Question Answering
Safetensors
upcycling-qwen2-moe
biology
medical
custom_code
BossRui commited on
Commit
29a1c7b
1 Parent(s): d54f686

Upload modeling_upcycling_qwen2_moe.py with huggingface_hub

Browse files
Files changed (1) hide show
  1. modeling_upcycling_qwen2_moe.py +1673 -0
modeling_upcycling_qwen2_moe.py ADDED
@@ -0,0 +1,1673 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # coding=utf-8
2
+ # Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
3
+ #
4
+ # This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
5
+ # and OPT implementations in this library. It has been modified from its
6
+ # original forms to accommodate minor architectural differences compared
7
+ # to GPT-NeoX and OPT used by the Meta AI team that trained the model.
8
+ #
9
+ # Licensed under the Apache License, Version 2.0 (the "License");
10
+ # you may not use this file except in compliance with the License.
11
+ # You may obtain a copy of the License at
12
+ #
13
+ # http://www.apache.org/licenses/LICENSE-2.0
14
+ #
15
+ # Unless required by applicable law or agreed to in writing, software
16
+ # distributed under the License is distributed on an "AS IS" BASIS,
17
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
18
+ # See the License for the specific language governing permissions and
19
+ # limitations under the License.
20
+ """PyTorch Qwen2MoE model."""
21
+
22
+ import inspect
23
+ import math
24
+ from typing import List, Optional, Tuple, Union
25
+
26
+ import torch
27
+ import torch.nn.functional as F
28
+ import torch.utils.checkpoint
29
+ from torch import nn
30
+ from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
31
+
32
+ from transformers.activations import ACT2FN
33
+ from transformers.cache_utils import Cache, DynamicCache, StaticCache
34
+ from transformers.modeling_attn_mask_utils import (
35
+ AttentionMaskConverter,
36
+ )
37
+ from transformers.modeling_outputs import (
38
+ MoeCausalLMOutputWithPast,
39
+ MoeModelOutputWithPast,
40
+ )
41
+ from transformers.modeling_utils import PreTrainedModel
42
+ from transformers.utils import (
43
+ is_flash_attn_2_available,
44
+ is_flash_attn_greater_or_equal_2_10,
45
+ logging,
46
+ replace_return_docstrings,
47
+ )
48
+ from configuration_upcycling_qwen2_moe import UpcyclingQwen2MoeConfig
49
+ from transformers import AutoModelForCausalLM,AutoConfig,AutoModel
50
+
51
+
52
+ if is_flash_attn_2_available():
53
+ from flash_attn import flash_attn_func, flash_attn_varlen_func
54
+ from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa
55
+
56
+ _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
57
+
58
+ logger = logging.get_logger(__name__)
59
+
60
+ _CHECKPOINT_FOR_DOC = "UpcyclingQwen2MoE"
61
+ _CONFIG_FOR_DOC = "UpcyclingQwen2MoeConfig"
62
+
63
+
64
+ # Copied from transformers.models.mixtral.modeling_mixtral.load_balancing_loss_func
65
+ def load_balancing_loss_func(
66
+ gate_logits: torch.Tensor, num_experts: torch.Tensor = None, top_k=2, attention_mask: Optional[torch.Tensor] = None
67
+ ) -> float:
68
+ r"""
69
+ Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.
70
+
71
+ See Switch Transformer (https://arxiv.org/abs/2101.03961) for more details. This function implements the loss
72
+ function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
73
+ experts is too unbalanced.
74
+
75
+ Args:
76
+ gate_logits (Union[`torch.Tensor`, Tuple[torch.Tensor]):
77
+ Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
78
+ shape [batch_size X sequence_length, num_experts].
79
+ attention_mask (`torch.Tensor`, None):
80
+ The attention_mask used in forward function
81
+ shape [batch_size X sequence_length] if not None.
82
+ num_experts (`int`, *optional*):
83
+ Number of experts
84
+
85
+ Returns:
86
+ The auxiliary loss.
87
+ """
88
+ if gate_logits is None or not isinstance(gate_logits, tuple):
89
+ return 0
90
+
91
+ if isinstance(gate_logits, tuple):
92
+ compute_device = gate_logits[0].device
93
+ concatenated_gate_logits = torch.cat([layer_gate.to(compute_device) for layer_gate in gate_logits], dim=0)
94
+
95
+ routing_weights = torch.nn.functional.softmax(concatenated_gate_logits, dim=-1)
96
+
97
+ _, selected_experts = torch.topk(routing_weights, top_k, dim=-1)
98
+
99
+ expert_mask = torch.nn.functional.one_hot(selected_experts, num_experts)
100
+
101
+ if attention_mask is None:
102
+ # Compute the percentage of tokens routed to each experts
103
+ tokens_per_expert = torch.mean(expert_mask.float(), dim=0)
104
+
105
+ # Compute the average probability of routing to these experts
106
+ router_prob_per_expert = torch.mean(routing_weights, dim=0)
107
+ else:
108
+ batch_size, sequence_length = attention_mask.shape
109
+ num_hidden_layers = concatenated_gate_logits.shape[0] // (batch_size * sequence_length)
110
+
111
+ # Compute the mask that masks all padding tokens as 0 with the same shape of expert_mask
112
+ expert_attention_mask = (
113
+ attention_mask[None, :, :, None, None]
114
+ .expand((num_hidden_layers, batch_size, sequence_length, top_k, num_experts))
115
+ .reshape(-1, top_k, num_experts)
116
+ .to(compute_device)
117
+ )
118
+
119
+ # Compute the percentage of tokens routed to each experts
120
+ tokens_per_expert = torch.sum(expert_mask.float() * expert_attention_mask, dim=0) / torch.sum(
121
+ expert_attention_mask, dim=0
122
+ )
123
+
124
+ # Compute the mask that masks all padding tokens as 0 with the same shape of tokens_per_expert
125
+ router_per_expert_attention_mask = (
126
+ attention_mask[None, :, :, None]
127
+ .expand((num_hidden_layers, batch_size, sequence_length, num_experts))
128
+ .reshape(-1, num_experts)
129
+ .to(compute_device)
130
+ )
131
+
132
+ # Compute the average probability of routing to these experts
133
+ router_prob_per_expert = torch.sum(routing_weights * router_per_expert_attention_mask, dim=0) / torch.sum(
134
+ router_per_expert_attention_mask, dim=0
135
+ )
136
+
137
+ overall_loss = torch.sum(tokens_per_expert * router_prob_per_expert.unsqueeze(0))
138
+ return overall_loss * num_experts
139
+
140
+
141
+ # Copied from transformers.models.llama.modeling_llama._get_unpad_data
142
+ def _get_unpad_data(attention_mask):
143
+ seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
144
+ indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
145
+ max_seqlen_in_batch = seqlens_in_batch.max().item()
146
+ cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
147
+ return (
148
+ indices,
149
+ cu_seqlens,
150
+ max_seqlen_in_batch,
151
+ )
152
+
153
+
154
+ # Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2Moe
155
+ class Qwen2MoeRMSNorm(nn.Module):
156
+ def __init__(self, hidden_size, eps=1e-6):
157
+ """
158
+ Qwen2MoeRMSNorm is equivalent to T5LayerNorm
159
+ """
160
+ super().__init__()
161
+ self.weight = nn.Parameter(torch.ones(hidden_size))
162
+ self.variance_epsilon = eps
163
+
164
+ def forward(self, hidden_states):
165
+ input_dtype = hidden_states.dtype
166
+ hidden_states = hidden_states.to(torch.float32)
167
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
168
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
169
+ return self.weight * hidden_states.to(input_dtype)
170
+
171
+
172
+ # Copied from transformers.models.mixtral.modeling_mixtral.MixtralRotaryEmbedding with Mixtral->Qwen2Moe
173
+ class Qwen2MoeRotaryEmbedding(nn.Module):
174
+ def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
175
+ super().__init__()
176
+
177
+ self.dim = dim
178
+ self.max_position_embeddings = max_position_embeddings
179
+ self.base = base
180
+ inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64).float().to(device) / self.dim))
181
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
182
+
183
+ # Build here to make `torch.jit.trace` work.
