# coding=utf-8 # Copyright 2024 JetMoe AI 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 JetMoe model.""" import math from typing import Optional, Union import torch import torch.utils.checkpoint from torch import nn from torch.nn import functional as F from ...activations import ACT2FN from ...cache_utils import Cache, DynamicCache from ...generation import GenerationMixin from ...modeling_attn_mask_utils import AttentionMaskConverter from ...modeling_flash_attention_utils import flash_attn_supports_top_left_mask, is_flash_attn_available from ...modeling_layers import ( GenericForSequenceClassification, GradientCheckpointingLayer, ) from ...modeling_outputs import MoeCausalLMOutputWithPast, MoeModelOutputWithPast from ...modeling_rope_utils import ROPE_INIT_FUNCTIONS, dynamic_rope_update from ...modeling_utils import PreTrainedModel from ...utils import auto_docstring, can_return_tuple, is_torch_flex_attn_available, logging from .configuration_jetmoe import JetMoeConfig if is_torch_flex_attn_available(): from torch.nn.attention.flex_attention import BlockMask from ...integrations.flex_attention import make_flex_block_causal_mask if is_flash_attn_available(): from ...modeling_flash_attention_utils import _flash_attention_forward 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 JetMoeParallelExperts(nn.Module): def __init__(self, num_experts: int, input_size: int, output_size: int) -> None: """ Initialize the JetMoeParallelExperts module. The experts weights are stored in [num_experts, output_size, input_size] format. Such that it's compatible with many MoE libraries, such as [Megablock](https://github.com/databricks/megablocks) and [ScatterMoE](https://github.com/shawntan/scattermoe), as well as the [MoE kernel](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/fused_moe.py) used in vllm. Args: num_experts (int): Number of experts. input_size (int): Size of the input. output_size (int): Size of the output. """ super().__init__() self.weight = nn.Parameter(torch.empty(num_experts, output_size, input_size)) self.num_experts = num_experts self.input_size = input_size self.output_size = output_size def forward(self, inputs, expert_size): """ Forward pass of the JetMoeParallelExperts module. Args: inputs (Tensor): Input tensor. expert_size: Expert size information. Returns: Tensor: Output tensor. """ input_list = inputs.split(expert_size, dim=0) output_list = [] for i in range(self.num_experts): output_list.append(F.linear(input_list[i], self.weight[i])) results = torch.cat(output_list, dim=0) return results class JetMoeTopKGating(nn.Module): def __init__(self, input_size: int, num_experts: int, top_k: int): """ Initialize the top-k gating mechanism. Args: input_size (`int`): Size of the input. num_experts (`int`): Number of experts. top_k (`int`): Number of top experts to select. """ super().__init__() self.num_experts = num_experts self.input_size = input_size self.top_k = top_k self.layer = nn.Linear(input_size, num_experts, bias=False) def forward(self, hidden_states): # compute the top_k routing decision logits = self.layer(hidden_states).float() # [batch_size x seq_len, num_experts] top_k_logits, top_k_indices = logits.topk(self.top_k, dim=1) # [num_tokens, top_k] top_k_gates = torch.softmax(top_k_logits, dim=1).type_as(hidden_states) # [num_tokens, top_k] # compute number of input given to each expert zeros = torch.zeros( [top_k_gates.size(0), self.num_experts], dtype=top_k_gates.dtype, device=top_k_gates.device ) # [num_tokens, num_experts] gates = zeros.scatter(1, top_k_indices, 1) # [num_tokens, num_experts] expert_size = gates.long().sum(0) # [num_experts,] # (This cause torch.compile to fail with `torch._dynamo.exc.Unsupported: Backend compiler failed with a fake tensor exception at`) # (and `DataDependentOutputException`) expert_size = expert_size.tolist() # sort and group input tokens according to expert assignment top_k_experts = top_k_indices.flatten() # [num_tokens * top_k] _, index_sorted_experts = top_k_experts.sort(0) # [num_tokens * top_k] batch_index = index_sorted_experts.div(self.top_k, rounding_mode="trunc") # [num_tokens * top_k] # gather the gate values for grouped input tokens top_k_gates = top_k_gates.flatten() # [num_tokens * top_k] batch_gates = top_k_gates[index_sorted_experts] # [num_tokens * top_k] return index_sorted_experts, batch_index, batch_gates, expert_size, logits class JetMoeMoE(nn.Module): """ A Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as experts. Args: config: Configuration object with model hyperparameters. """ def __init__(self, config: JetMoeConfig): super().