"""PyTorch Dbrx model.""" import math import warnings from copy import deepcopy from functools import partial from typing import Any, Callable, Dict, Optional, Tuple, Union import torch import torch.nn.functional as F import torch.utils.checkpoint from torch import nn from transformers.cache_utils import Cache, DynamicCache, StaticCache from transformers.modeling_attn_mask_utils import AttentionMaskConverter from transformers.modeling_outputs import (MoeCausalLMOutputWithPast, MoeModelOutputWithPast) from transformers.modeling_utils import PreTrainedModel from transformers.utils import is_flash_attn_2_available, logging from .configuration_dbrx import DbrxAttentionConfig, DbrxConfig, DbrxFFNConfig if is_flash_attn_2_available(): try: from flash_attn import flash_attn_func, flash_attn_varlen_func from flash_attn.bert_padding import pad_input # noqa from flash_attn.bert_padding import index_first_axis, unpad_input except: pass logger = logging.get_logger(__name__) _CONFIG_FOR_DOC = 'DbrxConfig' ############################################################################# # Copied from LLaMaRotaryEmbedding ############################################################################# class DbrxRotaryEmbedding(nn.Module): def __init__(self, dim: int, max_position_embeddings: int = 2048, base: float = 10000.0, scaling_factor: float = 1.0): super().__init__() self.scaling_factor = scaling_factor self.dim = dim self.max_position_embeddings = max_position_embeddings self.base = base inv_freq = 1.0 / (self.base**( torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim)) self.register_buffer('inv_freq', inv_freq, persistent=False) # For BC we register cos and sin cached self.max_seq_len_cached = max_position_embeddings @torch.no_grad() def forward( self, x: torch.Tensor, position_ids: torch.LongTensor ) -> Tuple[torch.Tensor, torch.Tensor]: # x: [bs, num_attention_heads, seq_len, head_size] inv_freq_expanded = self.inv_freq[None, :, None].float().expand( position_ids.shape[0], -1, 1) position_ids_expanded = position_ids[:, None, :].float() # Force float32 since bfloat16 loses precision on long contexts # See https://github.com/huggingface/transformers/pull/29285 device_type = x.device.type device_type = device_type if isinstance( device_type, str) and device_type != 'mps' else 'cpu' with torch.autocast(device_type=device_type, enabled=False): freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() sin = emb.sin() return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) def rotate_half(x: torch.Tensor) -> torch.Tensor: """Rotates half the hidden dims of the input.""" x1 = x[..., :x.shape[-1] // 2] x2 = x[..., x.shape[-1] // 2:] return torch.cat((-x2, x1), dim=-1) def apply_rotary_pos_emb( q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor, unsqueeze_dim: int = 1) -> Tuple[torch.Tensor, torch.Tensor]: """Applies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos and sin so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos and sin have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos and sin broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. """ cos = cos.unsqueeze(unsqueeze_dim) sin = sin.unsqueeze(unsqueeze_dim) q_embed = (q * cos) + (rotate_half(q) * sin) k_embed = (k * cos) + (rotate_half(k) * sin) return q_embed, k_embed def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: """Equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) """ batch, num_key_value_heads, slen, head_dim = hidden_states.shape if n_rep == 1: return hidden_states hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) ############################################################################# ############################################################################# # Modified from modeling_mixtral ############################################################################# def load_balancing_loss_func( gate_logits: torch.Tensor, num_experts: int, top_k: int, attention_mask: Optional[torch.Tensor], ) -> torch.Tensor: r"""Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch. See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between experts is too unbalanced. Args: gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]): Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of shape [batch_size X sequence_length, num_experts]. num_experts (`int`): Number of experts. top_k (`int`): The number of experts each token is routed to. attention_mask (`torch.Tensor`, None): The attention_mask used in forward function shape [batch_size X sequence_length] if not None. Returns: The auxiliary loss. """ if gate_logits is None or not isinstance(gate_logits, tuple): return torch.tensor(0.0) if isinstance(gate_logits, tuple): compute_device = gate_logits[0].device concatenated_gate_logits = torch.cat( [layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0) routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1) _, selected_experts = torch.topk(routing_weights, top_k, dim=-1) expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts) if