# coding=utf-8 # Copyright 2024 Microsoft and the HuggingFace Inc. 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. """PyTorch Phimoe model.""" import math from typing import Optional, Union import torch import torch.utils.checkpoint from torch import nn from ...activations import ACT2FN from ...cache_utils import Cache, DynamicCache, SlidingWindowCache, StaticCache from ...generation import GenerationMixin from ...modeling_attn_mask_utils import AttentionMaskConverter, _prepare_4d_causal_attention_mask from ...modeling_flash_attention_utils import is_flash_attn_available from ...modeling_layers import ( GenericForSequenceClassification, GradientCheckpointingLayer, ) from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS from ...modeling_utils import PreTrainedModel from ...utils import auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging from .configuration_phimoe import PhimoeConfig if is_flash_attn_available(): from ...modeling_flash_attention_utils import _flash_attention_forward if is_torch_flex_attn_available(): from torch.nn.attention.flex_attention import BlockMask from ...integrations.flex_attention import make_flex_block_causal_mask # This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph. # It means that the function will not be traced through and simply appear as a node in the graph. _prepare_4d_causal_attention_mask = torch.fx.wrap(_prepare_4d_causal_attention_mask) logger = logging.get_logger(__name__) # Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func def load_balancing_loss_func( gate_logits: Union[torch.Tensor, tuple[torch.Tensor], None], num_experts: Optional[int] = None, top_k=2, attention_mask: Optional[torch.Tensor] = None, ) -> Union[torch.Tensor, int]: r""" Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch. See Switch Transformer (https://huggingface.co/papers/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: 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: Number of experts top_k: The number of experts to route per-token, can be also interpreted as the `top-k` routing parameter. attention_mask (`torch.Tensor`, *optional*): 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 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 class PhimoeRotaryEmbedding(nn.Module): def __init__( self, config: Optional[PhimoeConfig] = None, ): super().__init__() self.config = config if config.rope_scaling is not None: self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) self.short_mscale = config.rope_scaling.get("short_mscale") self.long_mscale = config.rope_scaling.get("long_mscale") else: self.rope_type = "default" self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] def forward(self, x, seq_len=None): mscale = None if self.config.rope_scaling and seq_len: mscale = ( self.long_mscale if seq_len > self.config.rope_scaling["original_max_position_embeddings"] else self.short_mscale ) inv_freq, attention_scaling = self.rope_init_fn(self.config, x.device, seq_len) mscale = attention_scaling if mscale is None else mscale t = torch.arange(seq_len, device=x.device, dtype=torch.float32) freqs = torch.outer(t, inv_freq) emb = torch.cat((freqs, freqs), dim=-1) return (emb.cos() * mscale).to(x.dtype), (emb.sin() * mscale).to(x.dtype) # Copied from transformers.models.llama.modeling_llama.rotate_half def rotate_half(x): """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, k, cos, sin, position_ids, unsqueeze_dim=1): """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. position_ids (`torch.Tensor`): The position indices of the tokens corresponding to the query and key tensors. For example, this can be used to pass offsetted position ids when working with a KV-cache. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] 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[position_ids] and sin[position_ids] 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[position_ids].unsqueeze(unsqueeze_dim) sin = sin[position_ids].unsqueeze(unsqueeze_dim) q_embed = (q * cos) + (rotate_half(q) * sin) k_embed = (k * cos) + (rotate_half(k) * sin) return q_embed, k_embed # Copied from transformers.models.llama.modeling_llama.repeat_kv def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: """ This is the 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) class PhimoeAttention(nn.Module): """ Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer and "Generating Long Sequences with Sparse Transformers". """ def __init__(self, config: PhimoeConfig, layer_idx: Optional[int] = None): super().