__init__() self.input_size = config.hidden_size self.hidden_size = config.intermediate_size self.activation = ACT2FN[config.activation_function] self.bias = torch.nn.Parameter(torch.empty(self.input_size)) self.input_linear = JetMoeParallelExperts(config.num_local_experts, self.input_size, self.hidden_size * 2) self.output_linear = JetMoeParallelExperts(config.num_local_experts, self.hidden_size, self.input_size) self.router = JetMoeTopKGating( input_size=self.input_size, num_experts=config.num_local_experts, top_k=config.num_experts_per_tok, ) def forward(self, layer_input): """ Forward pass of the mixture of experts layer. Args: layer_input (Tensor): Input tensor. Returns: Tensor: Output tensor. Tensor: Router logits. """ bsz, length, emb_size = layer_input.size() layer_input = layer_input.reshape(-1, emb_size) _, batch_index, batch_gates, expert_size, router_logits = self.router(layer_input) expert_inputs = layer_input[batch_index] hidden_states = self.input_linear(expert_inputs, expert_size) chunked_hidden_states = hidden_states.chunk(2, dim=-1) hidden_states = self.activation(chunked_hidden_states[0]) * chunked_hidden_states[1] expert_outputs = self.output_linear(hidden_states, expert_size) expert_outputs = expert_outputs * batch_gates[:, None] zeros = torch.zeros((bsz * length, self.input_size), dtype=expert_outputs.dtype, device=expert_outputs.device) layer_output = zeros.index_add(0, batch_index, expert_outputs) layer_output = layer_output.view(bsz, length, self.input_size) layer_output = layer_output + self.bias return layer_output, router_logits class JetMoeMoA(nn.Module): """ A Sparsely gated mixture of attention layer with pairs of query- and output-projections as experts. Args: config: Configuration object with model hyperparameters. """ def __init__(self, config: JetMoeConfig): super().__init__() self.num_experts = config.num_local_experts self.input_size = config.hidden_size self.hidden_size = config.kv_channels * config.num_key_value_heads self.top_k = config.num_experts_per_tok self.bias = torch.nn.Parameter(torch.empty(self.input_size)) self.input_linear = JetMoeParallelExperts(self.num_experts, self.input_size, self.hidden_size) self.output_linear = JetMoeParallelExperts(self.num_experts, self.hidden_size, self.input_size) self.router = JetMoeTopKGating( input_size=self.input_size, num_experts=self.num_experts, top_k=self.top_k, ) def map(self, layer_input): """ Map inputs to attention experts according to routing decision and compute query projection inside each experts. """ # Compute gating topology bsz, length, emb_size = layer_input.size() layer_input = layer_input.reshape(-1, emb_size) # [bsz * length, emb_size] index_sorted_experts, batch_index, batch_gates, expert_size, router_logits = self.router(layer_input) topo_info = (index_sorted_experts, batch_index, batch_gates, expert_size) # Group inputs according to topology and compute query projection expert_inputs = layer_input[batch_index] # [bsz * length * top_k, emb_size] expert_outputs = self.input_linear(expert_inputs, expert_size) # [bsz * length * top_k, hidden_size] # Ungroup queries back to original order zeros = torch.zeros( (bsz * length * self.top_k, self.hidden_size), dtype=expert_outputs.dtype, device=expert_outputs.device ) layer_output = zeros.index_add(0, index_sorted_experts, expert_outputs) layer_output = layer_output.view(bsz, length, self.top_k, -1) # [bsz, length, top_k, hidden_size] return layer_output, router_logits, topo_info def reduce(self, layer_input, topo_info): """ Compute output projection inside each attention experts and merge the outputs of different experts. """ bsz, length, k, hidden_size = layer_input.size() layer_input = layer_input.reshape(-1, hidden_size) # [bsz * length * k, hidden_size] index_sorted_experts, batch_index, batch_gates, expert_size = topo_info # Group inputs according to topology and compute output projection expert_inputs = layer_input[index_sorted_experts] # [bsz * length * top_k, hidden_size] expert_outputs = self.output_linear(expert_inputs, expert_size) # [bsz * length * top_k, emb_size] # Apply gates to attention expert outputs expert_outputs = expert_outputs * batch_gates[:, None] # Ungroup and merge outputs to original order zeros = torch.zeros((bsz * length, self.input_size), dtype=expert_outputs.dtype, device=expert_outputs.device) layer_output = zeros.index_add(0, batch_index, expert_outputs) layer_output = layer_output.view(bsz, length, self.input_size) layer_output = layer_output + self.bias return layer_output def forward(self, layer_input): raise NotImplementedError("This module doesn't support call and forward.") # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->JetMoe class JetMoeRMSNorm(nn.Module): def __init__(self, hidden_size, eps=1e-6): """ JetMoeRMSNorm is equivalent to T5LayerNorm """ super().