attention_mask is None: # Compute the percentage of tokens routed to each experts tokens_per_expert = torch.mean(expert_mask.float(), dim=0) # Compute the average probability of routing to these experts router_prob_per_expert = torch.mean(routing_weights, dim=0) else: batch_size, sequence_length = attention_mask.shape num_hidden_layers = concatenated_gate_logits.shape[0] // ( batch_size * sequence_length) # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask expert_attention_mask = (attention_mask[None, :, :, None, None].expand( (num_hidden_layers, batch_size, sequence_length, top_k, num_experts)).reshape(-1, top_k, num_experts).to(compute_device)) # Compute the percentage of tokens routed to each experts tokens_per_expert = torch.sum( expert_mask.float() * expert_attention_mask, dim=0) / torch.sum( expert_attention_mask, dim=0) # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert router_per_expert_attention_mask = ( attention_mask[None, :, :, None].expand( (num_hidden_layers, batch_size, sequence_length, num_experts)).reshape(-1, num_experts).to(compute_device)) # Compute the average probability of routing to these experts router_prob_per_expert = torch.sum( routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(router_per_expert_attention_mask, dim=0) overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0)) return overall_loss * num_experts ############################################################################# def resolve_ffn_act_fn( ffn_act_fn: dict) -> Callable[[torch.Tensor], torch.Tensor]: """Resolve the activation function for the feed-forward network. Args: ffn_act_fn (dict): The configuration dictionary for the activation function. The dict config must specify the 'name' of a torch.nn.functional activation function. All of other key values pairs are bound to the function as a partial. Returns: Callable[[torch.Tensor], torch.Tensor]: The activation function. """ config = deepcopy(ffn_act_fn) name = config.pop('name') if not hasattr(nn.functional, name): raise ValueError(f'Unrecognised activation function name ({name}).') act = getattr(nn.functional, name) return partial(act, **config) ############################################################################# # Copied from LLaMaAttention ############################################################################# def _get_unpad_data(attention_mask: torch.Tensor): seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32) indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten() max_seqlen_in_batch = seqlens_in_batch.max().item() cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0)) return ( indices, cu_seqlens, max_seqlen_in_batch, ) class DbrxAttention(nn.Module): """Multi-head self attention.""" def __init__(self, hidden_size: int, num_heads: int, max_position_embeddings: int, attn_config: DbrxAttentionConfig, block_idx: Optional[int] = None): super().__init__() self.hidden_size = hidden_size self.num_heads = num_heads self.head_dim = self.hidden_size // self.num_heads self.max_position_embeddings = max_position_embeddings self.block_idx = block_idx self.config = attn_config if block_idx is None: logger.warning_once( f'Instantiating {self.__class__.__name__} without passing a `block_idx` is not recommended and will ' + 'lead to errors during the forward call if caching is used. Please make sure to provide a `block_idx` ' + 'when creating this class.') self.attn_pdrop = attn_config.attn_pdrop self.clip_qkv = attn_config.clip_qkv self.num_key_value_heads = attn_config.kv_n_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.rope_theta = attn_config.rope_theta self.Wqkv = nn.Linear(self.hidden_size, self.hidden_size + 2 * self.num_key_value_heads * self.head_dim, bias=False) self.out_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False) self.rotary_emb = DbrxRotaryEmbedding( self.head_dim, max_position_embeddings=self.max_position_embeddings, base=self.rope_theta, ) def forward( self, hidden_states: torch.Tensor, position_ids: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, **kwargs: Any, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache]]: bsz, q_len, _ = hidden_states.size() qkv_states = self.Wqkv(hidden_states) if self.clip_qkv is not None: qkv_states = qkv_states.clamp(min=-self.clip_qkv, max=self.clip_qkv) query_states, key_states, value_states = qkv_states.split( [ self.hidden_size, self.num_key_value_heads * self.head_dim, self.num_key_value_heads * self.head_dim, ], dim=2, ) 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) past_key_value = getattr(self, 'past_key_value', past_key_value) 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; position_ids 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.block_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.attn_pdrop, 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.out_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value class DbrxFlashAttention2(DbrxAttention): """Dbrx flash attention module. This module inherits from `DbrxAttention` as the weights of the module stays untouched. The only required change would be on the forward pass where it calls the public API of flash attention. """ def __init__(self, *args: Any, **kwargs: Any): if not is_flash_attn_2_available(): raise ImportError( 'Flash Attention 2 is not available. Please install it with `pip install flash-attn`.' ) super().