184
+ self._set_cos_sin_cache(
185
+ seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
186
+ )
187
+
188
+ def _set_cos_sin_cache(self, seq_len, device, dtype):
189
+ self.max_seq_len_cached = seq_len
190
+ t = torch.arange(self.max_seq_len_cached, device=device, dtype=torch.int64).type_as(self.inv_freq)
191
+
192
+ freqs = torch.outer(t, self.inv_freq)
193
+ # Different from paper, but it uses a different permutation in order to obtain the same calculation
194
+ emb = torch.cat((freqs, freqs), dim=-1)
195
+ self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False)
196
+ self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False)
197
+
198
+ def forward(self, x, seq_len=None):
199
+ # x: [bs, num_attention_heads, seq_len, head_size]
200
+ if seq_len > self.max_seq_len_cached:
201
+ self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype)
202
+
203
+ return (
204
+ self.cos_cached[:seq_len].to(dtype=x.dtype),
205
+ self.sin_cached[:seq_len].to(dtype=x.dtype),
206
+ )
207
+
208
+
209
+ # Copied from transformers.models.llama.modeling_llama.rotate_half
210
+ def rotate_half(x):
211
+ """Rotates half the hidden dims of the input."""
212
+ x1 = x[..., : x.shape[-1] // 2]
213
+ x2 = x[..., x.shape[-1] // 2 :]
214
+ return torch.cat((-x2, x1), dim=-1)
215
+
216
+
217
+ # Copied from transformers.models.mixtral.modeling_mixtral.apply_rotary_pos_emb
218
+ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
219
+ """Applies Rotary Position Embedding to the query and key tensors.
220
+
221
+ Args:
222
+ q (`torch.Tensor`): The query tensor.
223
+ k (`torch.Tensor`): The key tensor.
224
+ cos (`torch.Tensor`): The cosine part of the rotary embedding.
225
+ sin (`torch.Tensor`): The sine part of the rotary embedding.
226
+ position_ids (`torch.Tensor`):
227
+ The position indices of the tokens corresponding to the query and key tensors. For example, this can be
228
+ used to pass offsetted position ids when working with a KV-cache.
229
+ unsqueeze_dim (`int`, *optional*, defaults to 1):
230
+ The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
231
+ sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
232
+ that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
233
+ k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
234
+ cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
235
+ the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
236
+ Returns:
237
+ `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
238
+ """
239
+ cos = cos[position_ids].unsqueeze(unsqueeze_dim)
240
+ sin = sin[position_ids].unsqueeze(unsqueeze_dim)
241
+ q_embed = (q * cos) + (rotate_half(q) * sin)
242
+ k_embed = (k * cos) + (rotate_half(k) * sin)
243
+ return q_embed, k_embed
244
+
245
+
246
+ # Modified from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2Moe
247
+ class Qwen2MoeMLP(nn.Module):
248
+ def __init__(self, config, intermediate_size=None):
249
+ super().__init__()
250
+ self.config = config
251
+ self.hidden_size = config.hidden_size
252
+ self.intermediate_size = intermediate_size
253
+ self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
254
+ self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
255
+ self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
256
+ self.act_fn = ACT2FN[config.hidden_act]
257
+
258
+ def forward(self, x,language_ids:Optional[torch.LongTensor]=None):
259
+ return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
260
+
261
+
262
+ # Copied from transformers.models.llama.modeling_llama.repeat_kv
263
+ def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
264
+ """
265
+ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
266
+ num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
267
+ """
268
+ batch, num_key_value_heads, slen, head_dim = hidden_states.shape
269
+ if n_rep == 1:
270
+ return hidden_states
271
+ hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
272
+ return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
273
+
274
+
275
+ # Copied from transformers.models.qwen2.modeling_qwen2.Qwen2Attention with Qwen2->Qwen2Moe
276
+ class Qwen2MoeAttention(nn.Module):
277
+ """
278
+ Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
279
+ and "Generating Long Sequences with Sparse Transformers".
280
+ """
281
+
282
+ def __init__(self, config: UpcyclingQwen2MoeConfig, layer_idx: Optional[int] = None):
283
+ super().__init__()
284
+ self.config = config
285
+ self.layer_idx = layer_idx
286
+ if layer_idx is None:
287
+ logger.warning_once(
288
+ f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
289
+ "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
290
+ "when creating this class."
291
+ )
292
+
293
+ self.hidden_size = config.hidden_size
294
+ self.num_heads = config.num_attention_heads
295
+ self.head_dim = self.hidden_size // self.num_heads
296
+ self.num_key_value_heads = config.num_key_value_heads
297
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
298
+ self.max_position_embeddings = config.max_position_embeddings
299
+ self.rope_theta = config.rope_theta
300
+ self.is_causal = True
301
+ self.attention_dropout = config.attention_dropout
302
+
303
+ if (self.head_dim * self.num_heads) != self.hidden_size:
304
+ raise ValueError(
305
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
306
+ f" and `num_heads`: {self.num_heads})."
307
+ )
308
+ self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
309
+ self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
310
+ self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
311
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
312
+
313
+ self.rotary_emb = Qwen2MoeRotaryEmbedding(
314
+ self.head_dim,
315
+ max_position_embeddings=self.max_position_embeddings,
316
+ base=self.rope_theta,
317
+ )
318
+
319
+ def forward(
320
+ self,
321
+ hidden_states: torch.Tensor,
322
+ attention_mask: Optional[torch.Tensor] = None,
323
+ position_ids: Optional[torch.LongTensor] = None,
324
+ past_key_value: Optional[Cache] = None,
325
+ output_attentions: bool = False,
326
+ use_cache: bool = False,
327
+ cache_position: Optional[torch.LongTensor] = None,
328
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
329
+ bsz, q_len, _ = hidden_states.size()
330
+
331
+ query_states = self.q_proj(hidden_states)
332
+ key_states = self.k_proj(hidden_states)
333
+ value_states = self.v_proj(hidden_states)
334
+
335
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
336
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
337
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
338
+
339
+ kv_seq_len = key_states.shape[-2]
340
+ if past_key_value is not None:
341
+ if self.layer_idx is None:
342
+ raise ValueError(
343
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
344
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
345
+ "with a layer index."