__init__() self.weight = nn.Parameter(torch.ones(hidden_size)) self.variance_epsilon = eps def forward(self, hidden_states): input_dtype = hidden_states.dtype hidden_states = hidden_states.to(torch.float32) variance = hidden_states.pow(2).mean(-1, keepdim=True) hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) return self.weight * hidden_states.to(input_dtype) def extra_repr(self): return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" # Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with Gemma->JetMoe class JetMoeRotaryEmbedding(nn.Module): def __init__(self, config: JetMoeConfig, device=None): super().__init__() # BC: "rope_type" was originally "type" if hasattr(config, "rope_scaling") and isinstance(config.rope_scaling, dict): self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type")) else: self.rope_type = "default" self.max_seq_len_cached = config.max_position_embeddings self.original_max_seq_len = config.max_position_embeddings self.config = config self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device) self.register_buffer("inv_freq", inv_freq, persistent=False) self.original_inv_freq = self.inv_freq @torch.no_grad() @dynamic_rope_update # power user: used with advanced RoPE types (e.g. dynamic rope) def forward(self, x, position_ids): inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device) position_ids_expanded = position_ids[:, None, :].float() device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu" with torch.autocast(device_type=device_type, enabled=False): # Force float32 freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) emb = torch.cat((freqs, freqs), dim=-1) cos = emb.cos() * self.attention_scaling sin = emb.sin() * self.attention_scaling return cos.to(dtype=x.dtype), sin.to(dtype=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) # Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, 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`, *optional*): Deprecated and unused. 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.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 class JetMoeAttention(nn.Module): """ Multi-headed attention from 'Attention Is All You Need' paper. """ def __init__(self, config: JetMoeConfig, layer_idx: Optional[int] = None): """ Initialize the JetMoeAttention module. Args: config: Configuration object with model hyperparameters. layer_idx: Index of the layer in the model. """ super().__init__() self.config = config self.layer_idx = layer_idx self.is_causal = True 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.top_k = config.num_experts_per_tok self.attention_dropout = config.attention_dropout self.kv_projection_size = config.kv_channels * config.num_key_value_heads self.num_key_value_heads = config.num_key_value_heads self.num_heads = config.num_attention_heads self.head_dim = config.kv_channels self.experts = JetMoeMoA(config) self.kv_proj = torch.nn.Linear(config.hidden_size, self.kv_projection_size * 2, bias=False) self.rotary_emb = JetMoeRotaryEmbedding(config) def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]]]: bsz, q_len, _ = hidden_states.size() query_states, router_logits, topo_info = self.experts.map(hidden_states) key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1) query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) cos, sin = self.rotary_emb(value_states, position_ids) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads for top-k attention experts key_states = key_states.repeat(1, self.top_k, 1, 1) value_states = value_states.repeat(1, self.top_k, 1, 1) 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.top_k, self.kv_projection_size) attn_output = self.experts.reduce(attn_output, topo_info) attn_output = attn_output.view(bsz, q_len, -1) if not output_attentions: attn_weights = None return attn_output, attn_weights, router_logits class JetMoeSdpaAttention(JetMoeAttention): """ JetMoe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from `JetMoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to SDPA API. """ # Adapted from JetMoeAttention.forward def forward( self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, output_attentions: bool = False, use_cache: bool = False, cache_position: Optional[torch.LongTensor] = None, ) -> tuple[torch.Tensor, Optional[torch.Tensor], Optional[tuple[torch.Tensor]], Optional[torch.Tensor]]: if output_attentions: # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. logger.warning_once( "JetMoeModel is using JetMoeSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' ) return super().forward( hidden_states=hidden_states, attention_mask=attention_mask, position_ids=position_ids, past_key_value=past_key_value, output_attentions=output_attentions, use_cache=use_cache, cache_position=cache_position, ) bsz, q_len, _ = hidden_states.size() query_states, router_logits, topo_info = self.experts.map(hidden_states) key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1) query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) cos, sin = self.rotary_emb(value_states, position_ids) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads for top-k attention experts key_states = key_states.repeat(1, self.top_k, 1, 1) value_states = value_states.repeat(1, self.top_k, 1, 1) causal_mask = attention_mask if attention_mask is not None: causal_mask = causal_mask[:, :, :, : key_states.shape[-2]] # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, # Reference: https://github.com/pytorch/pytorch/issues/112577. if query_states.device.type == "cuda" and causal_mask is not None: query_states = query_states.contiguous() key_states = key_states.contiguous() value_states = value_states.contiguous() # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. is_causal = True if causal_mask is None and q_len > 1 else False attn_output = torch.nn.functional.scaled_dot_product_attention( query_states, key_states, value_states, attn_mask=causal_mask, dropout_p=self.attention_dropout if self.training else 0.0, is_causal=is_causal, ) attn_output = attn_output.transpose(1, 2).contiguous() attn_output = attn_output.reshape(bsz, q_len, self.top_k, self.kv_projection_size) attn_output = self.experts.reduce(attn_output, topo_info) attn_output = attn_output.view(bsz, q_len, -1) return attn_output, None, router_logits class JetMoeFlashAttention2(JetMoeAttention): def __init__(self, *args, **kwargs): super().__init__(*args, **kwargs) # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). self._flash_attn_uses_top_left_mask = flash_attn_supports_top_left_mask() def forward( self, hidden_states: Optional[torch.FloatTensor], attention_mask: Optional[torch.FloatTensor] = None, position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[Cache] = None, use_cache: Optional[bool] = False, output_attentions: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, ) -> Union[ tuple[torch.Tensor, tuple[torch.Tensor]], Optional[tuple[torch.Tensor, tuple[torch.Tensor], tuple[torch.Tensor, ...]]], ]: """ Forward pass of the JetMoeAttention module. Args: hidden_states (Optional[torch.FloatTensor]): Input hidden states. attention_mask (Optional[torch.FloatTensor]): Attention mask. layer_past (Optional[tuple[torch.Tensor]]): Past layer state. use_cache (Optional[bool]): Whether to use cached states. output_attentions (Optional[bool]): Whether to output attention weights. cache_position (Optional[torch.LongTensor]): Position of the cache. Returns: Union[tuple[torch.Tensor, tuple[torch.Tensor]], Optional[tuple[...]]]: Tuple containing outputs. """ output_attentions = False bsz, q_len, hidden_size = hidden_states.size() # calculate query, key, values query_states, router_logits, topo_info = self.experts.map(hidden_states) key_states, value_states = self.kv_proj(hidden_states).chunk(2, dim=-1) query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) cos, sin = self.rotary_emb(value_states, position_ids) query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) if past_key_value is not None: # sin and cos are specific to RoPE models; cache_position needed for the static cache cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) # repeat