import math from typing import Optional, Tuple import torch from torch import nn from ...cache_utils import Cache from ...pytorch_utils import ALL_LAYERNORM_LAYERS from ...utils import is_flash_attn_2_available, is_flash_attn_greater_or_equal_2_10, logging from ..llama.modeling_llama import LlamaRMSNorm from ..olmo.configuration_olmo import OlmoConfig from ..olmo.modeling_olmo import ( OlmoAttention, OlmoDecoderLayer, OlmoFlashAttention2, OlmoForCausalLM, OlmoModel, OlmoPreTrainedModel, OlmoSdpaAttention, apply_rotary_pos_emb, repeat_kv, ) if is_flash_attn_2_available(): from ...modeling_flash_attention_utils import _flash_attention_forward logger = logging.get_logger(__name__) class Olmo2Config(OlmoConfig): r""" This is the configuration class to store the configuration of a [`Olmo2Model`]. It is used to instantiate an OLMo2 model according to the specified arguments, defining the model architecture. Instantiating a configuration with the defaults will yield a similar configuration to that of the [allenai/Olmo2-7B-1124-hf](https://huggingface.co/allenai/Olmo2-7B-1124-hf). Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from [`PretrainedConfig`] for more information. Args: vocab_size (`int`, *optional*, defaults to 50304): Vocabulary size of the Olmo2 model. Defines the number of different tokens that can be represented by the `inputs_ids` passed when calling [`Olmo2Model`] hidden_size (`int`, *optional*, defaults to 4096): Dimension of the hidden representations. intermediate_size (`int`, *optional*, defaults to 11008): Dimension of the MLP representations. num_hidden_layers (`int`, *optional*, defaults to 32): Number of hidden layers in the Transformer decoder. num_attention_heads (`int`, *optional*, defaults to 32): Number of attention heads for each attention layer in the Transformer decoder. num_key_value_heads (`int`, *optional*): This is the number of key_value heads that should be used to implement Grouped Query Attention. If `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed by meanpooling all the original heads within that group. For more details checkout [this paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `num_attention_heads`. hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): The non-linear activation function (function or string) in the decoder. max_position_embeddings (`int`, *optional*, defaults to 2048): The maximum sequence length that this model might ever be used with. initializer_range (`float`, *optional*, defaults to 0.02): The standard deviation of the truncated_normal_initializer for initializing all weight matrices. use_cache (`bool`, *optional*, defaults to `True`): Whether or not the model should return the last key/values attentions (not used by all models). Only relevant if `config.is_decoder=True`. pad_token_id (`int`, *optional*, defaults to 1): Padding token id. bos_token_id (`int`, *optional*): Beginning of stream token id. eos_token_id (`int`, *optional*, defaults to 50279): End of stream token id. tie_word_embeddings (`bool`, *optional*, defaults to `False`): Whether to tie weight embeddings rope_theta (`float`, *optional*, defaults to 10000.0): The base period of the RoPE embeddings. rope_scaling (`Dict`, *optional*): Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update `max_position_embeddings` to the expected new maximum. See the following thread for more information on how these scaling strategies behave: https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an experimental feature, subject to breaking API changes in future versions. attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): Whether to use a bias in the query, key, value and output projection layers during self-attention. attention_dropout (`float`, *optional*, defaults to 0.0): The dropout ratio for the attention probabilities. rms_norm_eps (`float`, *optional*, defaults to 1e-05): The epsilon used by the rms normalization layers. ```python >>> from transformers import Olmo2Model, Olmo2Config >>> # Initializing a Olmo2 7B style configuration >>> configuration = Olmo2Config() >>> # Initializing a model from the Olmo2 7B style configuration >>> model = Olmo2Model(configuration) >>> # Accessing the model configuration >>> configuration = model.config ``` """ model_type = "olmo2" def __init__( self, vocab_size=50304, hidden_size=4096, intermediate_size=11008, num_hidden_layers=32, num_attention_heads=32, num_key_value_heads=None, hidden_act="silu", max_position_embeddings=2048, initializer_range=0.02, use_cache=True, pad_token_id=1, bos_token_id=None, eos_token_id=50279, tie_word_embeddings=False, rope_theta=10000.0, rope_scaling=None, attention_bias=False, attention_dropout=0.0, rms_norm_eps=1e-5, **kwargs, ): super().__init__( vocab_size=vocab_size, hidden_size=hidden_size, intermediate_size=intermediate_size, num_hidden_layers=num_hidden_layers, num_attention_heads=num_attention_heads, num_key_value_heads=num_key_value_heads, hidden_act=hidden_act, max_position_embeddings=max_position_embeddings, initializer_range=initializer_range, use_cache=use_cache, pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, rope_theta=rope_theta, rope_scaling=rope_scaling, attention_bias=attention_bias, attention_dropout=attention_dropout, **kwargs, ) self.rms_norm_eps = rms_norm_eps del self.clip_qkv class Olmo2RMSNorm(LlamaRMSNorm): pass ALL_LAYERNORM_LAYERS.append(Olmo2RMSNorm) # Olmo2 attention is identical to OLMo attention except: # - Norm is applied to attention queries and keys. # - No qkv clipping. class Olmo2Attention(OlmoAttention): def __init__(self, config: Olmo2Config, layer_idx: Optional[int] = None): super().