__init__() self.config = config self.layer_idx = layer_idx if layer_idx is None: logger.warning_once( f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will " "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` " "when creating this class." ) self.hidden_size = config.hidden_size self.num_heads = config.num_attention_heads self.head_dim = self.hidden_size // self.num_heads self.num_key_value_heads = config.num_key_value_heads self.num_key_value_groups = self.num_heads // self.num_key_value_heads self.max_position_embeddings = config.max_position_embeddings self.rope_theta = config.rope_theta self.is_causal = True self.attention_dropout = config.attention_dropout if (self.head_dim * self.num_heads) != self.hidden_size: raise ValueError( f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}" f" and `num_heads`: {self.num_heads})." ) self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=self.config.attention_bias) self.k_proj = nn.Linear( self.hidden_size, self.num_key_value_heads * self.head_dim, bias=self.config.attention_bias ) self.v_proj = nn.Linear( self.hidden_size, self.num_key_value_heads * self.head_dim, bias=self.config.attention_bias ) self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=self.config.attention_bias) def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int): return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous() 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, position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None, ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states) key_states = 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 = position_embeddings query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads 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 class PhimoeFlashAttention2(PhimoeAttention): """ Phimoe flash attention module. This module inherits from `PhimoeAttention` 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 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, position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None, ): bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states) key_states = 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 = position_embeddings query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads if n_kv_heads < n_heads key_states = repeat_kv(key_states, self.num_key_value_groups) value_states = repeat_kv(value_states, self.num_key_value_groups) dropout_rate = 0.0 if not self.training else self.attention_dropout # 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 float16 just to be sure everything works as expected. input_dtype = query_states.dtype device_type = query_states.device.type if query_states.device.type != "mps" else "cpu" if input_dtype == torch.float32: if torch.is_autocast_enabled(): target_dtype = ( torch.get_autocast_dtype(device_type) if hasattr(torch, "get_autocast_dtype") else 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) # Reashape to the expected shape for Flash Attention query_states = query_states.transpose(1, 2) key_states = key_states.transpose(1, 2) value_states = value_states.transpose(1, 2) attn_output = _flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, position_ids=position_ids, dropout=dropout_rate, sliding_window=getattr(self.config, "sliding_window", None), 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 class PhimoeSdpaAttention(PhimoeAttention): """ Phimoe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from `PhimoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to SDPA API. """ # Adapted from PhimoeAttention.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, position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = 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( "PhimoeModel is using PhimoeSdpaAttention, 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, position_embeddings=position_embeddings, ) bsz, q_len, _ = hidden_states.size() query_states = self.q_proj(hidden_states) key_states = 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 = position_embeddings query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) if past_key_value is not None: cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models 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: # no matter the length, we just slice it causal_mask = attention_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 attention_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. # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. 