__init__(*args, **kwargs) 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: Any, ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: logger.info( 'Implicitly setting `output_attentions` to False as it is not supported in Flash Attention.' ) output_attentions = False bsz, q_len, _ = hidden_states.size() qkv_states = self.Wqkv(hidden_states) if self.clip_qkv is not None: qkv_states = qkv_states.clamp(min=-self.clip_qkv, max=self.clip_qkv) query_states, key_states, value_states = qkv_states.split( [ self.hidden_size, self.num_key_value_heads * self.head_dim, self.num_key_value_heads * self.head_dim, ], dim=2, ) # 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) past_key_value = getattr(self, 'past_key_value', past_key_value) 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.block_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.attn_pdrop 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. (LlamaRMSNorm 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 = query_states.dtype logger.warning_once( f'The input hidden states seems to be silently casted in float32, this might be ' + f'related to the fact you have upcasted embedding or layer norm layers in ' + f'float32. We will cast back the input in {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 = self._flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate, ) attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous() attn_output = self.out_proj(attn_output) if not output_attentions: attn_weights = None return attn_output, attn_weights, past_key_value # type: ignore def _flash_attention_forward( self, query_states: torch.Tensor, key_states: torch.Tensor, value_states: torch.Tensor, attention_mask: Union[torch.LongTensor, None], query_length: int, dropout: float = 0.0, softmax_scale: Optional[float] = None, ): """Use FlashAttention, stripping padding tokens if necessary. Args: query_states (torch.Tensor): Input query states to be passed to Flash Attention API key_states (torch.Tensor): Input key states to be passed to Flash Attention API value_states (torch.Tensor): Input value states to be passed to Flash Attention API attention_mask (torch.LongTensor | None): The padding mask - corresponds to a tensor of size (batch_size, seq_len) where 0 stands for the position of padding tokens and 1 for the position of non-padding tokens. query_length (int): The length of the query sequence dropout (float): Attention dropout softmax_scale (float, optional): The scaling of QK^T before applying softmax. Defaults to 1 / sqrt(head_dim) """ causal = True # Contains at least one padding token in the sequence if attention_mask is not None: batch_size = query_states.shape[0] query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( query_states, key_states, value_states, attention_mask, query_length) cu_seqlens_q, cu_seqlens_k = cu_seq_lens max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens attn_output_unpad = flash_attn_varlen_func( query_states, key_states, value_states, cu_seqlens_q=cu_seqlens_q, cu_seqlens_k=cu_seqlens_k, max_seqlen_q=max_seqlen_in_batch_q, max_seqlen_k=max_seqlen_in_batch_k, dropout_p=dropout, softmax_scale=softmax_scale, causal=causal, ) attn_output = pad_input( attn_output_unpad, indices_q, batch_size, query_length, ) else: attn_output = flash_attn_func( query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal, ) return attn_output def _upad_input(self, query_layer: torch.Tensor, key_layer: torch.Tensor, value_layer: torch.Tensor, attention_mask: torch.Tensor, query_length: int): indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data( attention_mask) batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape key_layer = index_first_axis( key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k) value_layer = index_first_axis( value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k) if query_length == kv_seq_len: query_layer = index_first_axis( query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k) cu_seqlens_q = cu_seqlens_k max_seqlen_in_batch_q = max_seqlen_in_batch_k indices_q = indices_k elif query_length == 1: max_seqlen_in_batch_q = 1 cu_seqlens_q = torch.arange( batch_size + 1, dtype=torch.int32, device=query_layer.device ) # There is a memcpy here, that is very bad. indices_q = cu_seqlens_q[:-1] query_layer = query_layer.squeeze(1) else: # The -q_len: slice assumes left padding. attention_mask = attention_mask[:, -query_length:] query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input( query_layer, attention_mask) return ( query_layer, key_layer, value_layer, indices_q, (cu_seqlens_q, cu_seqlens_k), (max_seqlen_in_batch_q, max_seqlen_in_batch_k), ) DBRX_ATTENTION_CLASSES = { 'eager': DbrxAttention, 'flash_attention_2': DbrxFlashAttention2, } class DbrxNormAttentionNorm(nn.Module): def __init__( self, hidden_size: int, num_heads: int, max_position_embeddings: int, resid_pdrop: float, attn_implementation: str, attn_config: DbrxAttentionConfig, block_idx: Optional[int] = None, ): super().