, qk_head_dim: int = 128, qk_pos_emb_head_dim: int = 64, v_head_dim: int = 128, rope_type: str = 'yarn', rotary_base: float = 10000, rotary_percent: float = 1.0, rotary_scaling_factor: float = 40, max_position_embeddings: int = 4096, original_max_position_embeddings: int = 4096, beta_fast: float = 32, beta_slow: float = 1, mscale: float = 1.0, mscale_all_dim: float = 0.0, cache_mla_latents: bool = False)

Bases: core.transformer.transformer_config.TransformerConfig

Configuration object for megatron-core Multi-Latent Attention (MLA) transformers.

The initialization function has an argument for each parameter, including those in ModelParallelConfig. Included YaRN RoPE parameters that is fused in MLA.

beta_fast: float = 32

Beta fast for YaRN RoPE, used by yarn.

beta_slow: float = 1

Beta slow for YaRN RoPE, used by yarn.

cache_mla_latents: bool = False

Cache the low dimensional tensors for MLA rather than full KV cache. This is only for the dynamic inference backend and requires that Flash MLA is installed.

kv_lora_rank: int = 512

Rank of Key and Value tensors’ low rank representation.

max_position_embeddings: int = 4096

This arg is not used, will be deprecated.

mscale: float = 1.0

Mscale for YaRN RoPE in Multi-Latent Attention, used by yarn.

mscale_all_dim: float = 0.0

Mscale all dimensions for YaRN RoPE in Multi-Latent Attention, used by yarn.

multi_latent_attention: bool = True

Whether to use Multi-Latent Attention.

normalization: str = 'RMSNorm'

Default normalization layer for MLA models is RMSNorm.

original_max_position_embeddings: int = 4096

Original maximum position embeddings for the original model, used by yarn.

q_lora_rank: int = 512

Rank of Query tensor’s low rank representation.

qk_head_dim: int = 128

Dimension of the head in the QK projection. q_head_dim = qk_head_dim + qk_pos_emb_head_dim

qk_pos_emb_head_dim: int = 64

Dimension of the position embedding in the QK projection.

rope_type: str = 'yarn'

Type of RoPE to use. Default to yarn, options are rope and yarn.

rotary_base: float = 10000

Rotary base for the rotary embeddings, used by rope and yarn.

rotary_percent: float = 1.0

Rotary percent for the rotary embeddings, used by rope.

rotary_scaling_factor: float = 40

Rotary scaling factor for the rotary embeddings, used by yarn.

v_head_dim: int = 128

Dimension of the head in the V projection.