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 PHIMOE_ATTENTION_CLASSES = { "eager": PhimoeAttention, "flash_attention_2": PhimoeFlashAttention2, "sdpa": PhimoeSdpaAttention, } # Copied from transformers.models.mixtral.modeling_mixtral.MixtralBlockSparseTop2MLP with Mixtral->Phimoe class PhimoeBlockSparseTop2MLP(nn.Module): def __init__(self, config: PhimoeConfig): super().__init__() self.ffn_dim = config.intermediate_size self.hidden_dim = config.hidden_size self.w1 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False) self.w2 = nn.Linear(self.ffn_dim, self.hidden_dim, bias=False) self.w3 = nn.Linear(self.hidden_dim, self.ffn_dim, bias=False) self.act_fn = ACT2FN[config.hidden_act] def forward(self, hidden_states): current_hidden_states = self.act_fn(self.w1(hidden_states)) * self.w3(hidden_states) current_hidden_states = self.w2(current_hidden_states) return current_hidden_states class MultiplierProcessor(torch.autograd.Function): @staticmethod def forward( ctx, scores: torch.Tensor, multiplier: torch.Tensor, selected_experts: torch.Tensor, masked_gates: torch.Tensor, mask_for_one: torch.Tensor, ): """ Forward pass for the custom autograd function. Args: ctx: Context object to save information for backward computation. scores (torch.Tensor): Input scores tensor. multiplier (torch.Tensor): Multiplier tensor. selected_experts (torch.Tensor): Tensor of selected experts. masked_gates (torch.Tensor): Masked gates tensor. mask_for_one (torch.Tensor): Mask for one tensor. Returns: torch.Tensor: Result of the forward pass. """ ctx.save_for_backward(multiplier, selected_experts, masked_gates) return multiplier * mask_for_one @staticmethod def backward( ctx, grad_at_output: torch.Tensor, ): """ Backward pass for the custom autograd function. Args: ctx: Context object with saved tensors from the forward pass. grad_at_output (torch.Tensor): Gradient at the output. Returns: tuple[torch.Tensor, None, None, None, None]: Gradients for the inputs. """ multiplier, selected_experts, masked_gates = ctx.saved_tensors grad_at_output = grad_at_output * multiplier grad_at_scores_expanded = masked_gates * grad_at_output.mul(-1) grad_at_scores_expanded.scatter_add_( dim=-1, index=selected_experts, src=grad_at_output, ) return ( grad_at_scores_expanded, None, None, None, None, ) def sparsemixer(scores, jitter_eps, training, top_k=2): """ Sparse mixer function to select top-k experts and compute multipliers. Based on the paper: https://huggingface.co/papers/2409.12136 We first replace the TopK(ยท) function as random sampling of discrete variables in model training. Then, following Liu et al. (2023a) and Liu et al. (2023b), we apply Heun's third order method to approximate the expert routing gradient and construct a modified back-propagation to give a mathematically sound gradient estimation for expert routing. Args: scores (torch.Tensor): Input scores tensor. jitter_eps (float): Jitter epsilon for numerical stability. training (bool): Flag indicating if the model is in training mode. top_k (int): Number of top experts to select. Returns: tuple[torch.Tensor, torch.Tensor]: Multiplier and selected experts tensors. """ if top_k != 2: raise ValueError("top_k must be equal to 2") # first expert with torch.no_grad(): # Compute mask for sparsity mask_logits_threshold, max_ind = scores.max(dim=-1, keepdim=True) factor = scores.abs().clamp(min=mask_logits_threshold) mask_logits_threshold = ((mask_logits_threshold - scores) / factor) > (2 * jitter_eps) # Apply mask masked_gates = scores.masked_fill(mask_logits_threshold, float("-inf")) if training: selected_experts = ( ( masked_gates - torch.empty_like(masked_gates, memory_format=torch.legacy_contiguous_format).exponential_().log() ) .max(dim=-1)[1] .unsqueeze(-1) ) # Gumbel sampling, more robust than the multinomial method else: selected_experts = max_ind # Compute scores for gradients masked_gates = torch.softmax(masked_gates, dim=-1) multiplier_o = masked_gates.gather(dim=-1, index=selected_experts) if training: # Compute midpoint mask max_scores, max_ind = masked_gates.max(dim=-1, keepdim=True) mask_for_one = torch.logical_or( selected_experts == max_ind, torch.rand_like(max_scores) > 0.75, # Heun's third-order method ) # 1 -> 1.0 & 0 -> 1./3: lambda x: (x + 0.5) / 1.5 mask_for_one = torch.add(0.3333, mask_for_one, alpha=0.6667).type_as(masked_gates) multiplier = MultiplierProcessor.apply( scores, multiplier_o, selected_experts, masked_gates, mask_for_one, ) else: multiplier = multiplier_o # Masked out first expert masked_scores = torch.scatter( scores, -1, selected_experts, float("-inf"), ) with torch.no_grad(): # Compute mask for sparsity mask_logits_threshold, max_ind = masked_scores.max(dim=-1, keepdim=True) factor = scores.abs().clamp(min=mask_logits_threshold) mask_logits_threshold = ((mask_logits_threshold - scores) / factor) > (2 * jitter_eps) # Apply mask masked_gates_top2 = masked_scores.masked_fill(mask_logits_threshold, float("-inf")) if training: selected_experts_top2 = ( ( masked_gates_top2 - torch.empty_like(masked_gates_top2, memory_format=torch.legacy_contiguous_format) .exponential_() .log() ) .max(dim=-1)[1] .unsqueeze(-1) ) # Gumbel sampling, more robust than the multinomial method else: selected_experts_top2 = max_ind # Compute scores for gradients masked_gates_top2 = torch.softmax(masked_gates_top2, dim=-1) multiplier_top2_o = masked_gates_top2.gather(dim=-1, index=selected_experts_top2) if training: # Compute midpoint mask max_scores, max_ind = masked_gates_top2.max(dim=-1, keepdim=True) mask_for_one_top2 = torch.logical_or( selected_experts_top2 == max_ind, torch.rand_like(max_scores).uniform_() > 0.75, # Heun's third-order method ) # 1 -> 1.0 & 0 -> 1./3: lambda x: (x + 0.5) / 1.5 mask_for_one_top2 = torch.add(0.3333, mask_for_one_top2, alpha=0.6667).type_as(masked_gates_top2) multiplier_top2 = MultiplierProcessor.apply( scores, multiplier_top2_o, selected_experts_top2, masked_gates_top2, mask_for_one_top2, ) else: multiplier_top2 = multiplier_top2_o multiplier = torch.concat((multiplier, multiplier_top2), dim=-1) selected_experts = torch.concat((selected_experts, selected_experts_top2), dim=-1) return ( multiplier, selected_experts, ) class PhimoeSparseMoeBlock(nn.Module): """ This implementation is strictly equivalent to standard MoE with full capacity (no dropped tokens). It's faster since it formulates MoE operations in terms of block-sparse operations to accommodate imbalanced assignments of tokens to experts, whereas standard MoE either (1) drop tokens at the cost of reduced performance or (2) set capacity factor to number of experts and thus waste computation and memory on padding. """ def __init__(self, config): super().__init__() self.hidden_dim = config.hidden_size self.ffn_dim = config.intermediate_size self.num_experts = config.num_local_experts self.top_k = config.num_experts_per_tok # gating self.gate = nn.Linear(self.hidden_dim, self.num_experts, bias=False) self.experts = nn.ModuleList([PhimoeBlockSparseTop2MLP(config) for _ in range(self.num_experts)]) # Jitter parameters self.router_jitter_noise = config.router_jitter_noise self.input_jitter_noise = config.input_jitter_noise def forward(self, hidden_states: torch.Tensor) -> torch.Tensor: """ """ batch_size, sequence_length, hidden_dim = hidden_states.shape if self.training and self.input_jitter_noise > 0: hidden_states *= torch.empty_like(hidden_states).uniform_( 1.0 - self.input_jitter_noise, 1.0 + self.input_jitter_noise ) hidden_states = hidden_states.view(-1, hidden_dim) router_logits = self.gate(hidden_states) routing_weights, selected_experts = sparsemixer( router_logits, jitter_eps=self.router_jitter_noise, training=self.training, ) final_hidden_states = torch.zeros( (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device ) # One hot encode the selected experts to create an expert mask # this will be used to easily index which expert is going to be sollicitated expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0) # Loop over all available experts in the model and perform the computation on each expert for expert_idx in range(self.num_experts): expert_layer = self.experts[expert_idx] idx, top_x = torch.where(expert_mask[expert_idx]) if top_x.shape[0] == 0: continue # Index the correct hidden states and compute the expert hidden state for # the current expert. We need to make sure to multiply the output hidden # states by `routing_weights` on the corresponding tokens (top-1 and top-2) current_state = hidden_states[None, top_x].reshape(-1, hidden_dim) current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None] # However `index_add_` only support torch tensors for indexing so we'll use # the `top_x` tensor here. final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype)) final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim) return final_hidden_states, router_logits class PhimoeDecoderLayer(GradientCheckpointingLayer): def __init__(self, config: PhimoeConfig, layer_idx: int): super().