__init__() self.block_idx = block_idx self.resid_pdrop = resid_pdrop self.norm_1 = nn.LayerNorm(hidden_size, bias=False) self.attn = DBRX_ATTENTION_CLASSES[attn_implementation]( hidden_size=hidden_size, num_heads=num_heads, max_position_embeddings=max_position_embeddings, attn_config=attn_config, block_idx=block_idx, ) self.norm_2 = nn.LayerNorm(hidden_size, bias=False) def forward( self, hidden_states: torch.Tensor, position_ids: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, **kwargs: Any, ) -> Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor], Optional[Cache]]: residual_states = hidden_states hidden_states = self.norm_1(hidden_states).to(hidden_states.dtype) hidden_states, attn_weights, past_key_value = 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 = nn.functional.dropout(hidden_states, p=self.resid_pdrop, training=self.training) hidden_states = hidden_states + residual_states residual_states = hidden_states hidden_states = self.norm_2(hidden_states).to(hidden_states.dtype) return residual_states, hidden_states, attn_weights, past_key_value class DbrxRouter(nn.Module): def __init__(self, hidden_size: int, moe_num_experts: int, moe_top_k: int, moe_jitter_eps: Optional[float], moe_normalize_expert_weights: Optional[float], uniform_expert_assignment: bool): super().__init__() self.hidden_size = hidden_size self.moe_num_experts = moe_num_experts self.moe_top_k = moe_top_k self.moe_jitter_eps = moe_jitter_eps self.moe_normalize_expert_weights = moe_normalize_expert_weights self.uniform_expert_assignment = uniform_expert_assignment self.layer = nn.Linear(self.hidden_size, self.moe_num_experts, bias=False) def jitter(self, x: torch.Tensor) -> torch.Tensor: if self.moe_jitter_eps is None: raise RuntimeError('The router does not have moe_jitter_eps set.') low = 1.0 - self.moe_jitter_eps high = 1.0 + self.moe_jitter_eps noise = torch.rand(x.size(), dtype=x.dtype, device=x.device) return low + noise * (high - low) def forward( self, x: torch.Tensor ) -> Tuple[torch.Tensor, torch.Tensor, torch.LongTensor]: if self.training and self.moe_jitter_eps is not None: x = x * self.jitter(x) weights = self.layer(x.view(-1, x.shape[-1])).softmax(dim=-1, dtype=torch.float32) top_weights, top_experts = torch.topk(weights, self.moe_top_k, dim=-1) if self.moe_normalize_expert_weights: top_weights = top_weights / torch.norm( top_weights, p=self.moe_normalize_expert_weights, dim=-1, keepdim=True) if self.uniform_expert_assignment: with torch.no_grad(): uniform_tensor = torch.arange( 0, top_experts.numel(), device=top_experts.device, dtype=top_experts.dtype) % self.moe_num_experts top_experts = uniform_tensor.reshape(top_experts.shape) # Note, weights and top_weights are not changed weights = weights.to(x.dtype) top_weights = top_weights.to(x.dtype) return weights, top_weights, top_experts # type: ignore class DbrxExpertGLU(nn.Module): def __init__(self, hidden_size: int, ffn_hidden_size: int, moe_num_experts: int, ffn_act_fn: dict): super().__init__() self.hidden_size = hidden_size self.ffn_hidden_size = ffn_hidden_size self.moe_num_experts = moe_num_experts self.w1 = nn.Parameter( torch.empty(moe_num_experts * ffn_hidden_size, hidden_size)) self.v1 = nn.Parameter( torch.empty(moe_num_experts * ffn_hidden_size, hidden_size)) self.w2 = nn.Parameter( torch.empty(moe_num_experts * ffn_hidden_size, hidden_size)) self.activation_fn = resolve_ffn_act_fn(ffn_act_fn) def forward(self, x: torch.Tensor, expert_idx: int) -> torch.Tensor: expert_w1 = self.w1.view(self.moe_num_experts, self.ffn_hidden_size, self.hidden_size)[expert_idx] expert_v1 = self.v1.view(self.moe_num_experts, self.ffn_hidden_size, self.hidden_size)[expert_idx] expert_w2 = self.w2.view(self.moe_num_experts, self.ffn_hidden_size, self.hidden_size)[expert_idx] x1 = x.matmul(expert_w1.t()) x2 = x.matmul(expert_v1.t()) x1 = self.activation_fn(x1) x1 = x1 * x2 x1 = x1.matmul(expert_w2) return x1 class DbrxExperts(nn.Module): def __init__(self, hidden_size: int, ffn_hidden_size: int, moe_num_experts: int, ffn_act_fn: dict): super().__init__() self.moe_num_experts = moe_num_experts self.mlp = DbrxExpertGLU(hidden_size=hidden_size, ffn_hidden_size=ffn_hidden_size, moe_num_experts=moe_num_experts, ffn_act_fn=ffn_act_fn) def forward(self, x: torch.Tensor, weights: torch.Tensor, top_weights: torch.Tensor, top_experts: torch.LongTensor) -> torch.Tensor: bsz, q_len, hidden_size = x.shape x = x.view(-1, hidden_size) out = torch.zeros_like(x) expert_mask = nn.functional.one_hot( top_experts, num_classes=self.moe_num_experts).permute(2, 1, 0) for expert_idx in range(0, self.moe_num_experts): topk_idx, token_idx = torch.where(expert_mask[expert_idx]) if token_idx.shape[0] == 0: continue token_list = token_idx.tolist() topk_list = topk_idx.tolist() expert_tokens = x[None, token_list].reshape(-1, hidden_size) expert_out = self.mlp( expert_tokens, expert_idx) * top_weights[token_list, topk_list, None] out.index_add_(0, token_idx, expert_out) out = out.reshape(bsz, q_len, hidden_size) return out class DbrxFFN(nn.Module): def __init__(self, hidden_size: int, ffn_config: DbrxFFNConfig): super().