k/v heads for top-k attention experts key_states = key_states.repeat(1, self.top_k, 1, 1) value_states = value_states.repeat(1, self.top_k, 1, 1) # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache # to be able to avoid many of these transpose/reshape/view. query_states = query_states.transpose(1, 2) key_states = key_states.transpose(1, 2) value_states = value_states.transpose(1, 2) dropout_rate = self.attention_dropout if self.training else 0.0 # In PEFT, usually we cast the layer norms in float32 for training stability reasons # therefore the input hidden states gets silently casted in float32. Hence, we need # cast them back in the correct dtype just to be sure everything works as expected. # This might slowdown training & inference so it is recommended to not cast the LayerNorms # in fp32. (LlamaRMSNorm handles it correctly) 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.kv_proj.weight.dtype logger.warning_once( f"The input hidden states seems to be silently casted in float32, this might be related to" f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" f" {target_dtype}." ) query_states = query_states.to(target_dtype) key_states = key_states.to(target_dtype) value_states = value_states.to(target_dtype) attn_output = _flash_attention_forward( query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate, use_top_left_mask=self._flash_attn_uses_top_left_mask, is_causal=self.is_causal, ).to(input_dtype) # output projection attn_output = attn_output.reshape(bsz, q_len, self.top_k, self.kv_projection_size) attn_output = self.experts.reduce(attn_output, topo_info) attn_output = attn_output.view(bsz, q_len, hidden_size) # re-assemble all head outputs side by side if not output_attentions: attn_weights = None return attn_output, attn_weights, router_logits JETMOE_ATTENTION_CLASSES = { "eager": JetMoeAttention, "flash_attention_2": JetMoeFlashAttention2, "sdpa": JetMoeSdpaAttention, } class JetMoeBlock(GradientCheckpointingLayer): def __init__(self, config: JetMoeConfig, layer_idx: Optional[int] = None): """ Initialize the JetMoeBlock module. Args: config: Configuration object with model hyperparameters. """ super().__init__() self.input_layernorm = JetMoeRMSNorm(config.hidden_size) self.self_attention = JETMOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx) self.post_attention_layernorm = JetMoeRMSNorm(config.hidden_size) self.mlp = JetMoeMoE(config) def forward( self, hidden_states: Optional[torch.FloatTensor], position_ids: Optional[torch.LongTensor] = None, past_key_value: Optional[tuple[torch.Tensor]] = None, attention_mask: Optional[torch.FloatTensor] = None, output_attentions: Optional[bool] = False, output_router_logits: Optional[bool] = False, use_cache: Optional[bool] = False, cache_position: Optional[torch.LongTensor] = None, ) -> Union[tuple[torch.Tensor], Optional[tuple[torch.Tensor, tuple[torch.FloatTensor, ...]]]]: # Self Attention attn_output, self_attn_weights, attn_router_logits = self.self_attention( hidden_states=self.input_layernorm(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, ) hidden_states = hidden_states + attn_output x_mlp, mlp_router_logits = self.mlp(self.post_attention_layernorm(hidden_states)) hidden_states = hidden_states + x_mlp outputs = (hidden_states,) if output_attentions: outputs += (self_attn_weights,) if output_router_logits: outputs += attn_router_logits, mlp_router_logits return outputs @auto_docstring class JetMoePreTrainedModel(PreTrainedModel): config: JetMoeConfig base_model_prefix = "transformer" supports_gradient_checkpointing = False _no_split_modules = ["JetMoeBlock"] _skip_keys_device_placement = ["past_key_values"] _supports_flash_attn = True _supports_sdpa = True def _init_weights(self, module): """Initialize the weights.""" if isinstance(module, (nn.Linear,)): # Slightly different from Mesh Transformer JAX which uses truncated_normal for initialization # cf https://github.com/pytorch/pytorch/pull/5617 module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) if module.bias is not None: module.bias.data.zero_() elif isinstance(module, nn.Embedding): module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) if module.padding_idx is not None: module.weight.data[module.padding_idx].zero_() elif isinstance(module, JetMoeRMSNorm): module.weight.data.fill_(1.0) elif isinstance(module, JetMoeParallelExperts): module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) elif