class core.transformer.transformer_config.TransformerConfig(tensor_model_parallel_size: int = 1, pipeline_model_parallel_comm_backend: Optional[str] = None, pipeline_model_parallel_size: int = 1, virtual_pipeline_model_parallel_size: Optional[int] = None, sequence_parallel: bool = False, context_parallel_size: int = 1, hierarchical_context_parallel_sizes: Optional[list[int]] = None, expert_model_parallel_size: int = 1, expert_tensor_parallel_size: Optional[int] = None, moe_extended_tp: bool = False, perform_initialization: bool = True, use_cpu_initialization: bool = False, fp16: bool = False, bf16: bool = False, params_dtype: torch.dtype = torch.float32, timers: Optional[Callable] = None, finalize_model_grads_func: Optional[Callable] = None, grad_scale_func: Optional[Callable] = None, no_sync_func: Optional[Callable] = None, grad_sync_func: Optional[Callable] = None, param_sync_func: Optional[Callable] = None, deterministic_mode: bool = False, enable_autocast: bool = False, autocast_dtype: Optional[torch.dtype] = None, num_microbatches_with_partial_activation_checkpoints: Optional[int] = None, gradient_accumulation_fusion: bool = False, async_tensor_model_parallel_allreduce: bool = False, use_te_rng_tracker: bool = False, tp_comm_overlap: bool = False, tp_comm_bulk_wgrad: bool = True, tp_comm_bulk_dgrad: bool = True, tp_comm_overlap_ag: bool = True, tp_comm_overlap_rs: bool = True, tp_comm_overlap_rs_dgrad: bool = False, tp_comm_split_ag: bool = True, tp_comm_atomic_ag: bool = False, tp_comm_split_rs: bool = True, tp_comm_atomic_rs: bool = False, cross_entropy_loss_fusion: bool = False, cross_entropy_fusion_impl: str = 'native', tp_comm_overlap_disable_qkv: bool = False, tp_comm_overlap_disable_fc1: bool = False, tp_comm_bootstrap_backend: str = 'nccl', overlap_moe_expert_parallel_comm: bool = False, delay_wgrad_compute: bool = False, pipeline_dtype: Optional[torch.dtype] = None, variable_seq_lengths: bool = False, overlap_p2p_comm: bool = False, batch_p2p_comm: bool = True, batch_p2p_sync: bool = True, use_ring_exchange_p2p: bool = False, deallocate_pipeline_outputs: bool = False, defer_embedding_wgrad_compute: bool = False, wgrad_deferral_limit: int = 0, overlap_p2p_comm_warmup_flush: bool = False, microbatch_group_size_per_vp_stage: Optional[int] = None, cpu_offloading: bool = False, cpu_offloading_num_layers: int = 0, _cpu_offloading_context: Optional[ContextManager] = None, cpu_offloading_activations: bool = True, cpu_offloading_weights: bool = False, cpu_offloading_double_buffering: bool = False, barrier_with_L1_time: bool = True, num_layers: int = 0, mtp_num_layers: Optional[int] = None, mtp_loss_scaling_factor: Optional[float] = None, num_layers_in_first_pipeline_stage: Optional[int] = None, num_layers_in_last_pipeline_stage: Optional[int] = None, pipeline_model_parallel_layout: Optional[Union[str, list, megatron.core.transformer.pipeline_parallel_layer_layout.PipelineParallelLayerLayout]] = None, account_for_embedding_in_pipeline_split: bool = False, account_for_loss_in_pipeline_split: bool = False, hidden_size: int = 0, num_attention_heads: int = 0, attention_backend: megatron.core.transformer.enums.AttnBackend = megatron.core.transformer.enums.AttnBackend.auto, softmax_scale: Optional[float] = None, softmax_type: Literal['vanilla', 'off-by-one', 'learnable'] = 'vanilla', num_query_groups: Optional[int] = None, ffn_hidden_size: Optional[int] = None, kv_channels: Optional[int] = None, hidden_dropout: float = 0.1, attention_dropout: float = 0.1, fp32_residual_connection: bool = False, apply_residual_connection_post_layernorm: bool = False, layernorm_epsilon: float = 1e-05, layernorm_zero_centered_gamma: bool = False, add_bias_linear: bool = True, add_qkv_bias: bool = False, gated_linear_unit: bool = False, activation_func: Callable = torch.nn.functional.gelu, activation_func_fp8_input_store: bool = False, glu_linear_offset: float = 0.0, activation_func_clamp_value: Optional[float] = None, num_moe_experts: Optional[int] = None, rotary_interleaved: bool = False, window_size: Optional[Tuple[int, int]] = None, normalization: str = 'LayerNorm', qk_layernorm: bool = False, test_mode: bool = False, calculate_per_token_loss: bool = False, multi_latent_attention: bool = False, no_rope_freq: Optional[Union[int, List[int]]] = None, moe_deepep_num_sms: int = 20, init_method: Optional[Callable] = None, output_layer_init_method: Optional[Callable] = None, init_method_std: float = 0.02, embedding_init_method: Optional[Callable] = None, embedding_init_method_std: Optional[float] = None, init_model_with_meta_device: bool = False, apply_query_key_layer_scaling: bool = False, attention_softmax_in_fp32: bool = True, disable_bf16_reduced_precision_matmul: bool = False, bias_activation_fusion: bool = False, masked_softmax_fusion: bool = False, persist_layer_norm: bool = False, memory_efficient_layer_norm: bool = False, bias_dropout_fusion: bool = False, apply_rope_fusion: bool = False, use_fused_weighted_squared_relu: bool = False, recompute_granularity: Optional[str] = None, recompute_method: Optional[str] = None, recompute_num_layers: Optional[int] = None, distribute_saved_activations: Optional[bool] = None, recompute_modules: Optional[List[str]] = None, fp8: Optional[str] = None, fp8_recipe: Optional[str] = 'delayed', fp8_param: bool = False, fp8_margin: int = 0, fp8_interval: int = 1, fp8_amax_history_len: int = 1, fp8_amax_compute_algo: str = 'most_recent', fp8_wgrad: bool = True, fp8_dot_product_attention: bool = False, fp8_multi_head_attention: bool = False, tp_only_amax_red: bool = False, first_last_layers_bf16: bool = False, num_layers_at_start_in_bf16: int = 1, num_layers_at_end_in_bf16: int = 1, use_kitchen: bool = False, fp4: Optional[str] = None, fp4_recipe: Optional[str] = 'nvfp4', fp4_param: bool = False, moe_shared_expert_intermediate_size: Optional[int] = None, moe_shared_expert_overlap: bool = False, moe_layer_freq: Union[int, List[int]] = 1, moe_ffn_hidden_size: Optional[int] = None, moe_router_load_balancing_type: Union[str, List[str]] = 'aux_loss', moe_router_topk: int = 2, moe_router_topk_limited_devices: Optional[int] = None, moe_router_padding_for_fp8: Optional[bool] = False, moe_router_num_groups: Optional[int] = None, moe_router_group_topk: Optional[int] = None, moe_router_pre_softmax: bool = False, moe_router_topk_scaling_factor: Optional[float] = None, moe_router_score_function: str = 'softmax', moe_router_dtype: Optional[str] = None, moe_router_enable_expert_bias: bool = False, moe_router_bias_update_rate: float = 0.001, moe_router_force_load_balancing: bool = False, moe_grouped_gemm: bool = False, moe_use_legacy_grouped_gemm: bool = False, moe_aux_loss_coeff: Union[float, List[float]] = 0.0, moe_z_loss_coeff: Optional[float] = None, moe_input_jitter_eps: Optional[float] = None, moe_token_dropping: bool = False, moe_token_dispatcher_type: str = 'allgather', moe_enable_deepep: bool = False, moe_per_layer_logging: bool = False, moe_expert_capacity_factor: Optional[float] = None, moe_pad_expert_input_to_capacity: bool = False, moe_token_drop_policy: str = 'probs', moe_layer_recompute: bool = False, moe_permute_fusion: bool = False, moe_router_fusion: bool = False, moe_apply_probs_on_input: bool = False, cp_comm_type: Optional[Union[str, List[str]]] = None, enable_cuda_graph: bool = False, cuda_graph_use_single_mempool: bool = False, cuda_graph_retain_backward_graph: bool = False, cuda_graph_warmup_steps: int = 3, external_cuda_graph: bool = False, cuda_graph_scope: str = 'full', clone_scatter_output_in_embedding: bool = True, disable_parameter_transpose_cache: bool = False, config_logger_dir: str = '', flash_decode: bool = False, use_te_activation_func: bool = False, inference_rng_tracker: bool = False, inference_sampling_seed: int = 42, symmetric_ar_type: Optional[str] = None, mrope_section: Optional[List[int]] = None, is_hybrid_model: bool = False, mamba_state_dim: int = 128, mamba_head_dim: int = 64, mamba_num_groups: int = 8, mamba_num_heads: Optional[int] = None, use_mamba_mem_eff_path: bool = True, mlp_chunks_for_prefill: int = 1, heterogeneous_block_specs: bool = False, hetereogenous_dist_checkpoint: bool = False, quant_recipe: Optional[megatron.core.quantization.quant_config.RecipeConfig] = None, transformer_impl: str = 'transformer_engine')