346
+ )
347
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
348
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
349
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
350
+
351
+ if past_key_value is not None:
352
+ cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
353
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
354
+
355
+ # repeat k/v heads if n_kv_heads < n_heads
356
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
357
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
358
+
359
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
360
+
361
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
362
+ raise ValueError(
363
+ f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
364
+ f" {attn_weights.size()}"
365
+ )
366
+
367
+ if attention_mask is not None: # no matter the length, we just slice it
368
+ causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
369
+ attn_weights = attn_weights + causal_mask
370
+
371
+ # upcast attention to fp32
372
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
373
+ attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
374
+ attn_output = torch.matmul(attn_weights, value_states)
375
+
376
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
377
+ raise ValueError(
378
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
379
+ f" {attn_output.size()}"
380
+ )
381
+
382
+ attn_output = attn_output.transpose(1, 2).contiguous()
383
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
384
+
385
+ attn_output = self.o_proj(attn_output)
386
+
387
+ if not output_attentions:
388
+ attn_weights = None
389
+
390
+ return attn_output, attn_weights, past_key_value
391
+
392
+
393
+ # Copied from transformers.models.qwen2.modeling_qwen2.Qwen2FlashAttention2 with Qwen2->Qwen2Moe
394
+ class Qwen2MoeFlashAttention2(Qwen2MoeAttention):
395
+ """
396
+ Qwen2Moe flash attention module, following Qwen2Moe attention module. This module inherits from `Qwen2MoeAttention`
397
+ as the weights of the module stays untouched. The only required change would be on the forward pass
398
+ where it needs to correctly call the public API of flash attention and deal with padding tokens
399
+ in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
400
+ config.max_window_layers layers.
401
+ """
402
+
403
+ # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
404
+ def __init__(self, *args, **kwargs):
405
+ super().__init__(*args, **kwargs)
406
+
407
+ # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
408
+ # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
409
+ # 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).
410
+ self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
411
+
412
+ def forward(
413
+ self,
414
+ hidden_states: torch.Tensor,
415
+ attention_mask: Optional[torch.Tensor] = None,
416
+ position_ids: Optional[torch.LongTensor] = None,
417
+ past_key_value: Optional[Cache] = None,
418
+ output_attentions: bool = False,
419
+ use_cache: bool = False,
420
+ cache_position: Optional[torch.LongTensor] = None,
421
+ ):
422
+ bsz, q_len, _ = hidden_states.size()
423
+
424
+ query_states = self.q_proj(hidden_states)
425
+ key_states = self.k_proj(hidden_states)
426
+ value_states = self.v_proj(hidden_states)
427
+
428
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
429
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
430
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
431
+
432
+ kv_seq_len = key_states.shape[-2]
433
+ if past_key_value is not None:
434
+ if self.layer_idx is None:
435
+ raise ValueError(
436
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
437
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
438
+ "with a layer index."
439
+ )
440
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
441
+
442
+ # Because the input can be padded, the absolute sequence length depends on the max position id.
443
+ rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
444
+ cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len)
445
+
446
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
447
+
448
+ use_sliding_windows = (
449
+ _flash_supports_window_size
450
+ and getattr(self.config, "sliding_window", None) is not None
451
+ and kv_seq_len > self.config.sliding_window
452
+ and self.config.use_sliding_window
453
+ )
454
+
455
+ if not _flash_supports_window_size:
456
+ logger.warning_once(
457
+ "The current flash attention version does not support sliding window attention, for a more memory efficient implementation"
458
+ " make sure to upgrade flash-attn library."
459
+ )
460
+
461
+ if past_key_value is not None:
462
+ # Activate slicing cache only if the config has a value `sliding_windows` attribute
463
+ cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
464
+ if (
465
+ getattr(self.config, "sliding_window", None) is not None
466
+ and kv_seq_len > self.config.sliding_window
467
+ and cache_has_contents
468
+ ):
469
+ slicing_tokens = 1 - self.config.sliding_window
470
+
471
+ past_key = past_key_value[self.layer_idx][0]
472
+ past_value = past_key_value[self.layer_idx][1]
473
+
474
+ past_key = past_key[:, :, slicing_tokens:, :].contiguous()
475
+ past_value = past_value[:, :, slicing_tokens:, :].contiguous()
476
+
477
+ if past_key.shape[-2] != self.config.sliding_window - 1:
478
+ raise ValueError(
479
+ f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
480
+ f" {past_key.shape}"
481
+ )
482
+
483
+ if attention_mask is not None:
484
+ attention_mask = attention_mask[:, slicing_tokens:]
485
+ attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
486
+
487
+ cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
488
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
489
+
490
+ # repeat k/v heads if n_kv_heads < n_heads
491
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
492
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
493
+ dropout_rate = 0.0 if not self.training else self.attention_dropout
494
+
495
+ # In PEFT, usually we cast the layer norms in float32 for training stability reasons
496
+ # therefore the input hidden states gets silently casted in float32. Hence, we need
497
+ # cast them back in float16 just to be sure everything works as expected.
498
+ input_dtype = query_states.dtype
499
+ if input_dtype == torch.float32:
500
+ if torch.is_autocast_enabled():
501
+ target_dtype = torch.get_autocast_gpu_dtype()
502
+ # Handle the case where the model is quantized
503
+ elif hasattr(self.config, "_pre_quantization_dtype"):
504
+ target_dtype = self.config._pre_quantization_dtype
505
+ else:
506
+ target_dtype = self.q_proj.weight.dtype
507
+
508
+ logger.warning_once(
509
+ f"The input hidden states seems to be silently casted in float32, this might be related to"
510
+ f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
511
+ f" {target_dtype}."
