# Copyright 2025 the HuggingFace Team. All rights reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. from collections.abc import Callable import torch import torch.nn as nn from huggingface_hub.dataclasses import strict from ...cache_utils import Cache, DynamicCache from ...masking_utils import create_causal_mask, create_sliding_window_causal_mask from ...modeling_outputs import BaseModelOutputWithPast from ...modeling_utils import ALL_ATTENTION_FUNCTIONS from ...processing_utils import Unpack from ...utils import auto_docstring from ...utils.generic import TransformersKwargs from ..gemma2.modeling_gemma2 import Gemma2RotaryEmbedding from ..olmo2.configuration_olmo2 import Olmo2Config from ..olmo2.modeling_olmo2 import ( Olmo2Attention, Olmo2DecoderLayer, Olmo2ForCausalLM, Olmo2Model, Olmo2PreTrainedModel, Olmo2RMSNorm, apply_rotary_pos_emb, eager_attention_forward, ) @auto_docstring(checkpoint="allenai/Olmo-3-7B-Instruct") @strict class Olmo3Config(Olmo2Config): r""" Example: ```python >>> from transformers import Olmo3Model, Olmo3Config >>> # Initializing a Olmo3 7B style configuration >>> configuration = Olmo3Config() >>> # Initializing a model from the Olmo3 7B style configuration >>> model = Olmo3Model(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` """ model_type = "olmo3" keys_to_ignore_at_inference = ["past_key_values"] base_model_tp_plan = { "layers.*.self_attn.q_proj": "colwise_gather_output", # we need to replicate here due to the added norm on q and k "layers.*.self_attn.k_proj": "colwise_gather_output", # we need to replicate here due to the added norm on q and k "layers.*.self_attn.v_proj": "colwise_gather_output", # we need to replicate here due to the added norm on q and k "layers.*.self_attn.o_proj": "rowwise_split_input", # input is replicated due to the added norm on q and k "layers.*.mlp.gate_proj": "colwise", "layers.*.mlp.up_proj": "colwise", "layers.*.mlp.down_proj": "rowwise", } base_model_pp_plan = { "embed_tokens": (["input_ids"], ["inputs_embeds"]), "layers": (["hidden_states", "attention_mask"], ["hidden_states"]), "norm": (["hidden_states"], ["hidden_states"]), } sliding_window: int | None = 4096 layer_types: list[str] | None = None def __post_init__(self, **kwargs): if self.num_key_value_heads is None: self.num_key_value_heads = self.num_attention_heads if self.layer_types is None: self.layer_types = [ "sliding_attention" if (i + 1) % 4 != 0 else "full_attention" for i in range(self.num_hidden_layers) ] super().__post_init__(**kwargs) class Olmo3RMSNorm(Olmo2RMSNorm): pass # Olmo3 attention is identical to OLMo 2 attention except: # - Sliding window attention is used for 3 out of 4 layers. class Olmo3Attention(Olmo2Attention): def __init__(self, config: Olmo3Config, layer_idx: int): super().__init__(config, layer_idx=layer_idx) self.attention_type = config.layer_types[layer_idx] self.sliding_window = config.sliding_window if self.attention_type == "sliding_attention" else None def forward( self, hidden_states: torch.Tensor, position_embeddings: tuple[torch.Tensor, torch.Tensor], attention_mask: torch.Tensor | None, past_key_values: Cache | None = None, **kwargs: Unpack[TransformersKwargs], ) -> tuple[torch.Tensor, torch.Tensor | None]: input_shape = hidden_states.shape[:-1] hidden_shape = (*input_shape, -1, self.head_dim) query_states = self.q_norm(self.q_proj(hidden_states)) key_states = self.k_norm(self.k_proj(hidden_states)) value_states = self.v_proj(hidden_states) query_states = query_states.view(hidden_shape).transpose(1, 2) key_states = key_states.view(hidden_shape).transpose(1, 2) value_states = value_states.view(hidden_shape).transpose(1, 2) cos, sin = position_embeddings query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_values is not None: key_states, value_states = past_key_values.update(key_states, value_states, self.layer_idx) attention_interface: Callable = ALL_ATTENTION_FUNCTIONS.get_interface( self.config._attn_implementation, eager_attention_forward ) attn_output, attn_weights = attention_interface( self, query_states, key_states, value_states, attention_mask, dropout=0.0 if not self.training else self.attention_dropout, scaling=self.scaling, sliding_window=self.sliding_window, **kwargs, ) attn_output = attn_output.reshape(*input_shape, -1).contiguous() attn_output = self.o_proj(attn_output) return attn_output, attn_weights class Olmo3DecoderLayer(Olmo2DecoderLayer): pass class Olmo3RotaryEmbedding(Gemma2RotaryEmbedding): pass class Olmo3PreTrainedModel(Olmo2PreTrainedModel): pass # The OLMo 3 model is identical to the OLMo 2 model, except: # - Sliding window attention is used for 3 out of 4 layers. # - RoPE scaling is not applied to sliding window attention layers. class Olmo3Model(Olmo2Model): def __init__(self, config: Olmo3Config): super().__init__(config) self.norm = Olmo3RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.layers = nn.ModuleList( [Olmo3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.rotary_emb = Olmo3RotaryEmbedding(config=config) def forward( self, input_ids: torch.LongTensor | None = None, attention_mask: torch.Tensor | None = None, position_ids: torch.LongTensor | None = None, past_key_values: Cache | None = None, inputs_embeds: torch.FloatTensor | None = None, use_cache: bool | None = None, **kwargs: Unpack[TransformersKwargs], ) -> BaseModelOutputWithPast: if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") if inputs_embeds is None: inputs_embeds: torch.Tensor = self.embed_tokens(input_ids) if use_cache and past_key_values is None: past_key_values = DynamicCache(config=self.config) if position_ids is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 position_ids = torch.arange(inputs_embeds.shape[1], device=inputs_embeds.device) + past_seen_tokens position_ids = position_ids.unsqueeze(0) # It may already have been prepared by e.g. `generate` if not isinstance(causal_mask_mapping := attention_mask, dict): # Prepare mask arguments mask_kwargs = { "config": self.config, "inputs_embeds": inputs_embeds, "attention_mask": attention_mask, "past_key_values": past_key_values, "position_ids": position_ids, } # Create the masks causal_mask_mapping = { "full_attention": create_causal_mask(**mask_kwargs), "sliding_attention": create_sliding_window_causal_mask(**mask_kwargs), } hidden_states = inputs_embeds position_embeddings = self.rotary_emb(hidden_states, position_ids) for i, decoder_layer in enumerate(self.layers[: self.config.num_hidden_layers]): hidden_states = decoder_layer( hidden_states, attention_mask=causal_mask_mapping[self.config.layer_types[i]], position_ids=position_ids, past_key_values=past_key_values, position_embeddings=position_embeddings, **kwargs, ) hidden_states = self.norm(hidden_states) return BaseModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values, ) class Olmo3ForCausalLM(Olmo2ForCausalLM): pass __all__ = [ "Olmo3Config", "Olmo3ForCausalLM", "Olmo3Model", "Olmo3PreTrainedModel", ]