Bases: core.model_parallel_config.ModelParallelConfig

Configuration object for megatron-core transformers.

The initialization function has an argument for each parameter, including those in ModelParallelConfig.

account_for_embedding_in_pipeline_split: bool = False

If set, the embedding layer will be treated as a standard transformer layer in the context of partition and placement for pipeline parallelism.

account_for_loss_in_pipeline_split: bool = False

If set, the loss layer will be treated as a standard transformer layer in the context of partition and placement for pipeline parallelism.

activation_func: Callable

Activation function to use for the non-linearity in the MLP.

activation_func_clamp_value: Optional[float] = None

Clamp the output of the linear_fc1 in the activation function. Only used when activation_func is quick_gelu.

activation_func_fp8_input_store: bool = False

Store the input of MLP activation function in FP8 for backprop to save memory. The stored input is casted back to the original precision before backprop compuatation.

add_bias_linear: bool = True

Include a bias term in all linear layers (QKV projections, after core attention, and two in MLP layer).

add_qkv_bias: bool = False

Add a bias term only for QKV projections.

apply_query_key_layer_scaling: bool = False

If true, scale Q * K^T by 1 / layer-number. This improve numeric stability when training with fp16.

apply_residual_connection_post_layernorm: bool = False

If True, uses the original BERT residule connection ordering.

apply_rope_fusion: bool = False

If True, use fused RoPE kernel.

attention_backend: megatron.core.transformer.enums.AttnBackend

Attention backend to run. By default we let transformer engine decide the best backend to run (except in the case of local). If attention backend is local we use the local pytorch implementation in mcore. Users can specify exact backend by changing this config.

attention_dropout: float = 0.1

Post attention dropout probability.

attention_softmax_in_fp32: bool = True

If True, run attention masking and softmax in fp32. This should be True if apply_query_key_layer_scaling is True.

bias_activation_fusion: bool = False

If True, fuses bias addition and the activation function when possible.

bias_dropout_fusion: bool = False

If True, uses bias dropout fusion.

calculate_per_token_loss: bool = False

Whether cross entropy loss is calculated over the actual number of non-padded tokens in the global batch, versus the default behavior of assuming all tokens are non-padded.

clone_scatter_output_in_embedding: bool = True

When set to True, clone the output of scatter_to_sequence_parallel_region in embedding layer to facilitate garbage collection of input.

config_logger_dir: str = ''

When non-empty, dumps entry-point configs to config_logger_dir

cp_comm_type: Optional[Union[str, List[str]]] = None

Inter-gpu communication type for context parallelism. str: all layers share same communication type. List[str]: each layer has its separate communication type. cp_comm_type of each layer can be “p2p” or “all_gather” or “a2a” or “a2a+p2p”. “p2p”: Exchange KV chunks with P2P communications in ring topology. P2P is async and can be overlapped with attention compute. “all_gather”: All-gather to get full sequence of KV before attention. The all-gather is not async, and cannot be overlapped. “a2a”: Like DeepSpeed Ulysses, scatter attention heads across the CP group, and gather to get full sequence of QKV. “a2a+p2p”: A hierarchical implementation of context parallelism to attention. It uses A2A communications in low-level CP groups (e.g., via NVLink), and P2P communications in high-level CP groups (e.g., via IBLink).

cuda_graph_retain_backward_graph: bool = False

When set to true, cudagraph backward passes will be graph captured with ‘retain_grad=True’ This may enable cudagraphs for certain modules that are not completely cudagraph safe. For more details, see: https://pytorch.org/docs/stable/generated/torch.Tensor.backward.html.

cuda_graph_scope: str = 'full'

Determines the CUDA graphs capturing scope. When external_cuda_graph is set to true, valid values are “full” and “attn”. “Full” scope captures a whole Transformer layer. “Attn” scope only captures operations in TransformerLayer._forward_attention(). When enable_cuda_graph is set to true, “full_iteration” can be specified as cuda_graph_scope to enable whole iteration CUDA graph. All other values enable layerwise CUDA graph.

cuda_graph_use_single_mempool: bool = False

When set to true, cudagraphs will be captured inside a single mempool, in which all cudagraphs may only be used once per step. If false, cudagraphs may be reused across microbatches. Enabling may reduce cudagraph memory overheads due to memory fragmentation, however may greatly increase the number of cudagraphs created when the number of microbatches is high.