512
+ )
513
+
514
+ query_states = query_states.to(target_dtype)
515
+ key_states = key_states.to(target_dtype)
516
+ value_states = value_states.to(target_dtype)
517
+
518
+ # Reashape to the expected shape for Flash Attention
519
+ query_states = query_states.transpose(1, 2)
520
+ key_states = key_states.transpose(1, 2)
521
+ value_states = value_states.transpose(1, 2)
522
+
523
+ attn_output = self._flash_attention_forward(
524
+ query_states,
525
+ key_states,
526
+ value_states,
527
+ attention_mask,
528
+ q_len,
529
+ dropout=dropout_rate,
530
+ use_sliding_windows=use_sliding_windows,
531
+ )
532
+
533
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
534
+ attn_output = self.o_proj(attn_output)
535
+
536
+ if not output_attentions:
537
+ attn_weights = None
538
+
539
+ return attn_output, attn_weights, past_key_value
540
+
541
+ def _flash_attention_forward(
542
+ self,
543
+ query_states,
544
+ key_states,
545
+ value_states,
546
+ attention_mask,
547
+ query_length,
548
+ dropout=0.0,
549
+ softmax_scale=None,
550
+ use_sliding_windows=False,
551
+ ):
552
+ """
553
+ Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
554
+ first unpad the input, then computes the attention scores and pad the final attention scores.
555
+
556
+ Args:
557
+ query_states (`torch.Tensor`):
558
+ Input query states to be passed to Flash Attention API
559
+ key_states (`torch.Tensor`):
560
+ Input key states to be passed to Flash Attention API
561
+ value_states (`torch.Tensor`):
562
+ Input value states to be passed to Flash Attention API
563
+ attention_mask (`torch.Tensor`):
564
+ The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
565
+ position of padding tokens and 1 for the position of non-padding tokens.
566
+ dropout (`float`):
567
+ Attention dropout
568
+ softmax_scale (`float`, *optional*):
569
+ The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
570
+ use_sliding_windows (`bool`, *optional*):
571
+ Whether to activate sliding window attention.
572
+ """
573
+ if not self._flash_attn_uses_top_left_mask:
574
+ causal = self.is_causal
575
+ else:
576
+ # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
577
+ causal = self.is_causal and query_length != 1
578
+
579
+ # Decide whether to use SWA or not by layer index.
580
+ if use_sliding_windows and self.layer_idx >= self.config.max_window_layers:
581
+ use_sliding_windows = False
582
+
583
+ # Contains at least one padding token in the sequence
584
+ if attention_mask is not None:
585
+ batch_size = query_states.shape[0]
586
+ query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
587
+ query_states, key_states, value_states, attention_mask, query_length
588
+ )
589
+
590
+ cu_seqlens_q, cu_seqlens_k = cu_seq_lens
591
+ max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
592
+
593
+ if not use_sliding_windows:
594
+ attn_output_unpad = flash_attn_varlen_func(
595
+ query_states,
596
+ key_states,
597
+ value_states,
598
+ cu_seqlens_q=cu_seqlens_q,
599
+ cu_seqlens_k=cu_seqlens_k,
600
+ max_seqlen_q=max_seqlen_in_batch_q,
601
+ max_seqlen_k=max_seqlen_in_batch_k,
602
+ dropout_p=dropout,
603
+ softmax_scale=softmax_scale,
604
+ causal=causal,
605
+ )
606
+ else:
607
+ attn_output_unpad = flash_attn_varlen_func(
608
+ query_states,
609
+ key_states,
610
+ value_states,
611
+ cu_seqlens_q=cu_seqlens_q,
612
+ cu_seqlens_k=cu_seqlens_k,
613
+ max_seqlen_q=max_seqlen_in_batch_q,
614
+ max_seqlen_k=max_seqlen_in_batch_k,
615
+ dropout_p=dropout,
616
+ softmax_scale=softmax_scale,
617
+ causal=causal,
618
+ window_size=(self.config.sliding_window, self.config.sliding_window),
619
+ )
620
+
621
+ attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
622
+ else:
623
+ if not use_sliding_windows:
624
+ attn_output = flash_attn_func(
625
+ query_states,
626
+ key_states,
627
+ value_states,
628
+ dropout,
629
+ softmax_scale=softmax_scale,
630
+ causal=causal,
631
+ )
632
+ else:
633
+ attn_output = flash_attn_func(
634
+ query_states,
635
+ key_states,
636
+ value_states,
637
+ dropout,
638
+ softmax_scale=softmax_scale,
639
+ causal=causal,
640
+ window_size=(self.config.sliding_window, self.config.sliding_window),
641
+ )
642
+
643
+ return attn_output
644
+
645
+ # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
646
+ def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
647
+ batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
648
+
649
+ # On the first iteration we need to properly re-create the padding mask
650
+ # by slicing it on the proper place
651
+ if kv_seq_len != attention_mask.shape[-1]:
652
+ attention_mask_num_tokens = attention_mask.shape[-1]
653
+ attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
654
+
655
+ indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
656
+
657
+ key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
658
+ value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
659
+
660
+ if query_length == kv_seq_len:
661
+ query_layer = index_first_axis(
662
+ query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
663
+ )
664
+ cu_seqlens_q = cu_seqlens_k
665
+ max_seqlen_in_batch_q = max_seqlen_in_batch_k
666
+ indices_q = indices_k
667
+ elif query_length == 1:
668
+ max_seqlen_in_batch_q = 1
669
+ cu_seqlens_q = torch.arange(
670
+ batch_size + 1, dtype=torch.int32, device=query_layer.device
671
+ ) # There is a memcpy here, that is very bad.
672
+ indices_q = cu_seqlens_q[:-1]
673
+ query_layer = query_layer.squeeze(1)
674
+ else:
675
+ # The -q_len: slice assumes left padding.
676
+ attention_mask = attention_mask[:, -query_length:]
677
+ query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
678
+
679
+ return (
680
+ query_layer,
681
+ key_layer,
682
+ value_layer,
683
+ indices_q,
684
+ (cu_seqlens_q, cu_seqlens_k),
685
+ (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
686
+ )
687
+
688
+
689
+ # Copied from transformers.models.mixtral.modeling_mixtral.MixtralSdpaAttention with Mixtral->Qwen2Moe
690
+ class Qwen2MoeSdpaAttention(Qwen2MoeAttention):
691
+ """
692
+ Qwen2Moe attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
693
+ `Qwen2MoeAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
694
+ SDPA API.
695
+ """
696
+
697
+ # Adapted from Qwen2MoeAttention.forward
698
+ def forward(
699
+ self,
700
+ hidden_states: torch.Tensor,
701
+ attention_mask: Optional[torch.Tensor] = None,
702
+ position_ids: Optional[torch.LongTensor] = None,
703
+ past_key_value: Optional[Cache] = None,
704
+ output_attentions: bool = False,
705
+ use_cache: bool = False,
706
+ cache_position: Optional[torch.LongTensor] = None,
707
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
708
+ if output_attentions:
709
+ # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
710
+ logger.warning_once(
711
+ "Qwen2MoeModel is using Qwen2MoeSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
712
+ '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.'