__init__() self.hidden_size = config.hidden_size self.self_attn = PHIMOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx) self.block_sparse_moe = PhimoeSparseMoeBlock(config) self.input_layernorm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True) self.post_attention_layernorm = nn.LayerNorm( config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True ) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[tuple[torch.Tensor]] = None, output_attentions: Optional[bool] = False, output_router_logits: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, position_embeddings: Optional[tuple[torch.Tensor, torch.Tensor]] = None, **kwargs, ) -> tuple[torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]]: """ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, sequence_length)` where padding elements are indicated by 0. past_key_value (`Tuple(torch.FloatTensor)`, *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 logits of all the routers. They are useful for computing the router loss, and should not be returned during inference. 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` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. kwargs (`dict`, *optional*): Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code into the model """ residual = hidden_states hidden_states = self.input_layernorm(hidden_states) # Self Attention hidden_states, self_attn_weights = 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, position_embeddings=position_embeddings, ) hidden_states = residual + hidden_states # Fully Connected residual = hidden_states hidden_states = self.post_attention_layernorm(hidden_states) hidden_states, router_logits = self.block_sparse_moe(hidden_states) hidden_states = residual + hidden_states outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if output_router_logits: outputs += (router_logits,) return outputs @auto_docstring class PhimoePreTrainedModel(PreTrainedModel): config: PhimoeConfig base_model_prefix = "model" supports_gradient_checkpointing = True _no_split_modules = ["PhimoeDecoderLayer"] _skip_keys_device_placement = ["past_key_values"] _supports_flash_attn = True _supports_sdpa = True _can_compile_fullgraph = False # MoE models don't work with torch.compile (`torch.where(condition)` not supported) def _init_weights(self, 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.bias.data.zero_() module.weight.data.fill_(1.0) @auto_docstring class PhimoeModel(PhimoePreTrainedModel): """ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`PhimoeDecoderLayer`] Args: config: PhimoeConfig """ def __init__(self, config: PhimoeConfig): super().__init__(config) self.padding_idx = config.pad_token_id self.vocab_size = config.vocab_size self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) self.layers = nn.ModuleList( [PhimoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] ) self._attn_implementation = config._attn_implementation self.norm = nn.LayerNorm(config.hidden_size, eps=config.rms_norm_eps, elementwise_affine=True) self.rotary_emb = PhimoeRotaryEmbedding(config=config) self.gradient_checkpointing = False # Initialize weights and apply final processing self.post_init() @can_return_tuple @auto_docstring 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.FloatTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, output_router_logits: Optional[bool] = None, cache_position: Optional[torch.LongTensor] = None, ) -> MoeModelOutputWithPast: output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_router_logits = ( output_router_logits if output_router_logits is not None else self.config.output_router_logits ) output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) use_cache = use_cache if use_cache is not None else self.config.use_cache 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: if use_cache: logger.warning_once( "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." ) use_cache = False # TODO (joao): remove this exception in v4.56 -- it exists for users that try to pass a legacy cache if not isinstance(past_key_values, (type(None), Cache)): raise ValueError("The `past_key_values` should be either a `Cache` object or `None`.") if use_cache and past_key_values is None: past_key_values = DynamicCache() if inputs_embeds is None: inputs_embeds = self.embed_tokens(input_ids) if cache_position is None: past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 cache_position = torch.arange( 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) causal_mask = self._update_causal_mask( attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions ) hidden_states = inputs_embeds position_embeddings = self.rotary_emb(hidden_states, seq_len=cache_position[-1] + 1) # 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 for decoder_layer in self.layers: if output_hidden_states: all_hidden_states += (hidden_states,) layer_outputs = decoder_layer( 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, position_embeddings=position_embeddings, ) hidden_states = layer_outputs[0] if