__init__() self.router = DbrxRouter( hidden_size, moe_num_experts=ffn_config.moe_num_experts, moe_top_k=ffn_config.moe_top_k, moe_jitter_eps=ffn_config.moe_jitter_eps, moe_normalize_expert_weights=ffn_config. moe_normalize_expert_weights, uniform_expert_assignment=ffn_config.uniform_expert_assignment, ) self.experts = DbrxExperts( hidden_size=hidden_size, ffn_hidden_size=ffn_config.ffn_hidden_size, moe_num_experts=ffn_config.moe_num_experts, ffn_act_fn=ffn_config.ffn_act_fn, ) def forward(self, x: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: weights, top_weights, top_experts = self.router(x) out = self.experts(x, weights, top_weights, top_experts) return out, weights class DbrxBlock(nn.Module): def __init__(self, config: DbrxConfig, block_idx: int): super().__init__() self.hidden_size = config.d_model self.resid_pdrop = config.resid_pdrop self.block_idx = block_idx self.norm_attn_norm = DbrxNormAttentionNorm( hidden_size=config.d_model, num_heads=config.n_heads, max_position_embeddings=config.max_seq_len, resid_pdrop=config.resid_pdrop, attn_implementation=config._attn_implementation, attn_config=config.attn_config, block_idx=block_idx, ) self.ffn = DbrxFFN(hidden_size=config.d_model, ffn_config=config.ffn_config) def forward( self, hidden_states: torch.Tensor, position_ids: torch.LongTensor, attention_mask: Optional[torch.Tensor] = None, past_key_value: Optional[Cache] = None, output_attentions: Optional[bool] = False, output_router_logits: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, **kwargs: Any, ) -> Union[Tuple[torch.Tensor], Tuple[torch.Tensor, Optional[torch.Tensor]], Tuple[torch.Tensor, Optional[Cache]], Tuple[ torch.Tensor, Optional[torch.Tensor], Optional[Cache]], Tuple[torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]], Tuple[ torch.Tensor, Optional[Cache], Optional[torch.Tensor]], Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Cache], Optional[torch.Tensor]],]: """Forward function for DbrxBlock. Args: hidden_states (`torch.Tensor`): input to the layer of shape `(batch, seq_len, embed_dim)` position_ids (`torch.LongTensor`): position ids of shape `(batch, seq_len)` attention_mask (`torch.Tensor`, optional): attention mask of size (batch_size, sequence_length) if flash attention is used or (batch_size, 1, query_sequence_length, key_sequence_length) if default attention is used. past_key_value (`Tuple(torch.Tensor)`, optional): cached past key and value projection states output_attentions (`bool`, optional): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_router_logits (`bool`, optional): Whether or not to return the router logits. use_cache (`bool`, optional): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). cache_position (`torch.LongTensor`, optional): position ids of the cache """ if 'padding_mask' in kwargs: warnings.warn( 'Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`' ) # Norm + Attention + Norm resid_states, hidden_states, self_attn_weights, present_key_value = self.norm_attn_norm( 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, ) # Fully Connected hidden_states, router_logits = self.ffn(hidden_states) hidden_states = nn.functional.dropout(hidden_states, p=self.resid_pdrop, training=self.training) hidden_states = resid_states + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if use_cache: outputs += (present_key_value,) if output_router_logits: outputs += (router_logits,) return outputs class DbrxPreTrainedModel(PreTrainedModel): config_class = DbrxConfig base_model_prefix = 'transformer' supports_gradient_checkpointing = True _no_split_modules = ['DbrxBlock'] _skip_keys_device_placement = ['past_key_values'] _supports_flash_attn_2 = True _supports_sdpa = False _supports_cache_class = True def _init_weights(self, module: nn.Module): std = self.config.initializer_range if isinstance(module, nn.Linear): module.weight.data.normal_(mean=0.0, std=std) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=std) if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_() elif isinstance(module, nn.LayerNorm): module.weight.data.normal_(mean=0.0, std=std) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, DbrxExpertGLU): module.w1.data.normal_(mean=0.0, std=std) module.v1.data.normal_(mean=0.0, std=std) module.w2.data.normal_(mean=0.0, std=std) def _setup_cache(self, cache_cls: Any, max_batch_size: int, max_cache_len: int): # TODO: how to set var type of class? if self.config._attn_implementation == 'flash_attention_2' and