713
+ )
714
+ return super().forward(
715
+ hidden_states=hidden_states,
716
+ attention_mask=attention_mask,
717
+ position_ids=position_ids,
718
+ past_key_value=past_key_value,
719
+ output_attentions=output_attentions,
720
+ use_cache=use_cache,
721
+ )
722
+
723
+ bsz, q_len, _ = hidden_states.size()
724
+
725
+ query_states = self.q_proj(hidden_states)
726
+ key_states = self.k_proj(hidden_states)
727
+ value_states = self.v_proj(hidden_states)
728
+
729
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
730
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
731
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
732
+
733
+ kv_seq_len = key_states.shape[-2]
734
+ if past_key_value is not None:
735
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
736
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
737
+
738
+ query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
739
+
740
+ if past_key_value is not None:
741
+ cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} # Specific to RoPE models
742
+ key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
743
+
744
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
745
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
746
+
747
+ causal_mask = attention_mask
748
+ if attention_mask is not None: # no matter the length, we just slice it
749
+ causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
750
+
751
+ # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
752
+ # Reference: https://github.com/pytorch/pytorch/issues/112577.
753
+ if query_states.device.type == "cuda" and attention_mask is not None:
754
+ query_states = query_states.contiguous()
755
+ key_states = key_states.contiguous()
756
+ value_states = value_states.contiguous()
757
+
758
+ # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
759
+ # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
760
+ # 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.
761
+ is_causal = True if causal_mask is None and q_len > 1 else False
762
+
763
+ attn_output = torch.nn.functional.scaled_dot_product_attention(
764
+ query_states,
765
+ key_states,
766
+ value_states,
767
+ attn_mask=causal_mask,
768
+ dropout_p=self.attention_dropout if self.training else 0.0,
769
+ is_causal=is_causal,
770
+ )
771
+
772
+ attn_output = attn_output.transpose(1, 2).contiguous()
773
+ attn_output = attn_output.view(bsz, q_len, self.hidden_size)
774
+
775
+ attn_output = self.o_proj(attn_output)
776
+
777
+ return attn_output, None, past_key_value
778
+
779
+
780
+ QWEN2MOE_ATTENTION_CLASSES = {
781
+ "eager": Qwen2MoeAttention,
782
+ "flash_attention_2": Qwen2MoeFlashAttention2,
783
+ "sdpa": Qwen2MoeSdpaAttention,
784
+ }
785
+
786
+
787
+ class Qwen2MoeSparseMoeBlock(nn.Module):
788
+ def __init__(self, config):
789
+ super().__init__()
790
+ self.num_experts = config.num_experts
791
+ self.top_k = config.num_experts_per_tok
792
+ self.norm_topk_prob = config.norm_topk_prob
793
+
794
+ # gating
795
+ self.gate = nn.Linear(config.hidden_size, config.num_experts, bias=False)
796
+ self.experts = nn.ModuleList(
797
+ [Qwen2MoeMLP(config, intermediate_size=config.moe_intermediate_size) for _ in range(self.num_experts)]
798
+ )
799
+ #share
800
+ self.share_flag=config.share_flag
801
+
802
+ if self.share_flag:
803
+ self.shared_expert = Qwen2MoeMLP(config, intermediate_size=config.shared_expert_intermediate_size)
804
+ self.shared_expert_gate = torch.nn.Linear(config.hidden_size, 1, bias=False)
805
+
806
+ #language-specific
807
+ self.language_gate=config.language_gate
808
+
809
+ def forward(self, hidden_states: torch.Tensor,language_ids:Optional[torch.LongTensor] = None) -> torch.Tensor:
810
+
811
+ batch_size, sequence_length, hidden_dim = hidden_states.shape
812
+ hidden_states = hidden_states.view(-1, hidden_dim)
813
+ if self.language_gate and self.training :
814
+ if language_ids is None:
815
+ raise ValueError('language_ids is not initialized')
816
+ language_ids=language_ids.view(batch_size*sequence_length,-1)
817
+ # router_logits: (batch * sequence_length, n_experts)
818
+ router_logits = self.gate(hidden_states)
819
+
820
+ routing_weights = F.softmax(router_logits, dim=1, dtype=torch.float)
821
+
822
+ _, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
823
+
824
+ #language specific select one expert
825
+ if self.language_gate and self.training:
826
+ if language_ids is None:
827
+ raise ValueError('language_ids is not initialized')
828
+ assert language_ids.shape[0]==selected_experts.shape[0],f'{language_ids.shape},{selected_experts.shape}'
829
+ language_experts=language_ids.to(selected_experts.dtype)
830
+ mask=torch.sum((language_experts==selected_experts).int(),dim=1,keepdims=True).bool()
831
+ selected_experts[:,-1]=torch.where(mask.squeeze(),selected_experts[:,-1],language_experts.squeeze())
832
+ routing_weights=torch.gather(routing_weights,1,selected_experts)
833
+ else:
834
+ routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
835
+
836
+ if self.norm_topk_prob:
837
+ routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
838
+ # we cast back to the input dtype
839
+ routing_weights = routing_weights.to(hidden_states.dtype)
840
+
841
+ final_hidden_states = torch.zeros(
842
+ (batch_size * sequence_length, hidden_dim), dtype=hidden_states.dtype, device=hidden_states.device
843
+ )
844
+
845
+ # One hot encode the selected experts to create an expert mask
846
+ # this will be used to easily index which expert is going to be sollicitated
847
+ expert_mask = torch.nn.functional.one_hot(selected_experts, num_classes=self.num_experts).permute(2, 1, 0)
848
+
849
+ # Loop over all available experts in the model and perform the computation on each expert
850
+ for expert_idx in range(self.num_experts):
851
+ expert_layer = self.experts[expert_idx]
852
+ idx, top_x = torch.where(expert_mask[expert_idx])
853
+
854
+ # Index the correct hidden states and compute the expert hidden state for
855
+ # the current expert. We need to make sure to multiply the output hidden
856
+ # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
857
+ current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
858
+ current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
859
+
860
+ # However `index_add_` only support torch tensors for indexing so we'll use
861
+ # the `top_x` tensor here.
862
+ final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
863
+
864
+ if self.share_flag:
865
+
866
+ shared_expert_output = self.shared_expert(hidden_states)
867
+ shared_expert_output = F.sigmoid(self.shared_expert_gate(hidden_states)) * shared_expert_output
868
+
869
+ final_hidden_states = final_hidden_states + shared_expert_output
870
+
871
+ final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
872
+ return final_hidden_states, router_logits
873
+
874
+
875
+ class Qwen2MoeDecoderLayer(nn.Module):
876
+ def __init__(self, config: UpcyclingQwen2MoeConfig, layer_idx: int):
877
+ super().__init__()
878
+ self.hidden_size = config.hidden_size
879
+
880
+ self.self_attn = QWEN2MOE_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
881
+
882
+ if (layer_idx not in config.mlp_only_layers) and (
883
+ config.num_experts > 0 and (layer_idx + 1) % config.decoder_sparse_step == 0
884
+ ):
885
+ self.mlp = Qwen2MoeSparseMoeBlock(config)
886
+ else:
887
+ self.mlp = Qwen2MoeMLP(config, intermediate_size=config.intermediate_size)
888
+
889
+ self.input_layernorm = Qwen2MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
890
+ self.post_attention_layernorm = Qwen2MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
891
+
892
+ def forward(
893
+ self,
894
+ hidden_states: torch.Tensor,
895
+ language_ids:Optional[torch.LongTensor] = None,
896
+ attention_mask: Optional[torch.Tensor] = None,
897
+ position_ids: Optional[torch.LongTensor] = None,
898
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
899
+ output_attentions: Optional[bool] = False,
900
+ output_router_logits: Optional[bool] = False,
901
+ use_cache: Optional[bool] = False,
902
+ cache_position: Optional[torch.LongTensor] = None,
903
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
904
+ """
905
+ Args:
906
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
907
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
908
+ `(batch, sequence_length)` where padding elements are indicated by 0.