cuda_graph_warmup_steps: int = 3

Number of warmup steps for CUDA graphs

disable_bf16_reduced_precision_matmul: bool = False

If True, sets torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction=False to prevent matmul from using reduced precision accumulation when using BF16.

disable_parameter_transpose_cache: bool = False

When set to true, the parameter transposes are not cached for subsequent iterations.

distribute_saved_activations: Optional[bool] = None

If True, distribute recomputed activations across the model parallel group.

embedding_init_method: Optional[Callable] = None

Method to initialize weights of the embedding layer. If None, will be set as described in init_method above.

embedding_init_method_std: Optional[float] = None

Standard deviation of the zero mean normal for the default initialization method for the embedding layer. If None, will be set to init_method_std.

enable_cuda_graph: bool = False

When set to true, either partial CUDA graph (1/many CUDA graph per layer) or full iteration CUDA graph (1 CUDA graph for whole iteration excluding optimizer) is enabled. –cuda-graph-scope determines the scope of graph capture.

external_cuda_graph: bool = False

When set to true, TransformerLayer layers are swapped with user provided CUDA graphs.

ffn_hidden_size: Optional[int] = None

Transformer Feed-Forward Network hidden size. This is set to 4*hidden_size if not provided.

first_last_layers_bf16: bool = False

If True, retains first and last N TransformerBlocks in BF16 as opposed to FP8.

flash_decode: bool = False

Use the optimized flash decoding kernel during inference.

fp32_residual_connection: bool = False

If true, move residual connections to fp32.

fp4: Optional[str] = None

If set, enables the use of FP4 precision through Transformer Engine. Currently only supports ‘nvfp4’ which uses NVFP4BlockScaling recipe (requires TE >= 2.7.0.dev0).

fp4_param: bool = False

If set, keep the parameters in fp4 precision to save memory. This option must be used together with fp4 mode (i.e., TransformerConfig.fp4 is not None). Note that not all parameters will be converted to fp4; for example, biases will remain unchanged.

fp4_recipe: Optional[str] = 'nvfp4'

If set, enables the use of FP4 precision through Transformer Engine. Currently only ‘nvfp4’ is supported which uses NVFP4BlockScaling recipe for Blackwell+ architecture.

fp8: Optional[str] = None

If set, enables the use of FP8 precision through Transformer Engine. There are 2 predefined choices (1) ‘e4m3’ uniformly uses e4m3 for all FP8 tensors, (2) ‘hybrid’ uses e4m3 for all FP8 activation and weight tensors and e5m2 for all FP8 output activation gradient tensors.

fp8_amax_compute_algo: str = 'most_recent'

Algorithm used for choosing the amax value for the scaling factor computation. There are 2 predefined choices: max chooses the largest amax in the history window, while most_recent always chooses the most recently seen value.

fp8_amax_history_len: int = 1

The length of the amax history window used for scaling factor computation.

fp8_dot_product_attention: bool = False

When set to True, use the FP8 implementation of Dot Product Attention.

fp8_interval: int = 1

DEPRECATED from TransformerEngine v1.8.0. This flag is ignored. Controls how often the scaling factor is recomputed.

fp8_margin: int = 0

Margin for the scaling factor computation.

fp8_multi_head_attention: bool = False

When set to True, use the FP8 implementation of Multi Head Attention.

fp8_param: bool = False

If set, keep the parameters in fp8 precision to save memory. This option must be used together with fp8 mode (i.e., TransformerConfig.fp8 is not None). Note that not all parameters will be converted to fp8; for example, biases will remain unchanged. The parameters affected are primarily the weights of GEMMs. The specific parameters that will be converted to fp8 are determined by TE.

fp8_recipe: Optional[str] = 'delayed'

If set, enables the use of FP8 precision through Transformer Engine. There are 3 predefined choices (1) ‘tensorwise’ uses per tensor current scaling recipe, (2) ‘delayed’ uses delayed scaling recipe, 3) ‘mxfp8’ for Blackwell architecture only, 4) ‘blockwise’ for blockwise scaling recipe.

fp8_wgrad: bool = True

When set to False, override FP8 config options and do the wgrad computation in higher precision.

gated_linear_unit: bool = False

Use a gated linear unit for the first linear layer in the MLP.

glu_linear_offset: float = 0.0

activation_func(x[0]) * (x[1] + offset). Only used when gated_linear_unit is True

Type

Offset term in the GLU activation function

hetereogenous_dist_checkpoint: bool = False

Whether to use heterogenous layers in distributed checkpoint.

heterogeneous_block_specs: bool = False

Whether to use heterogeneous block specs (nemotron-nas architecture).

hidden_dropout: float = 0.1

Dropout probability for transformer hidden state.

hidden_size: int = 0

Transformer hidden size.

inference_rng_tracker: bool = False

Whether we should instantiate a separate RNG tracker for inference.

inference_sampling_seed: int = 42

Random seed to use for sampling during inference.

init_method: Optional[Callable] = None

Method to initialize weights. Note that bias is always set to zero. Should be a function that takes a single Tensor and initializes it. If None, will be set to megatron.core.utils.init_method_normal(init_method_std) which is torch nn init normal with mean=0.0 and std=init_method_std.

init_method_std: float = 0.02

Standard deviation of the zero mean normal for the default initialization method, not used if init_method and output_layer_init_method are provided.