cache_cls == StaticCache: raise ValueError( '`static` cache implementation is not compatible with ' + '`attn_implementation==flash_attention_2`. Make sure to use ' + '`spda` in the mean time and open an issue at https://github.com/huggingface/transformers.' ) for block in self.transformer.blocks: device = block.norm_attn_norm.norm_1.weight.device if hasattr(self.config, '_pre_quantization_dtype'): dtype = self.config._pre_quantization_dtype else: dtype = block.norm_attn_norm.attn.out_proj.weight.dtype block.norm_attn_norm.attn.past_key_value = cache_cls(self.config, max_batch_size, max_cache_len, device=device, dtype=dtype) def _reset_cache(self): for block in self.transformer.blocks: block.norm_attn_norm.attn.past_key_value = None class DbrxModel(DbrxPreTrainedModel): """Transformer decoder consisting of *config.num_hidden_layers* [`DbrxBlock`] layers. Args: config: DbrxConfig """ def __init__(self, config: DbrxConfig): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.emb_pdrop = config.emb_pdrop self.wte = nn.Embedding(config.vocab_size, config.d_model, self.padding_idx) self.blocks = nn.ModuleList([ DbrxBlock(config, block_idx) for block_idx in range(config.n_layers) ]) self.norm_f = nn.LayerNorm(config.d_model, bias=False) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> nn.Embedding: return self.wte def set_input_embeddings(self, value: nn.Embedding): self.wte = value def _autocast_input_embeddings(self, inputs_embeds: torch.Tensor) -> torch.Tensor: if inputs_embeds.device.type == 'cuda' and torch.is_autocast_enabled(): return inputs_embeds.to(dtype=torch.get_autocast_gpu_dtype()) elif inputs_embeds.device.type == 'cpu' and torch.is_autocast_cpu_enabled( ): return inputs_embeds.to(dtype=torch.get_autocast_cpu_dtype()) else: return inputs_embeds def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.Tensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, MoeModelOutputWithPast]: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = (output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states) output_router_logits = (output_router_logits if output_router_logits is not None else self.config.output_router_logits) use_cache = use_cache if use_cache is not None else self.config.use_cache return_dict = return_dict if return_dict is not None else self.config.use_return_dict if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError( 'You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one' ) if self.gradient_checkpointing and self.training and use_cache: logger.warning_once( '`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.' ) use_cache = False if inputs_embeds is None: inputs_embeds = self.wte(input_ids) inputs_embeds = self._autocast_input_embeddings( inputs_embeds) # type: ignore inputs_embeds = nn.functional.dropout(inputs_embeds, p=self.emb_pdrop, training=self.training) past_seen_tokens = 0 if use_cache: # kept for BC (cache positions) if not isinstance(past_key_values, StaticCache): past_key_values = DynamicCache.from_legacy_cache( past_key_values) past_seen_tokens = past_key_values.get_seq_length( # type: ignore ) if cache_position is None: if isinstance(past_key_values, StaticCache): raise ValueError( 'cache_position is a required argument when using StaticCache.' ) cache_position = torch.arange( # type: ignore past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device) if position_ids is None: position_ids = cache_position.unsqueeze(0) # type: ignore causal_mask = self._update_causal_mask(attention_mask, inputs_embeds, cache_position) # type: ignore # embed positions hidden_states = inputs_embeds # decoder layers all_hidden_states = () if output_hidden_states else None all_self_attns = () if output_attentions else None all_router_logits = () if output_router_logits else None next_decoder_cache = None for block in self.blocks: if output_hidden_states: all_hidden_states += (hidden_states,) # type: ignore if self.gradient_checkpointing and self.training: block_outputs = self._gradient_checkpointing_func( block.__call__, hidden_states, attention_mask=causal_mask, position_ids=position_ids, past_key_values=past_key_values, output_attentions=output_attentions, output_router_logits=output_router_logits, use_cache=use_cache, cache_position=cache_position, ) else: block_outputs = block( hidden_states, attention_mask=causal_mask, position_ids=position_ids, past_key_value=past_key_values, output_attentions=output_attentions, output_router_logits=output_router_logits, use_cache=use_cache, cache_position=cache_position, ) hidden_states = block_outputs[0] if use_cache: next_decoder_cache = block_outputs[ 2 if output_attentions else 1] if output_attentions: all_self_attns += (block_outputs[1],) # type: ignore if output_router_logits: all_router_logits += (block_outputs[-1],) # type: ignore hidden_states = self.norm_f(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) # type: ignore next_cache = None if use_cache: next_cache = ( next_decoder_cache.to_legacy_cache() # type: ignore if