909
+ output_attentions (`bool`, *optional*):
910
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
911
+ returned tensors for more detail.
912
+ output_router_logits (`bool`, *optional*):
913
+ Whether or not to return the logits of all the routers. They are useful for computing the router loss,
914
+ and should not be returned during inference.
915
+ use_cache (`bool`, *optional*):
916
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
917
+ (see `past_key_values`).
918
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
919
+ cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
920
+ Indices depicting the position of the input sequence tokens in the sequence.
921
+ """
922
+
923
+ residual = hidden_states
924
+
925
+ hidden_states = self.input_layernorm(hidden_states)
926
+
927
+ # Self Attention
928
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
929
+ hidden_states=hidden_states,
930
+ attention_mask=attention_mask,
931
+ position_ids=position_ids,
932
+ past_key_value=past_key_value,
933
+ output_attentions=output_attentions,
934
+ use_cache=use_cache,
935
+ cache_position=cache_position,
936
+ )
937
+ hidden_states = residual + hidden_states
938
+
939
+ # Fully Connected
940
+ residual = hidden_states
941
+ hidden_states = self.post_attention_layernorm(hidden_states)
942
+
943
+ hidden_states = self.mlp(hidden_states,language_ids)
944
+ if isinstance(hidden_states, tuple):
945
+ hidden_states, router_logits = hidden_states
946
+ else:
947
+ router_logits = None
948
+
949
+ hidden_states = residual + hidden_states
950
+
951
+ outputs = (hidden_states,)
952
+
953
+ if output_attentions:
954
+ outputs += (self_attn_weights,)
955
+
956
+ if use_cache:
957
+ outputs += (present_key_value,)
958
+
959
+ if output_router_logits:
960
+ outputs += (router_logits,)
961
+
962
+ return outputs
963
+
964
+
965
+ class UpcyclingQwen2MoePreTrainedModel(PreTrainedModel):
966
+ config_class = UpcyclingQwen2MoeConfig
967
+ base_model_prefix = "model"
968
+ supports_gradient_checkpointing = True
969
+ _no_split_modules = ["Qwen2MoeDecoderLayer"]
970
+ _skip_keys_device_placement = "past_key_values"
971
+ _supports_flash_attn_2 = True
972
+ _supports_sdpa = True
973
+ _supports_cache_class = True
974
+
975
+ def _init_weights(self, module):
976
+ std = self.config.initializer_range
977
+ if isinstance(module, nn.Linear):
978
+ module.weight.data.normal_(mean=0.0, std=std)
979
+ if module.bias is not None:
980
+ module.bias.data.zero_()
981
+ elif isinstance(module, nn.Embedding):
982
+ module.weight.data.normal_(mean=0.0, std=std)
983
+ if module.padding_idx is not None:
984
+ module.weight.data[module.padding_idx].zero_()
985
+
986
+ @classmethod
987
+ def from_qwen(cls, pretrained_model_name_or_path, *model_args, **kwargs):
988
+ share_flag=kwargs.pop('share_flag')
989
+ attn_init_change=kwargs.pop('attn_init_change')
990
+ language_gate=kwargs.pop('language_gate')
991
+
992
+ config = cls.config_class.from_pretrained(pretrained_model_name_or_path)
993
+
994
+ config.share_flag=True if isinstance(share_flag,bool) and share_flag else False
995
+ config.attn_init_change=True if isinstance(attn_init_change,bool) and attn_init_change else False
996
+ config.language_gate=True if isinstance(language_gate,bool) and language_gate else False
997
+
998
+ print('share_flag',config.share_flag)
999
+ print('attn_init_change',config.attn_init_change)
1000
+ print('language_gate',config.language_gate)
1001
+
1002
+ config.num_experts_per_tok = config.num_experts_per_tok if not config.share_flag else config.num_experts_per_tok-1
1003
+ config.num_experts = config.num_experts if not config.share_flag else config.num_experts-1
1004
+
1005
+ base_model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
1006
+ base_cls = type(base_model)
1007
+
1008
+ print(cls.config_class,cls)
1009
+
1010
+ #create auto_map
1011
+ #allows you to use your custom model with the auto-API (but doesn’t share any custom code with other users).