isinstance(module, JetMoeMoA): module.bias.data.zero_() elif isinstance(module, JetMoeMoE): module.bias.data.zero_() @auto_docstring class JetMoeModel(JetMoePreTrainedModel): """ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`JetMoeBlock`] Args: config: JetMoeConfig """ def __init__(self, config: JetMoeConfig): 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([JetMoeBlock(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]) self._attn_implementation = config._attn_implementation self.norm = JetMoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps) 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[Union[Cache, list[torch.FloatTensor]]] = 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_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 if (input_ids is None) ^ (inputs_embeds is not None): raise ValueError("You must specify exactly one of input_ids or inputs_embeds") 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.embed_tokens(input_ids) # 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 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) if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: batch_size = inputs_embeds.shape[0] is_padding_right = attention_mask[:, -1].sum().item() != batch_size 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 JetMoe. Make sure to " " call `tokenizer.padding_side = 'left'` before tokenizing the input. " ) causal_mask = self._update_causal_mask( attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions ) 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 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, ) hidden_states = layer_outputs[0] if output_attentions: all_self_attns += (layer_outputs[1],) if output_router_logits: all_router_logits += (layer_outputs[-2], 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, ) # Copied from transformers.models.gptj.modeling_gptj.GPTJModel._update_causal_mask 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 (attention_mask == 0.0).any(): 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_compilable_cache = past_key_values.is_compileable if past_key_values is not None else False # 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_compilable_cache and not output_attentions: if AttentionMaskConverter._ignore_causal_mask_sdpa( attention_mask, inputs_embeds=input_tensor, past_key_values_length=past_seen_tokens, is_training=self.training, ): return None dtype = input_tensor.dtype sequence_length = input_tensor.shape[1] if using_compilable_cache: target_length = past_key_values.get_max_cache_shape() 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], ) 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 min_dtype = torch.finfo(dtype).min causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) return causal_mask @staticmethod # Copied from transformers.models.gptj.modeling_gptj.GPTJModel._prepare_4d_causal_attention_mask_with_cache_position 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, **kwargs, ): """ 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. """ 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 ) if sequence_length != 1: causal_mask = torch.triu(causal_mask, diagonal=1) causal_mask *= torch.arange(target_length, device=cache_position.device) > cache_position.reshape(-1, 1) 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 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 JetMoeForCausalLM(JetMoePreTrainedModel, GenerationMixin): _tied_weights_keys = ["lm_head.weight"] def __init__(self, config): super().__init__(config) self.model = JetMoeModel(config) self.vocab_size = config.vocab_size self.aux_loss_coef = config.aux_loss_coef self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) self.tie_word_embeddings = config.tie_word_embeddings # 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]`. """ 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 ) # 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, 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, vocab_size=self.config.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.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, ) class JetMoeForSequenceClassification(GenericForSequenceClassification, JetMoePreTrainedModel): ... __all__ = ["JetMoeForCausalLM", "JetMoeModel", "JetMoePreTrainedModel", "JetMoeForSequenceClassification"]