init_model_with_meta_device: bool = False

If True, initializes the model with the meta device. This is helpful for training of very large models. This feature is only works when megatron fsdp is turned on.

is_hybrid_model: bool = False

Indicates whether this is a hybrid model.

kv_channels: Optional[int] = None

Projection weights dimension in multi-head attention. This is set to hidden_size // num_attention_heads if not provided.

layernorm_epsilon: float = 1e-05

Epsilon value for any LayerNorm operations.

layernorm_zero_centered_gamma: bool = False

If set to True, the LayerNorm is adjusted to center the gamma values around 0. This improves numerical stability.

mamba_head_dim: int = 64

The dimensionality of the heads in the Mamba layers.

mamba_num_groups: int = 8

The number of groups used in Mamba layers.

mamba_num_heads: Optional[int] = None

The number of heads used in Mamba layers. If None, the number of heads will be hidden_size * expand // mamba_head_dim.

mamba_state_dim: int = 128

The dimensionality of the state representation in Mamba layers.

masked_softmax_fusion: bool = False

If True, uses softmax fusion.

memory_efficient_layer_norm: bool = False

If True, and using local layers (not from TransformerEngine), tells Apex to use the memory efficient fused LayerNorm kernel. Ignored if not using LayerNorm.

mlp_chunks_for_prefill: int = 1

The number of chunks along the sequence dimension to use for MLP computation during prefill.

moe_apply_probs_on_input: bool = False

Apply probs on input of experts instead of applying after activation and glu.

moe_aux_loss_coeff: Union[float, List[float]] = 0.0

Scaling coefficient for the aux loss. A starting value of 1e-2 is recommended. If a list of load balancing types is provided for moe_router_load_balancing_type, a corresponding list of coefficients should be provided here.

moe_deepep_num_sms: int = 20

Number of SMs to use for DeepEP.

moe_enable_deepep: bool = False

[Experimental] Enable DeepEP for efficient token dispatching and combine in MoE models.

moe_expert_capacity_factor: Optional[float] = None

The capacity factor for each expert, None means no token will be dropped. The default is None.

Type

moe_expert_capacity_factor (float)

moe_ffn_hidden_size: Optional[int] = None

MoE Feed-Forward Network hidden size

moe_grouped_gemm: bool = False

When there are multiple experts per rank, compress multiple local (potentially small) gemms in a single kernel launch to improve the utilization and performance by leveraging the Grouped GEMM feature introduced since CUTLASS 2.8 (https://github.com/fanshiqing/grouped_gemm).

moe_input_jitter_eps: Optional[float] = None

Add noise to the input tensor by applying jitter with a specified epsilon value.

moe_layer_freq: Union[int, List[int]] = 1

Frequency between MoE layers and Dense layers. Accepts either: - An integer N: Represents a 1:N ratio, meaning one expert layer for every N-1 dense layers. - A list that defines a custom pattern, e.g.: [1,1,1,0,1,1,1,0,1,1,1,0]

moe_layer_recompute: bool = False

checkpointing moe_layer to save actiavtion memory.

Type

Memory optimization

moe_pad_expert_input_to_capacity: bool = False

If True, pads the input for each expert to match the expert capacity length, effective only after the moe_expert_capacity_factor is set. The default setting is False.

Type

moe_pad_expert_input_to_capacity (bool)

moe_per_layer_logging: bool = False

Enable per-layer logging for MoE, currently supports auxiliary loss and z loss.

moe_permute_fusion: bool = False

Fuse token rearrangement ops during token dispatching.

moe_router_bias_update_rate: float = 0.001

The expert bias is updated based on the number of assigned tokens to each expert in a global batch, where the bias is increased for the experts with less assigned tokens and decreased for the experts with more assigned tokens. The default value 1e-3 is same as that used in DeepSeekV3.

moe_router_dtype: Optional[str] = None

Data type for routing and expert output weighted averaging. Using fp32 or fp64 can improve stability especially when the number of experts is large (e.g. finegrained-moe). None means no changes for dtype.

moe_router_enable_expert_bias: bool = False

TopK routing with dynamic per-expert bias in the aux-loss-free load balancing strategy. The routing decision is based on the sum of the routing scores and the expert bias. See https://arxiv.org/abs/2408.15664 for details.

moe_router_force_load_balancing: bool = False

[Experimental] Force load balancing with random logits for MoE router, supports naive topk and group-limited topk. This is an experimental feature and only for benchmark.

moe_router_fusion: bool = False

Fuse ops in routing and aux loss calculation.

moe_router_group_topk: Optional[int] = None

Number of selected groups for group-limited routing.

moe_router_load_balancing_type: Union[str, List[str]] = 'aux_loss'

The load balancing strategy for the router. Options: - “aux_loss”: Load balancing loss used in GShard and SwitchTransformer, calculated at micro-batch level. - “seq_aux_loss”: Load balancing loss used in DeepSeekV2 and DeepSeekV3, computes loss for each individual sample. - “global_aux_loss”: Load balancing loss calculated at global batch level. - “sinkhorn”: Balancing algorithm used in S-BASE. - “none”: No load balancing. A list of strings can be provided to combine multiple aux-loss load balancing types. The default is “aux_loss”.