1012
+ cls.config_class.register_for_auto_class()
1013
+ cls.register_for_auto_class('AutoModelForCausalLM')
1014
+
1015
+ # assert base_cls.__name__ == "Qwen2ForCausalLM", f"Invalid convert base model type: {base_cls}"
1016
+
1017
+ model = cls(config)
1018
+ print(f"converting {base_cls.__name__} to {cls.__name__}")
1019
+
1020
+ #MoE architechture
1021
+ model_dict=model.state_dict()
1022
+ base_model_dict = base_model.state_dict()
1023
+
1024
+ #lm_head
1025
+ print('lm_head.weight',model_dict['lm_head.weight'],base_model_dict['lm_head.weight'])
1026
+
1027
+ shared_keys=set(model_dict)&set(base_model_dict)
1028
+ init_keys=[]
1029
+ #attention
1030
+ for k in shared_keys:
1031
+ if k not in init_keys and 'self_attn' in k:
1032
+ init_keys.append(k)
1033
+ if not config.attn_init_change:
1034
+ model_dict[k]=base_model_dict[k]
1035
+
1036
+ if config.attn_init_change:
1037
+ #initilization with upper and lower
1038
+ for layer_id in range(config.num_hidden_layers):
1039
+ if layer_id ==0 or config.num_hidden_layers-1:
1040
+ model_dict[f'model.layers.{layer_id}.self_attn.q_proj.bias']=base_model_dict[f'model.layers.{layer_id}.self_attn.q_proj.bias']
1041
+ model_dict[f'model.layers.{layer_id}.self_attn.q_proj.weight']=base_model_dict[f'model.layers.{layer_id}.self_attn.q_proj.weight']
1042
+ model_dict[f'model.layers.{layer_id}.self_attn.k_proj.bias']=base_model_dict[f'model.layers.{layer_id}.self_attn.k_proj.bias']
1043
+ model_dict[f'model.layers.{layer_id}.self_attn.k_proj.weight']=base_model_dict[f'model.layers.{layer_id}.self_attn.k_proj.weight']
1044
+ model_dict[f'model.layers.{layer_id}.self_attn.v_proj.bias']=base_model_dict[f'model.layers.{layer_id}.self_attn.v_proj.bias']
1045
+ model_dict[f'model.layers.{layer_id}.self_attn.v_proj.weight']=base_model_dict[f'model.layers.{layer_id}.self_attn.v_proj.weight']
1046
+ model_dict[f'model.layers.{layer_id}.self_attn.o_proj.weight']=base_model_dict[f'model.layers.{layer_id}.self_attn.o_proj.weight']
1047
+ else:
1048
+ model_dict[f'model.layers.{layer_id}.self_attn.q_proj.bias']=1/3*(base_model_dict[f'model.layers.{layer_id}.self_attn.q_proj.bias']+base_model_dict[f'model.layers.{layer_id+1}.self_attn.q_proj.bias']+base_model_dict[f'model.layers.{layer_id-1}.self_attn.q_proj.bias'])
1049
+ model_dict[f'model.layers.{layer_id}.self_attn.q_proj.weight']=1/3*(base_model_dict[f'model.layers.{layer_id}.self_attn.q_proj.weight']+base_model_dict[f'model.layers.{layer_id+1}.self_attn.q_proj.weight']+base_model_dict[f'model.layers.{layer_id-1}.self_attn.q_proj.weight'])
1050
+ model_dict[f'model.layers.{layer_id}.self_attn.k_proj.bias']=1/3*(base_model_dict[f'model.layers.{layer_id}.self_attn.k_proj.bias']+base_model_dict[f'model.layers.{layer_id+1}.self_attn.k_proj.bias']+base_model_dict[f'model.layers.{layer_id-1}.self_attn.k_proj.bias'])
1051
+ model_dict[f'model.layers.{layer_id}.self_attn.k_proj.weight']=1/3*(base_model_dict[f'model.layers.{layer_id}.self_attn.k_proj.weight']+base_model_dict[f'model.layers.{layer_id+1}.self_attn.k_proj.weight']+base_model_dict[f'model.layers.{layer_id-1}.self_attn.k_proj.weight'])
1052
+ model_dict[f'model.layers.{layer_id}.self_attn.v_proj.bias']=1/3*(base_model_dict[f'model.layers.{layer_id}.self_attn.v_proj.bias']+base_model_dict[f'model.layers.{layer_id+1}.self_attn.v_proj.bias']+base_model_dict[f'model.layers.{layer_id-1}.self_attn.v_proj.bias'])
1053
+ model_dict[f'model.layers.{layer_id}.self_attn.v_proj.weight']=1/3*(base_model_dict[f'model.layers.{layer_id}.self_attn.v_proj.weight']+base_model_dict[f'model.layers.{layer_id+1}.self_attn.v_proj.weight']+base_model_dict[f'model.layers.{layer_id-1}.self_attn.v_proj.weight'])
1054
+ model_dict[f'model.layers.{layer_id}.self_attn.o_proj.weight']=1/3*(base_model_dict[f'model.layers.{layer_id}.self_attn.o_proj.weight']+base_model_dict[f'model.layers.{layer_id+1}.self_attn.o_proj.weight']+base_model_dict[f'model.layers.{layer_id-1}.self_attn.o_proj.weight'])
1055
+
1056
+ #mlp
1057
+ if config.mlp_only_layers:
1058
+ for layer_id in config.mlp_only_layers:
1059
+ key_mapping=sum([
1060
+ [
1061
+ (f'model.layers.{layer_id}.mlp.down_proj.weight',f'model.layers.{layer_id}.mlp.down_proj.weight'),
1062
+ (f'model.layers.{layer_id}.mlp.gate_proj.weight',f'model.layers.{layer_id}.mlp.gate_proj.weight'),
1063
+ (f'model.layers.{layer_id}.mlp.up_proj.weight',f'model.layers.{layer_id}.mlp.up_proj.weight'),
1064
+ ]]
1065
+ ,[])
1066
+ for model_key,base_model_key in key_mapping:
1067
+ model_dict[model_key]=base_model_dict[base_model_key]
1068
+ init_keys.append(model_key)
1069
+ moe_only_layers=list(set(range(config.num_hidden_layers))-set(config.mlp_only_layers)) if config.mlp_only_layers else config.num_hidden_layers
1070
+ #moe-mlp-expert
1071
+ for layer_id in moe_only_layers:
1072
+ key_mapping=sum([
1073
+ [
1074
+ (f'model.layers.{layer_id}.mlp.experts.{expert_id}.down_proj.weight',f'model.layers.{layer_id}.mlp.down_proj.weight'),
1075
+ (f'model.layers.{layer_id}.mlp.experts.{expert_id}.gate_proj.weight',f'model.layers.{layer_id}.mlp.gate_proj.weight'),
1076
+ (f'model.layers.{layer_id}.mlp.experts.{expert_id}.up_proj.weight',f'model.layers.{layer_id}.mlp.up_proj.weight'),
1077
+ ] for expert_id in range(config.num_experts)]
1078
+ ,[])
1079
+ for model_key,base_model_key in key_mapping:
1080
+ model_dict[model_key]=base_model_dict[base_model_key]
1081
+ init_keys.append(model_key)
1082
+ #model_dict[f'model.layers.{layer_id}.mlp.gate.weight']
1083
+
1084
+ #share expert
1085
+ if config.share_flag:
1086
+ shared_key_mapping=sum([[
1087
+ (f'model.layers.{layer_id}.mlp.shared_expert.down_proj.weight',f'model.layers.{layer_id}.mlp.down_proj.weight'),
1088
+ (f'model.layers.{layer_id}.mlp.shared_expert.gate_proj.weight',f'model.layers.{layer_id}.mlp.gate_proj.weight'),
1089
+ (f'model.layers.{layer_id}.mlp.shared_expert.up_proj.weight',f'model.layers.{layer_id}.mlp.up_proj.weight'),
1090
+ ]for layer_id in range(config.num_hidden_layers)],
1091
+ [])
1092
+ for model_key,base_model_key in shared_key_mapping:
1093
+ model_dict[model_key]=base_model_dict[base_model_key]
1094
+ init_keys.append(model_key)
1095
+ # model_dict[f'model.layers.{layer_id}.mlp.shared_expert_gate.weight']
1096
+
1097
+ #norm
1098
+ for k in shared_keys:
1099
+ if k not in init_keys:
1100
+ #input_layernorm.weight,post_attention_layernorm.weight,norm.weight
1101
+ # embed_token.weight,lm_head.weight
1102
+ model_dict[k]=base_model_dict[k]
1103
+ init_keys.append(k)
1104
+
1105
+ gate_initialized = False
1106
+ shared_gate_initilizaed=False
1107
+ for key in model_dict.keys():
1108
+ if key in init_keys:
1109
+ continue
1110
+ if "mlp.gate.weight" in key:
1111
+ if gate_initialized:
1112
+ continue
1113
+ gate_initialized = True
1114
+ print(f"{cls.__name__} key [{cls.base_model_prefix}.layers.[0-{config.num_hidden_layers-1}].mlp.gate.weight] is not initialized from {base_cls.__name__}. e.g, {key}")
1115
+ continue
1116
+ if 'shared_expert_gate.weight' in key:
1117
+ if shared_gate_initilizaed:
1118
+ continue
1119
+ shared_gate_initilizaed = True
1120
+ print(f"{cls.__name__} key [{cls.base_model_prefix}.layers.[0-{config.num_hidden_layers-1}].mlp.shared_expert_gate.weight] is not initialized from {base_cls.__name__}. e.g, {key}")
1121
+ continue
1122
+
1123
+ raise NotImplementedError(f"{cls.__name__} key [{key}] is not correctly initilized from {base_cls.__name__}.")