isinstance(next_decoder_cache, Cache) else next_decoder_cache) if not return_dict: return tuple(v for v in [ hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits ] if v is not None) return MoeModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=next_cache, hidden_states=all_hidden_states, attentions=all_self_attns, router_logits=all_router_logits, ) # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 def _update_causal_mask( self, attention_mask: Optional[torch.Tensor], input_tensor: torch.Tensor, cache_position: torch.Tensor) -> Optional[torch.Tensor]: if self.config._attn_implementation == 'flash_attention_2': if attention_mask is not None and 0.0 in attention_mask: return attention_mask return None dtype, device = input_tensor.dtype, input_tensor.device min_dtype = torch.finfo(dtype).min sequence_length = input_tensor.shape[1] if hasattr(self.blocks[0].norm_attn_norm.attn, 'past_key_value'): # static cache target_length = self.config.max_position_embeddings else: # dynamic cache target_length = (attention_mask.shape[-1] if isinstance( attention_mask, torch.Tensor) else cache_position[-1] + 1) target_length = int(target_length) causal_mask = torch.full((sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device) if sequence_length != 1: causal_mask = torch.triu(causal_mask, diagonal=1) causal_mask *= torch.arange( target_length, device=device) > cache_position.reshape(-1, 1) causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) if attention_mask is not None: causal_mask = causal_mask.clone( ) # copy to contiguous memory for in-place edit if attention_mask.dim() == 2: mask_length = attention_mask.shape[-1] padding_mask = causal_mask[..., :mask_length].eq( 0.0) * attention_mask[:, None, None, :].eq(0.0) causal_mask[..., :mask_length] = causal_mask[ ..., :mask_length].masked_fill(padding_mask, min_dtype) elif attention_mask.dim() == 4: # backwards compatibility: we allow passing a 4D attention mask shorter than the input length with # cache. In that case, the 4D attention mask attends to the newest tokens only. if attention_mask.shape[ -2] < cache_position[0] + sequence_length: offset = cache_position[0] else: offset = 0 mask_shape = attention_mask.shape mask_slice = (attention_mask.eq(0.0)).to( dtype=dtype) * min_dtype causal_mask[:mask_shape[0], :mask_shape[1], offset:mask_shape[2] + offset, :mask_shape[3]] = mask_slice if (self.config._attn_implementation == 'sdpa' and attention_mask is not None and attention_mask.device.type == 'cuda'): # TODO: For dynamo, rather use a check on fullgraph=True once this is possible (https://github.com/pytorch/pytorch/pull/120400). is_tracing = ( torch.jit.is_tracing() or isinstance(input_tensor, torch.fx.Proxy) or # type: ignore (hasattr(torch, '_dynamo') and torch._dynamo.is_compiling())) if not is_tracing and torch.any(attention_mask != 1): # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. # Details: https://github.com/pytorch/pytorch/issues/110213 causal_mask = AttentionMaskConverter._unmask_unattended( causal_mask, min_dtype) return causal_mask class DbrxForCausalLM(DbrxPreTrainedModel): def __init__(self, config: DbrxConfig): super().__init__(config) self.transformer = DbrxModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.router_aux_loss_coef = config.router_aux_loss_coef self.num_experts = config.ffn_config.moe_num_experts self.num_experts_per_tok = config.ffn_config.moe_top_k # Initialize weights and apply final processing self.post_init() def get_input_embeddings(self) -> nn.Embedding: return self.transformer.get_input_embeddings() def set_input_embeddings(self, value: nn.Embedding): self.transformer.set_input_embeddings(value) def get_output_embeddings(self) -> nn.Linear: return self.lm_head def set_output_embeddings(self, new_embeddings: nn.Linear): self.lm_head = new_embeddings def set_decoder(self, decoder: DbrxModel): self.transformer = decoder def get_decoder(self) -> DbrxModel: return self.transformer def forward( self, input_ids: Optional[torch.LongTensor] = None, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Cache] = None, inputs_embeds: Optional[torch.Tensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, return_dict: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, ) -> Union[Tuple, MoeCausalLMOutputWithPast]: r"""Forward function for causal language modeling. Example: ```python >>> from transformers import AutoTokenizer, DbrxForCausalLM >>> model = DbrxForCausalLM.from_pretrained("databricks/dbrx") >>> tokenizer = AutoTokenizer.from_pretrained("databricks/dbrx") >>> prompt = "Hey, are you conscious? Can you talk to me?" >>> inputs = tokenizer(prompt, return_tensors="pt") >>> # Generate >>> generate_ids = model.generate(inputs.input_ids, max_length=30) >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." ``` """ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_hidden_states = (output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states) output_router_logits = (output_router_logits if output_router_logits