output_attentions: all_self_attns += (layer_outputs[1],) if output_router_logits: all_router_logits += (layer_outputs[-1],) hidden_states = self.norm(hidden_states) # add hidden states from the last decoder layer if output_hidden_states: all_hidden_states += (hidden_states,) return MoeModelOutputWithPast( last_hidden_state=hidden_states, past_key_values=past_key_values, hidden_states=all_hidden_states, attentions=all_self_attns, router_logits=all_router_logits, ) def _update_causal_mask( self, attention_mask: Union[torch.Tensor, "BlockMask"], input_tensor: torch.Tensor, cache_position: torch.Tensor, past_key_values: Cache, output_attentions: bool = False, ): if self.config._attn_implementation == "flash_attention_2": if attention_mask is not None and past_key_values is not None: is_padding_right = attention_mask[:, -1].sum().item() != input_tensor.size()[0] if is_padding_right: raise ValueError( "You are attempting to perform batched generation with padding_side='right'" " this may lead to unexpected behaviour for Flash Attention version of Phimoe. Make sure to " " call `tokenizer.padding_side = 'left'` before tokenizing the input. " ) if attention_mask is not None and 0.0 in attention_mask: return attention_mask return None if self.config._attn_implementation == "flex_attention": if isinstance(attention_mask, torch.Tensor): attention_mask = make_flex_block_causal_mask(attention_mask) return attention_mask # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail # to infer the attention mask. past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 using_static_cache = isinstance(past_key_values, StaticCache) using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache) # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward if ( self.config._attn_implementation == "sdpa" and not (using_static_cache or using_sliding_window_cache) and not output_attentions ): if AttentionMaskConverter._ignore_causal_mask_sdpa( attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens, sliding_window=self.config.sliding_window, is_training=self.training, ): return None dtype = input_tensor.dtype min_dtype = torch.finfo(dtype).min sequence_length = input_tensor.shape[1] # SlidingWindowCache or StaticCache if using_sliding_window_cache or using_static_cache: target_length = past_key_values.get_max_cache_shape() # DynamicCache or no cache else: target_length = ( attention_mask.shape[-1] if isinstance(attention_mask, torch.Tensor) else past_seen_tokens + sequence_length + 1 ) # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position( attention_mask, sequence_length=sequence_length, target_length=target_length, dtype=dtype, cache_position=cache_position, batch_size=input_tensor.shape[0], config=self.config, past_key_values=past_key_values, ) if ( self.config._attn_implementation == "sdpa" and attention_mask is not None and attention_mask.device.type in ["cuda", "xpu", "npu"] and not output_attentions ): # 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 @staticmethod def _prepare_4d_causal_attention_mask_with_cache_position( attention_mask: torch.Tensor, sequence_length: int, target_length: int, dtype: torch.dtype, cache_position: torch.Tensor, batch_size: int, config: PhimoeConfig, past_key_values: Cache, ): """ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. config (`PhimoeConfig`): The model's configuration class past_key_values (`Cache`): The cache class that is being used currently to generate """ if attention_mask is not None and attention_mask.dim() == 4: # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. causal_mask = attention_mask else: min_dtype = torch.finfo(dtype).min causal_mask = torch.full( (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=cache_position.device ) diagonal_attend_mask = torch.arange(target_length, device=cache_position.device) > cache_position.reshape( -1, 1 ) text_config = config.get_text_config() if getattr(text_config, "use_sliding_window", True) and text_config.sliding_window is not None: # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also # the check is needed to verify is current checkpoint was trained with sliding window or not if not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length: sliding_attend_mask = torch.arange(target_length, device=cache_position.device) <= ( cache_position.reshape(-1, 1) - text_config.sliding_window ) diagonal_attend_mask.bitwise_or_(sliding_attend_mask) causal_mask *= diagonal_attend_mask causal_mask = causal_mask[None, None, :, :].expand(batch_size, 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.shape[-1] > target_length: attention_mask = attention_mask[:, :target_length] mask_length = attention_mask.shape[-1] padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :].to( causal_mask.device ) padding_mask = padding_mask == 0 causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( padding_mask, min_dtype ) return causal_mask class PhimoeForCausalLM(PhimoePreTrainedModel, GenerationMixin): _tied_weights_keys = ["lm_head.weight"] def __init__(self, config): super().