OLM API Reference

`olm.models.meta.llama3`

Source: src/olm/models/meta/llama3.py:1

Classes

Llama3Block(embed_dim: int, intermediate_size: int, num_heads: int, num_kv_heads: int, max_seq_len: int, dropout: float, rope_theta: float)

Bases: olm.nn.structure.block.Block

Source: src/olm/models/meta/llama3.py:10

A single Transformer block for Llama 3.x architecture.

Similar to Llama 2 but parameterized for Llama 3's high-performance context.

Structure

x = x + GQA(RMSNorm(x)) x = x + SwiGLU(RMSNorm(x))

Parameters

  • embed_dim (int): Model dimension.
  • intermediate_size (int): FFN hidden dimension.
  • num_heads (int): Number of attention heads.
  • num_kv_heads (int): Number of KV heads.
  • max_seq_len (int): Max sequence length.
  • dropout (float): Dropout probability.
  • rope_theta (float): RoPE base.

Methods

forward(self, x: torch.Tensor) -> torch.Tensor (inherited from Block)

Source: src/olm/nn/structure/block.py:26

Apply each block to the input in sequence.

Parameters

  • x: Input tensor.

Returns

Output tensor after all blocks have been applied.

Llama3Model(vocab_size: int, embed_dim: int, intermediate_size: int, num_layers: int, num_heads: int, num_kv_heads: int, max_seq_len: int, rope_theta: float = 500000.0, dropout: float = 0.0, tie_weights: bool = True)

Bases: olm.nn.structure.block.Block

Source: src/olm/models/meta/llama3.py:75

Base class for Llama 3, 3.1, and 3.2 models.

Inherits from Block for pure sequential composition.

Implementation Note

This implementation uses standard Rotary Positional Embeddings (RoPE) parameterized via rope_theta. Llama 3.1/3.2 official checkpoints use specialized scaled RoPE behavior for long contexts, so exact long-context behavior may differ from the released Meta checkpoints.

Structure

Embedding -> [Llama3Block] x N -> RMSNorm -> tied OutputHead.

Forward

Accepts token IDs shaped [batch, seq_len] and returns logits shaped [batch, seq_len, vocab_size].

Methods

forward(self, x: torch.Tensor) -> torch.Tensor (inherited from Block)

Source: src/olm/nn/structure/block.py:26

Apply each block to the input in sequence.

Parameters

  • x: Input tensor.

Returns

Output tensor after all blocks have been applied.

Llama3_1_405B()

Bases: olm.models.meta.llama3.Llama3Model

Source: src/olm/models/meta/llama3.py:139

Llama 3.1 405B Model (Flagship).

Methods

forward(self, x: torch.Tensor) -> torch.Tensor (inherited from Block)

Source: src/olm/nn/structure/block.py:26

Apply each block to the input in sequence.

Parameters

  • x: Input tensor.

Returns

Output tensor after all blocks have been applied.

Llama3_1_70B()

Bases: olm.models.meta.llama3.Llama3Model

Source: src/olm/models/meta/llama3.py:155

Llama 3.1 70B Model.

Methods

forward(self, x: torch.Tensor) -> torch.Tensor (inherited from Block)

Source: src/olm/nn/structure/block.py:26

Apply each block to the input in sequence.

Parameters

  • x: Input tensor.

Returns

Output tensor after all blocks have been applied.

Llama3_1_8B()

Bases: olm.models.meta.llama3.Llama3Model

Source: src/olm/models/meta/llama3.py:171

Llama 3.1 8B Model.

Methods

forward(self, x: torch.Tensor) -> torch.Tensor (inherited from Block)

Source: src/olm/nn/structure/block.py:26

Apply each block to the input in sequence.

Parameters

  • x: Input tensor.

Returns

Output tensor after all blocks have been applied.

Llama3_2_1B()

Bases: olm.models.meta.llama3.Llama3Model

Source: src/olm/models/meta/llama3.py:206

Llama 3.2 1B Model (Pruned/Distilled).

Methods

forward(self, x: torch.Tensor) -> torch.Tensor (inherited from Block)

Source: src/olm/nn/structure/block.py:26

Apply each block to the input in sequence.

Parameters

  • x: Input tensor.

Returns

Output tensor after all blocks have been applied.

Llama3_2_3B()

Bases: olm.models.meta.llama3.Llama3Model

Source: src/olm/models/meta/llama3.py:190

Llama 3.2 3B Model (Edge-optimized).

Methods

forward(self, x: torch.Tensor) -> torch.Tensor (inherited from Block)

Source: src/olm/nn/structure/block.py:26

Apply each block to the input in sequence.

Parameters

  • x: Input tensor.

Returns

Output tensor after all blocks have been applied.