is not None else self.config.output_router_logits) return_dict = return_dict if return_dict is not None else self.config.use_return_dict # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs = self.transformer( input_ids=input_ids, attention_mask=attention_mask, position_ids=position_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, output_router_logits=output_router_logits, return_dict=return_dict, cache_position=cache_position, ) hidden_states = outputs[0] logits = self.lm_head(hidden_states) loss = None if labels is not None: # Shift so that tokens < n predict n shift_logits = logits[..., :-1, :].contiguous() shift_labels = labels[..., 1:].contiguous() # Flatten the tokens loss_fct = nn.CrossEntropyLoss() shift_logits = shift_logits.view(-1, self.config.vocab_size) shift_labels = shift_labels.view(-1) # Enable model parallelism shift_labels = shift_labels.to(shift_logits.device) loss = loss_fct(shift_logits, shift_labels) aux_loss = None if output_router_logits: aux_loss = load_balancing_loss_func( outputs.router_logits if return_dict else outputs[-1], self.num_experts, self.num_experts_per_tok, attention_mask, ) if labels is not None and loss is not None: loss += self.router_aux_loss_coef * aux_loss.to( loss.device) # make sure to reside in the same device if not return_dict: output = (logits,) + outputs[1:] return (loss,) + output if loss is not None else output return MoeCausalLMOutputWithPast( loss=loss, aux_loss=aux_loss, logits=logits, past_key_values=outputs.past_key_values, hidden_states=outputs.hidden_states, attentions=outputs.attentions, router_logits=outputs.router_logits, ) def prepare_inputs_for_generation( self, input_ids: torch.Tensor, past_key_values: Optional[Cache] = None, attention_mask: Optional[torch.Tensor] = None, inputs_embeds: Optional[torch.Tensor] = None, **kwargs: Any) -> Dict[str, Any]: past_length = 0 if past_key_values is not None: if isinstance(past_key_values, Cache): cache_length = past_key_values.get_seq_length() past_length = past_key_values.seen_tokens max_cache_length = past_key_values.get_max_length() else: cache_length = past_length = past_key_values[0][0].shape[2] max_cache_length = None # Keep only the unprocessed tokens: # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as # input) if attention_mask is not None and attention_mask.shape[ 1] > input_ids.shape[1]: input_ids = input_ids[:, -(attention_mask.shape[1] - past_length):] # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard # input_ids based on the past_length. elif past_length < input_ids.shape[1]: input_ids = input_ids[:, past_length:] # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. # If we are about to go beyond the maximum cache length, we need to crop the input attention mask. if (max_cache_length is not None and attention_mask is not None and cache_length + input_ids.shape[1] > max_cache_length): attention_mask = attention_mask[:, -max_cache_length:] position_ids = kwargs.get('position_ids', None) if attention_mask is not None and position_ids is None: # create position_ids on the fly for batch generation position_ids = attention_mask.long().cumsum(-1) - 1 position_ids.masked_fill_(attention_mask == 0, 1) if past_key_values: position_ids = position_ids[:, -input_ids.shape[1]:] if self.generation_config.cache_implementation == 'static': # generation with static cache cache_position = kwargs.get('cache_position', None) if cache_position is None: past_length = 0 else: past_length = cache_position[-1] + 1 input_ids = input_ids[:, past_length:] position_ids = position_ids[:, past_length:] if position_ids is not None else None # TODO @gante we should only keep a `cache_position` in generate, and do +=1. # same goes for position ids. Could also help with continued generation. input_length = position_ids.shape[ -1] if position_ids is not None else input_ids.shape[-1] cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device) position_ids = position_ids.contiguous( ) if position_ids is not None else None # if `inputs_embeds` are passed, we only want to use them in the 1st generation step if inputs_embeds is not None and past_key_values is None: model_inputs = {'inputs_embeds': inputs_embeds} else: # The `contiguous()` here is necessary to have a static stride during decoding. torchdynamo otherwise # recompiles graphs as the stride of the inputs is a guard. Ref: https://github.com/huggingface/transformers/pull/29114 # TODO: use `next_tokens` directly instead. model_inputs = {'input_ids': input_ids.contiguous()} model_inputs.update( { # type: ignore 'position_ids': position_ids, 'cache_position': cache_position, 'past_key_values': past_key_values, 'use_cache': kwargs.get('use_cache'), 'attention_mask': attention_mask, } ) return model_inputs @staticmethod def _reorder_cache(past_key_values: Cache, beam_idx: torch.LongTensor): reordered_past = () for layer_past in past_key_values: reordered_past += (tuple( past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),) return reordered_past