__init__(config, layer_idx=layer_idx) self.q_norm = Olmo2RMSNorm(self.num_heads * self.head_dim, config.rms_norm_eps) self.k_norm = Olmo2RMSNorm(self.num_key_value_heads * self.head_dim, config.rms_norm_eps) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() 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(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) cos, sin = self.rotary_emb(value_states, position_ids) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) if attention_mask is not None: # no matter the length, we just slice it causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] attn_weights = attn_weights + causal_mask # upcast attention to fp32 attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training) attn_output = torch.matmul(attn_weights, value_states) if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): raise ValueError( f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is" f" {attn_output.size()}" ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value class Olmo2FlashAttention2(OlmoFlashAttention2, Olmo2Attention): """ OLMo2 flash attention module. This module inherits from `Olmo2Attention` as the weights of the module stays untouched. The only required change would be on the forward pass where it needs to correctly call the public API of flash attention and deal with padding tokens in case the input contains any of them. """ def __init__(self, *args, **kwargs): Olmo2Attention.__init__(*args, **kwargs) # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: output_attentions = False bsz, q_len, _ = hidden_states.size() 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) # Flash attention requires the input to have the shape # batch_size x seq_length x head_dim x hidden_dim # therefore we just need to keep the original shape query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) cos, sin = self.rotary_emb(value_states, position_ids) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache # to be able to avoid many of these transpose/reshape/view. query_states = query_states.transpose(1, 2) key_states = key_states.transpose(1, 2) value_states = value_states.transpose(1, 2) dropout_rate = self.attention_dropout if self.training else 0.0 # In PEFT, usually we cast the layer norms in float32 for training stability reasons # therefore the input hidden states gets silently casted in float32. Hence, we need # cast them back in the correct dtype just to be sure everything works as expected. # This might slowdown training & inference so it is recommended to not cast the LayerNorms # in fp32. (OlmoRMSNorm handles it correctly) input_dtype = query_states.dtype if input_dtype == torch.float32: if torch.is_autocast_enabled(): target_dtype = torch.get_autocast_gpu_dtype() # Handle the case where the model is quantized elif hasattr(self.config, "_pre_quantization_dtype"): target_dtype = self.config._pre_quantization_dtype else: target_dtype = self.q_proj.weight.dtype logger.warning_once( f"The input hidden states seems to be silently casted in float32, this might be related to" f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" f" {target_dtype}." ) query_states = query_states.to(target_dtype) key_states = key_states.to(target_dtype) value_states = value_states.to(target_dtype) attn_output = _flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, position_ids=position_ids, dropout=dropout_rate, use_top_left_mask=self._flash_attn_uses_top_left_mask, is_causal=self.is_causal, ) attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous() attn_output = self.o_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value class Olmo2SdpaAttention(OlmoSdpaAttention, Olmo2Attention): # Adapted from Olmo2Attention.forward def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. logger.warning_once( "Olmo2Model is using Olmo2SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' ) return super().forward( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) bsz, q_len, _ = hidden_states.size() 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(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) cos, sin = self.rotary_emb(value_states, position_ids) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) causal_mask = attention_mask # if attention_mask is not None and cache_position is not None: if attention_mask is not None: causal_mask = causal_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. if query_states.device.type == "cuda" and causal_mask is not None: query_states = query_states.contiguous() key_states = key_states.contiguous() value_states = value_states.contiguous() # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.view(bsz, q_len, self.hidden_size) attn_output = self.o_proj(attn_output) return attn_output, None, past_key_value # The OLMo2 layers are identical to those of the OLMo model except: # - RMSNorm is used instead of standard layer norm. # - Norm is applied after attention/feedforward rather than before. class Olmo2DecoderLayer(OlmoDecoderLayer): def __init__(self, config: Olmo2Config, layer_idx: int): super().__init__(config, layer_idx=layer_idx) self.post_attention_layernorm = Olmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) self.post_feedforward_layernorm = Olmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) del self.input_layernorm def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, **kwargs, ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: residual = hidden_states # Self Attention hidden_states, self_attn_weights, present_key_value = self.self_attn( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, **kwargs, ) hidden_states = self.post_attention_layernorm(hidden_states) hidden_states = residual + hidden_states # Fully Connected residual = hidden_states hidden_states = self.mlp(hidden_states) hidden_states = self.post_feedforward_layernorm(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) return outputs class Olmo2PreTrainedModel(OlmoPreTrainedModel): pass # The OLMo2 model is identical to the OLMo model, except RMSNorm is used instead of # standard layer norm for the output norm. class Olmo2Model(OlmoModel): def __init__(self, config: Olmo2Config): super().__init__(config) self.layers = nn.ModuleList( [Olmo2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self.norm = Olmo2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) # The heads now only need to redefine the model inside to the correct `RobertaModel` class Olmo2ForCausalLM(OlmoForCausalLM): def __init__(self, config: Olmo2Config): super().__init__(config) self.model = Olmo2Model(config) __all__ = [ "Olmo2Config", "Olmo2ForCausalLM", "Olmo2Model", "Olmo2PreTrainedModel", ]