nemo_automodel.components.models.nemotron_v3.model#

Module Contents#

Classes#

NemotronV3Model

NemotronV3 base model (without LM head).

NemotronHForCausalLM

NemotronV3 model with language modeling head.

Data#

ModelClass

API#

class nemo_automodel.components.models.nemotron_v3.model.NemotronV3Model(
config,
backend: nemo_automodel.components.models.common.BackendConfig | None = None,
*,
moe_config: nemo_automodel.components.moe.config.MoEConfig | None = None,
moe_overrides: dict | None = None,
)#

Bases: torch.nn.Module

NemotronV3 base model (without LM head).

This is a hybrid architecture with Mamba2, Attention, MLP, and MoE layers.

Initialization

Initialize NemotronV3Model.

Parameters:

  • config – NemotronH config with model parameters

  • backend – Backend configuration for MoE and other components

  • moe_config – MoE configuration (optional, will create default if None)

  • moe_overrides – Optional dict of overrides to apply to the default MoE config

forward(
input_ids: torch.LongTensor | None = None,
*,
attention_mask: torch.Tensor | None = None,
causal_mask_mapping: dict[str, torch.Tensor] | None = None,
inputs_embeds: torch.Tensor | None = None,
past_key_values=None,
cache_position: torch.LongTensor | None = None,
**kwargs: Any,
) → torch.Tensor#

Forward pass through the model. Supports BSHD [B, S, H] and THD [T, H].

initialize_weights(buffer_device: torch.device | None = None) → None#

Initialize model weights according to NemotronV3 spec.

Parameters:

buffer_device – Device to use for buffer initialization

class nemo_automodel.components.models.nemotron_v3.model.NemotronHForCausalLM(
config,
backend: nemo_automodel.components.models.common.BackendConfig | None = None,
**kwargs,
)#

Bases: nemo_automodel.components.models.common.HFCheckpointingMixin, transformers.generation.GenerationMixin, torch.nn.Module, nemo_automodel.components.moe.fsdp_mixin.MoEFSDPSyncMixin

NemotronV3 model with language modeling head.

Supports .generate() from transformers.generation.GenerationMixin with O(1) per-step KV caching for attention layers and recurrent state caching for Mamba2 layers.

Initialization

Initialize NemotronV3ForCausalLM.

Parameters:

  • config – NemotronH config

  • backend – Backend configuration

  • **kwargs – Additional arguments

_is_stateful: bool#

True

main_input_name: str#

‘input_ids’

classmethod from_config(
config,
backend: nemo_automodel.components.models.common.BackendConfig | None = None,
**kwargs,
)#

Create model from config.

Parameters:

  • config – NemotronH config

  • backend – Backend configuration

  • **kwargs – Additional arguments

Returns:

NemotronHForCausalLM instance

classmethod from_pretrained(
pretrained_model_name_or_path: str,
*model_args,
**kwargs,
)#

Load pretrained model.

Parameters:

  • pretrained_model_name_or_path – Path or name of pretrained model

  • *model_args – Additional positional arguments

  • **kwargs – Additional keyword arguments

Returns:

NemotronHForCausalLM instance

property device: torch.device#

Return the device of the first model parameter (required by GenerationMixin).

property dtype: torch.dtype#

Return the dtype of the first model parameter (used by cache construction).

get_input_embeddings()#
set_input_embeddings(value)#
get_output_embeddings()#
set_output_embeddings(new_embeddings)#
forward(
input_ids: Optional[torch.LongTensor] = None,
attention_mask: Optional[torch.Tensor] = None,
causal_mask_mapping: Optional[dict[str, torch.Tensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
past_key_values: Optional[Any] = None,
use_cache: Optional[bool] = None,
cache_position: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
padding_mask: Optional[torch.Tensor] = None,
logits_to_keep: Union[int, torch.Tensor] = 0,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
**kwargs: Any,
) → transformers.modeling_outputs.CausalLMOutputWithPast#

Forward pass with optional loss computation.

Supports both BSHD format (input_ids shape [B, S]) and THD format (input_ids shape [T] after squeeze_input_for_thd). When kwargs["qkv_format"] == "thd", inputs are squeezed to THD before the base-model forward and logits are unsqueezed back to [1, T, V] on exit.

Parameters:

  • input_ids – Input token IDs. BSHD: [B, S]; THD: [1, T] (squeezed internally).

  • attention_mask – 2D padding mask [B, S].

  • causal_mask_mapping – Dict with precomputed 4D causal masks.

  • inputs_embeds – Pre-computed input embeddings (optional).

  • labels – Token IDs for loss computation [B, S] (optional).

  • past_key_values – Optional NemotronHybridCache for incremental decoding.

  • use_cache – Whether to return past_key_values for subsequent steps.

  • cache_position – Token position indices for cache updates.

  • position_ids – Unused – accepted for API compatibility with GenerationMixin.

  • padding_mask – Padding mask [B, S] used by THD squeeze helper.

  • logits_to_keep – If > 0, only compute logits for the last logits_to_keep token positions (avoids materialising the full logit matrix during generation).

  • output_hidden_states – Whether to return hidden states.

  • return_dict – Accepted for API compatibility (always returns CausalLMOutputWithPast).

  • **kwargs – Additional arguments forwarded to the base model (e.g. seq_idx, cu_seqlens, qkv_format, CP kwargs).

Returns:

class:

~transformers.modeling_outputs.CausalLMOutputWithPast with logits (float32), optional loss, past_key_values, and hidden_states.

static _make_causal_mask(
query_len: int,
kv_len: int,
batch_size: int,
dtype: torch.dtype,
device: torch.device,
) → torch.Tensor#

Build a 4D SDPA-compatible causal mask.

Prefill (query_len == kv_len): standard lower-triangular causal mask. Decode (query_len == 1): all-zeros row allowing attention to all cached positions.

prepare_inputs_for_generation(
input_ids: torch.LongTensor,
attention_mask: Optional[torch.Tensor] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
past_key_values: Optional[Any] = None,
cache_position: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = True,
**kwargs,
) → dict#

Prepare model inputs for each generation step.

On the first call (prefill), creates a :class:NemotronHybridCache and forwards the full prompt. On subsequent calls (decode), only the newly generated token is forwarded.

Parameters:

  • input_ids – Accumulated token ids [batch_size, current_seq_len].

  • attention_mask – Padding mask [batch_size, current_seq_len].

  • inputs_embeds – Pre-computed embeddings for the first step (optional).

  • past_key_values – NemotronHybridCache from the previous step (None on first call).

  • cache_position – Token position indices.

  • use_cache – Whether to use caching (default True).

  • **kwargs – Remaining model kwargs.

Returns:

Dict of keyword arguments to pass to :meth:forward.

initialize_weights(
buffer_device: torch.device | None = None,
dtype: torch.dtype = torch.bfloat16,
) → None#

Initialize model weights.

Parameters:

  • buffer_device – Device to use for buffer initialization

  • dtype – Target dtype for model weights

nemo_automodel.components.models.nemotron_v3.model.ModelClass#

None