moe_router_num_groups: Optional[int] = None

Number of groups to divide experts into for group-limited routing. When using group-limited routing: 1. Experts are divided into ‘moe_router_num_groups’ equal-sized groups 2. For each token, ‘moe_router_group_topk’ groups are selected based on sum of top-(‘moe_router_topk’/’moe_router_group_topk’) routing scores within each group 3. From these selected groups, ‘moe_router_topk’ individual experts are chosen Two common use cases: - Device-limited routing: Set ‘moe_router_num_groups’ equal to expert parallel size (EP) to limit each token to experts on a subset of devices (See DeepSeek-V2: https://arxiv.org/pdf/2405.04434) - Node-limited routing: Set ‘moe_router_num_groups’ equal to number of nodes in EP group to limit each token to experts on a subset of nodes (See DeepSeek-V3: https://arxiv.org/pdf/2412.19437)

moe_router_padding_for_fp8: Optional[bool] = False

Whether to pad the routing_map to make sure the number of tokens each expert received is a multiple of 16/32 for FP8 precision. This can remove the explicit padding in the GroupedMLP layer.

moe_router_pre_softmax: bool = False

Enable pre-softmax(pre-sigmoid) routing for MoE, which means softmax is before the top-k selection. By default, softmax is done after top-k.

moe_router_score_function: str = 'softmax'

Score function for MoE routing. Can be “softmax” or “sigmoid”.

moe_router_topk: int = 2

Number of experts to route to for each token.

moe_router_topk_limited_devices: Optional[int] = None

Number of EP ranks to consider for each token in group-limited routing, DEPRECATED and replaced by moe_router_num_groups and moe_router_group_topk.

moe_router_topk_scaling_factor: Optional[float] = None

Scaling factor for routing score in top-k selection, only works when moe_router_pre_softmax enabled. Defaults to None, which means no scaling.

moe_shared_expert_intermediate_size: Optional[int] = None

Shared expert total ffn hidden size. It should be equal to ‘num_shared_experts * ffn_size_of_each_shared_expert’ if there are multiple shared experts. None means no shared expert.

moe_shared_expert_overlap: bool = False

Enable overlapping between shared expert computations and dispatcher communications. Without this, the shared epxerts execute after the routed experts.

moe_token_dispatcher_type: str = 'allgather'

The type of token dispatcher to use. The default is ‘allgather’. Options are ‘allgather’,’alltoall’ and ‘flex’.

moe_token_drop_policy: str = 'probs'

The policy to drop tokens. Can be either “probs” or “position”. If “probs”, the tokens with the lowest probabilities will be dropped. If “position”, tokens at the end of each batch will be dropped.

moe_token_dropping: bool = False

This feature involves selectively dropping and padding tokens for each expert to achieve a specified capacity, similar to GShard, Switch-Transformer, and DeepSpeed-MoE. Note that this is currently unsupported so should remain False.

moe_use_legacy_grouped_gemm: bool = False

Use legacy GroupedMLP rather than TEGroupedMLP. Note: The legacy one will be deprecated soon.

moe_z_loss_coeff: Optional[float] = None

Scaling coefficient for the z-loss. A starting value of 1e-3 is recommended.

mrope_section: Optional[List[int]] = None

Multimodal rope section is for channel dimension of temporal, height and width in rope calculation.

mtp_loss_scaling_factor: Optional[float] = None

Weighting factor of Multi-Token Prediction (MTP) loss.

mtp_num_layers: Optional[int] = None

Number of Multi-Token Prediction (MTP) Layers.

multi_latent_attention: bool = False

Whether to use multi-latent attention.

no_rope_freq: Optional[Union[int, List[int]]] = None

Controls which layers perform Rotary Position Embedding (RoPE). Accepts either: An integer N: Creates a pattern where RoPE is skipped every N-1 layers. For example, no_rope=4 means RoPE is applied for 3 layers, then skipped for 1 layer, repeating this pattern. A list of integers: Defines a custom pattern where 1 means skip RoPE and 0 means apply RoPE. For example, [0,1,1,0] means: apply RoPE, skip RoPE, skip RoPE, apply RoPE.

normalization: str = 'LayerNorm'

Which norm to use for normalization layers, valid options are LayerNorm and RMSNorm.

num_attention_heads: int = 0

Number of transformer attention heads.

num_layers: int = 0

Number of transformer layers in a transformer block.

num_layers_at_end_in_bf16: int = 1

Number of layers at the end of the model to keep in BF16 precision when first_last_layers_bf16 is True.

num_layers_at_start_in_bf16: int = 1

Number of layers at the start of the model to keep in BF16 precision when first_last_layers_bf16 is True.

num_layers_in_first_pipeline_stage: Optional[int] = None

Number of transformer layers on first pipeline stage. None implies equal layer division across PP ranks.

num_layers_in_last_pipeline_stage: Optional[int] = None

Number of transformer layers on last pipeline stage. None implies equal layer division across PP ranks.

num_moe_experts: Optional[int] = None

Number of experts to use for MoE layer. When set, it replaces MLP with MoE layer. Set to None for no MoE.

num_query_groups: Optional[int] = None

Number of query groups for group query attention. If None, normal attention is used.

output_layer_init_method: Optional[Callable] = None

Method to initialize weights of the output layer of both attention and MLP blocks. If None, will be set to megatron.core.utils.scaled_init_method_normal(init_method_std) which is torch nn init normal with mean=0.0 and std=init_method_std / math.sqrt(2.0 * num_layers).

persist_layer_norm: bool = False

If True, uses the persistent fused layer norm kernel. This kernel only supports a fixed set of hidden sizes.