__init__(config) self.model = PhimoeModel(config) self.vocab_size = config.vocab_size self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=self.config.lm_head_bias) self.router_aux_loss_coef = config.router_aux_loss_coef self.num_experts = config.num_local_experts self.num_experts_per_tok = config.num_experts_per_tok # Initialize weights and apply final processing self.post_init() # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.set_decoder def set_decoder(self, decoder): self.model = decoder # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.get_decoder def get_decoder(self): return self.model @can_return_tuple @auto_docstring 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.FloatTensor] = 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, cache_position: Optional[torch.LongTensor] = None, logits_to_keep: Union[int, torch.Tensor] = 0, **kwargs, ) -> MoeCausalLMOutputWithPast: r""" labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. Example: ```python >>> from transformers import AutoTokenizer, PhimoeForCausalLM >>> model = PhimoeForCausalLM.from_pretrained("microsoft/Phi-3.5-MoE-instruct") >>> tokenizer = AutoTokenizer.from_pretrained("microsoft/Phi-3.5-MoE-instruct") >>> 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." ```""" if ( use_cache and self.config.rope_scaling and cache_position is not None and cache_position[0] == self.config.original_max_position_embeddings ): logger.warning( f"If you are not using the generate method, you may encounter nonsensical outputs after the {self.config.original_max_position_embeddings}th token, as the KV cache needs to be recomputed." ) output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions output_router_logits = ( output_router_logits if output_router_logits is not None else self.config.output_router_logits ) output_hidden_states = ( output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states ) # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) outputs: MoeModelOutputWithPast = self.model( 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, cache_position=cache_position, ) hidden_states = outputs.last_hidden_state # Only compute necessary logits, and do not upcast them to float if we are not computing the loss slice_indices = slice(-logits_to_keep, None) if isinstance(logits_to_keep, int) else logits_to_keep logits = self.lm_head(hidden_states[:, slice_indices, :]) loss = None if labels is not None: loss = self.loss_function(logits, labels, self.vocab_size, **kwargs) aux_loss = None if output_router_logits: aux_loss = load_balancing_loss_func( outputs.router_logits, self.num_experts, self.num_experts_per_tok, attention_mask, ) if labels is not None: loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device 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, ) # Copied from transformers.models.phi3.modeling_phi3.Phi3ForCausalLM.prepare_inputs_for_generation def prepare_inputs_for_generation( self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, cache_position=None, position_ids=None, use_cache=True, logits_to_keep=None, **kwargs, ): # Overwritten -- this model may need to switch between short and long rope, invalidating the cache in the # process # When the first time input length reached long and short factor switching point, enforce re-compute cache # It will cause downside of slower at this single token position, however, better than current failure. if ( past_key_values and self.config.rope_scaling and input_ids.shape[1] >= self.config.original_max_position_embeddings + 1 ): past_length = cache_position[0] if past_length <= self.config.original_max_position_embeddings: past_key_values = None model_inputs = super().prepare_inputs_for_generation( input_ids=input_ids, past_key_values=past_key_values, attention_mask=attention_mask, inputs_embeds=inputs_embeds, cache_position=cache_position, position_ids=position_ids, use_cache=use_cache, logits_to_keep=logits_to_keep, **kwargs, ) return model_inputs class PhimoeForSequenceClassification(GenericForSequenceClassification, PhimoePreTrainedModel): ... __all__ = [ "PhimoePreTrainedModel", "PhimoeModel", "PhimoeForCausalLM", "PhimoeForSequenceClassification", ]