pipeline_model_parallel_layout: Optional[Union[str, list, megatron.core.transformer.pipeline_parallel_layer_layout.PipelineParallelLayerLayout]] = None

Custom definition of the pipeline parallel partitioning. Support type: - str: e.g., ‘Et*3|(tt|)*29,m|L’. Stages are split by ‘|’, replicated stages or layers can be described with multiplication. Commas can be used cosmetically. - list: e.g., [[‘embedding’, ‘decoder’], [‘decoder’, ‘decoder’, ‘decoder’, ‘loss’]]. - PipelineParallelLayerLayout: a PipelineParallelLayerLayout object. If given either a string or a list, it will be transferred into a PipelineParallelLayerLayout in post init. Let i = a * pp_size + b, then layout[i] gives a list of the layers in the a-th vpp stage and the b-th pp stage, i.e., vpp(0)pp(0), vpp(0)pp(1), …, vpp(i)pp(j), vpp(i)pp(j+1), …, vpp(-1)pp(-2), vpp(-1)pp(-1). In the inner lists of layers, ‘embedding’ or ‘E’ denotes the embedding layer, ‘loss’ or ‘L’ denotes the loss function, and ‘decoder’ or ‘t’ denotes the transformer decoder layer. .. rubric:: Examples

[[‘embedding’, ‘decoder’], [‘decoder’, ‘decoder’, ‘decoder’, ‘loss’]]: pp = 2, vpp = None pp rank 0 holds: embedding, decoder pp rank 1 holds: decoder*3, loss ‘E|(tt|)*2,(t|)*4,mL’: pp = 2, vpp = 4 vpp rank 0 pp rank 0 holds: embedding vpp rank 0 pp rank 1~2 holds: decoder*2 vpp rank 0 pp rank 3 holds: decoder vpp rank 1 pp rank 0~2 holds: decoder vpp rank 1 pp rank 3 holds: mtp, loss

qk_layernorm: bool = False

Whether to apply normalization type of normalization to the query and key embeddings.

quant_recipe: Optional[megatron.core.quantization.quant_config.RecipeConfig] = None

Configuration of any quantization to be applied to the model

recompute_granularity: Optional[str] = None

Determines which type of activation recompute to use. Megatron-core supports ‘selective’ activation checkpointing where the submodules set in –recompute-modules is checkpointed. The default is “core_attn” which is the memory intensive part of attention. These memory intensive activations are also less compute intensive which makes activation checkpointing more efficient for LLMs (20B+). See Reducing Activation Recomputation in Large Transformer Models (https://arxiv.org/abs/2205.05198) for more details. ‘full’ will checkpoint the entire transformer layer. If None, no recompute is performed and all activations are saved. If set, must be ‘selective’ or ‘full’. ‘selective’ always uses all layers.

recompute_method: Optional[str] = None

Determines which transformer layers will be recomputed. uniform will uniformly divide the total number of transformer layers in a transformer block and recompute the input activation of each divided chunk at the specified granularity. block will recompute the input activations for only a set number of transformer layers per pipeline stage. The rest of the layers in the pipeline stage will not have any activations recomputed. If None, and recompute is enabled, all layers will do recomputation. If set, must be ‘uniform’ or ‘block’.

recompute_modules: Optional[List[str]] = None

The submodules to recompute. choices: “core_attn”, “moe_act”, “layernorm”, “mla_up_proj”, “mlp”, “moe”, “shared_experts”. default: [“core_attn”]. “core_attn”: recompute the core attention part of the transformer layer. “moe_act”: recompute the MoE MLP activation function. “layernorm”: recompute the input_layernorm and pre_mlp_layernorm. “mla_up_proj”: recompute the MLA up projection and RoPE applying parts. “mlp”: recompute the dense MLP submodule. “moe”: recompute the MoE layer. “shared_experts”: recompute the shared experts in the MoE layer. “moe_act”, “layernorm”, and “mla_up_proj” use output-discarding checkpointing, “core_attn”, “mlp”, “moe”, and “shared_experts” use normal checkpointing.

recompute_num_layers: Optional[int] = None

When recompute_method is uniform, recompute_num_layers is the number of transformer layers in each uniformly divided recompute unit. When recompute_method is block, recompute_num_layers is the number of transformer layers to recompute within each pipeline stage. Must be None for ‘selective’ activation checkpointing.

rotary_interleaved: bool = False

True is rotate pairs of even and odd dimensions (RoFormer style), False is rotate pairs of first half and second half (LLaMa style). Default to False.

softmax_scale: Optional[float] = None

Softmax scale for attention scaling.

softmax_type: Literal['vanilla', 'off-by-one', 'learnable'] = 'vanilla'

//www.evanmiller.org/attention-is-off-by-one.html. Supports both TE FusedAttention and local unfused attention. Supports both a fixed offset and and learnable offset.

Type

Applies modified softmax from https

symmetric_ar_type: Optional[str] = None

Type of symmetric all reduce to use

test_mode: bool = False

Whether to run real-time tests.

tp_only_amax_red: bool = False

When set to True, reduce the FP8 AMAX only in the TP or TP-CP domain

transformer_impl: str = 'transformer_engine'

Transformer implementation to use. Options are ‘transformer_engine’ for Transformer Engine and ‘local’ for MCore.

use_fused_weighted_squared_relu: bool = False

If True, uses fused weighted squared relu kernel when using MoE.

use_kitchen: bool = False

Use the kitchen extension for transformer quantization.

use_mamba_mem_eff_path: bool = True

If True, use the memory efficient path for Mamba layers.

use_te_activation_func: bool = False

Whether to use ffn activation functions implemented by TransformerEngine

use_te_rng_tracker: bool = False

Whether to use the TE or MCore version of the RNG tracker.

window_size: Optional[Tuple[int, int]] = None

If not None, then will use sliding window attention. The size of the window is specified by the numbers inside the tuple; -1 is special value meaning “infinite window size”.

A single standard transformer layer including attention and MLP blocks.

class core.transformer.transformer_layer.BaseTransformerLayer

Bases: abc.ABC

A common parent class for TransformerLayer like implementations.

A dummy class that is subclassed by similar TransformerLayer`s e.g. the `TransformerLayer in this file and possibly other TransformerLayer implementations that aim to use TransformerBlock as the base module. The main purpose is to check if any layer (or module) provided in the spec is a subclass of this class to allow fanning-out of that spec for all the layers in the TransformerBlock. See _get_block_submodules method implementation in transformer_block.py file for more details.

class core.transformer.transformer_layer.TransformerLayer(*args: Any, **kwargs: Any)

Bases: megatron.core.transformer.module.MegatronModule, core.transformer.transformer_layer.BaseTransformerLayer

A single transformer layer.

Transformer layer takes input with size [s, b, h] and returns an output of the same size.

forward(*args, **kwargs)

Perform a forward pass through the transformer layer.

This method calls the core computation of a transformer layer, including self-attention, cross-attention (if applicable), and feed-forward operations.

get_layer_static_inputs(seq_length, micro_batch_size)

Get the static inputs for the transformer layer.

Returns

A dictionary containing the static inputs for the layer.

Return type

Dict[str, torch.Tensor]

setup_manual_hooks(make_hook_func)

Set CUDA Graph manual hooks for the modules that contain direct parameters and are covered by cudagraphs.

sharded_state_dict(prefix: str = '', sharded_offsets: tuple = (), metadata: Optional[dict] = None) → megatron.core.dist_checkpointing.mapping.ShardedStateDict

Generate a sharded state dictionary for the transformer layer.

Parameters

• prefix (str, optional) – Prefix to be added to all keys in the state dict.

• sharded_offsets (tuple, optional) – Tuple of sharding offsets.

• metadata (Optional[dict], optional) – Additional metadata for sharding.

Returns

A dictionary containing the sharded state of the transformer layer.

Return type

ShardedStateDict

class core.transformer.transformer_layer.TransformerLayerSubmodules(input_layernorm: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityOp, self_attention: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityOp, self_attn_bda: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityFuncOp, pre_cross_attn_layernorm: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityOp, cross_attention: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityOp, cross_attn_bda: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityFuncOp, pre_mlp_layernorm: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityOp, mlp: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityOp, mlp_bda: typing.Union[megatron.core.transformer.spec_utils.ModuleSpec, type] = megatron.core.transformer.identity_op.IdentityFuncOp, sharded_state_dict_keys_map: typing.Dict[str, str] = <factory>)

Bases: object

Configuration class for specifying the submodules of a transformer layer.

This class defines the structure and default implementations for various components of a transformer layer, allowing for flexible customization of the layer’s architecture.

Parameters

• input_layernorm (Union[ModuleSpec, type]) – Specification for the input layer normalization.

• self_attention (Union[ModuleSpec, type]) – Specification for the self-attention mechanism.

• self_attn_bda (Union[ModuleSpec, type]) – Specification for the bias-dropout-add operation after self-attention.

• pre_cross_attn_layernorm (Union[ModuleSpec, type]) – Specification for the layer normalization before cross-attention.

• cross_attention (Union[ModuleSpec, type]) – Specification for the cross-attention mechanism.

• cross_attn_bda (Union[ModuleSpec, type]) – Specification for the bias-dropout-add operation after cross-attention.

• pre_mlp_layernorm (Union[ModuleSpec, type]) – Specification for the layer normalization before the MLP.

• mlp (Union[ModuleSpec, type]) – Specification for the MLP in Dense layer.

• mlp_bda (Union[ModuleSpec, type]) – Specification for the bias-dropout-add operation after the MLP.

• sharded_state_dict_keys_map (Dict[str, str]) – Mapping for sharded tensor keys to be applied in the sharded_state_dict method.

sharded_state_dict_keys_map: Dict[str, str]

core.transformer.transformer_layer.get_transformer_layer_offset(config: megatron.core.transformer.transformer_config.TransformerConfig, vp_stage: Optional[int] = None, pp_rank: Optional[int] = None)

Get the index offset of current pipeline stage, given the level of pipelining.

Various utilities used in the transformer implementation.

Utilities for transformer layers.

core.transformer.utils.attention_mask_func(attention_scores, attention_mask)

core.transformer.utils.erf_gelu(x)

core.transformer.utils.gelu_impl(x)

OpenAI’s gelu implementation.

core.transformer.utils.get_default_causal_mask(sq: int) → torch.Tensor

Return the causal upper triangular mask for softmax input.

core.transformer.utils.get_linear_layer(rows, columns, init_method, perform_initialization=True)

Simple linear layer with weight initialization.

core.transformer.utils.init_cuda_graph_cache(model)

Initialize the cache of modules for cuda graphs