1124
+
1125
+ model.load_state_dict(model_dict)
1126
+ print(f"Done converted, alreadly check all parameters of {cls.__name__} are initialized from {base_cls.__name__}.")
1127
+
1128
+ del base_model
1129
+ return model
1130
+
1131
+ @classmethod
1132
+ def from_btx(cls, pretrained_model_name_or_path, *model_args, **kwargs):
1133
+ share_flag=kwargs.pop('share_flag')
1134
+ attn_init_change=kwargs.pop('attn_init_change')
1135
+ language_gate=kwargs.pop('language_gate')
1136
+
1137
+ config = cls.config_class.from_pretrained(pretrained_model_name_or_path)
1138
+
1139
+ config.share_flag=True if isinstance(share_flag,bool) and share_flag else False
1140
+ config.attn_init_change=True if isinstance(attn_init_change,bool) and attn_init_change else False
1141
+ config.language_gate=True if isinstance(language_gate,bool) and language_gate else False
1142
+
1143
+ print('share_flag',config.share_flag)
1144
+ print('attn_init_change',config.attn_init_change)
1145
+ print('language_gate',config.language_gate)
1146
+
1147
+ config.num_experts_per_tok = config.num_experts_per_tok if not config.share_flag else config.num_experts_per_tok-1
1148
+ config.num_experts = config.num_experts if not config.share_flag else config.num_experts-1
1149
+
1150
+ base_model = AutoModelForCausalLM.from_pretrained(pretrained_model_name_or_path, *model_args, **kwargs)
1151
+ base_cls = type(base_model)
1152
+
1153
+ print(cls.config_class,cls)
1154
+
1155
+ #create auto_map
1156
+ #allows you to use your custom model with the auto-API (but doesn’t share any custom code with other users).
1157
+ cls.config_class.register_for_auto_class()
1158
+ cls.register_for_auto_class('AutoModelForCausalLM')
1159
+
1160
+ # assert base_cls.__name__ == "Qwen2ForCausalLM", f"Invalid convert base model type: {base_cls}"
1161
+
1162
+ model = cls(config)
1163
+ print(f"converting {base_cls.__name__} to {cls.__name__}")
1164
+
1165
+ #MoE architechture
1166
+ model_dict=model.state_dict()
1167
+ base_model_dict = base_model.state_dict()
1168
+
1169
+ #lm_head
1170
+ print('lm_head.weight',model_dict['lm_head.weight'],base_model_dict['lm_head.weight'])
1171
+
1172
+ shared_keys=set(model_dict)&set(base_model_dict)
1173
+ init_keys=[]
1174
+ #attention
1175
+ for k in shared_keys:
1176
+ init_keys.append(k)
1177
+ model_dict[k]=base_model_dict[k]
1178
+
1179
+ gate_initialized = False
1180
+ shared_gate_initilizaed=False
1181
+ for key in model_dict.keys():
1182
+ if key in init_keys:
1183
+ continue
1184
+ if "mlp.gate.weight" in key:
1185
+ if gate_initialized:
1186
+ continue
1187
+ gate_initialized = True
1188
+ print(f"{cls.__name__} key [{cls.base_model_prefix}.layers.[0-{config.num_hidden_layers-1}].mlp.gate.weight] is not initialized from {base_cls.__name__}. e.g, {key}")
1189
+ continue
1190
+ if 'shared_expert_gate.weight' in key:
1191
+ if shared_gate_initilizaed:
1192
+ continue
1193
+ shared_gate_initilizaed = True
1194
+ print(f"{cls.__name__} key [{cls.base_model_prefix}.layers.[0-{config.num_hidden_layers-1}].mlp.shared_expert_gate.weight] is not initialized from {base_cls.__name__}. e.g, {key}")
1195
+ continue
1196
+
1197
+ raise NotImplementedError(f"{cls.__name__} key [{key}] is not correctly initilized from {base_cls.__name__}.")
1198
+
1199
+ model.load_state_dict(model_dict)
1200
+ print(f"Done converted, alreadly check all parameters of {cls.__name__} are initialized from {base_cls.__name__}.")
1201
+
1202
+ del base_model
1203
+ return model
1204
+
1205
+ class UpcyclingQwen2MoeModel(UpcyclingQwen2MoePreTrainedModel):
1206
+ """
1207
+ Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2MoeDecoderLayer`]
1208
+
1209
+ Args:
1210
+ config: Qwen2MoeConfig
1211
+ """
1212
+
1213
+ def __init__(self, config: UpcyclingQwen2MoeConfig):
1214
+ super().__init__(config)
1215
+ self.padding_idx = config.pad_token_id
1216
+ self.vocab_size = config.vocab_size
1217
+
1218
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
1219
+ self.layers = nn.ModuleList(
1220
+ [Qwen2MoeDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
1221
+ )
1222
+ self._attn_implementation = config._attn_implementation
1223
+ self.norm = Qwen2MoeRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
1224
+
1225
+ self.gradient_checkpointing = False
1226
+ # Initialize weights and apply final processing
1227
+ self.post_init()
1228
+
1229
+ def get_input_embeddings(self):
1230
+ return self.embed_tokens
1231
+
1232
+ def set_input_embeddings(self, value):
1233
+ self.embed_tokens = value
1234
+
1235
+ def forward(
1236
+ self,
1237
+ input_ids: torch.LongTensor = None,
1238
+ language_ids :Optional[torch.LongTensor]= None,
1239
+ attention_mask: Optional[torch.Tensor] = None,
1240
+ position_ids: Optional[torch.LongTensor] = None,
1241
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1242
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1243
+ use_cache: Optional[bool] = None,
1244
+ output_attentions: Optional[bool] = None,
1245
+ output_hidden_states: Optional[bool] = None,
1246
+ output_router_logits: Optional[bool] = None,
1247
+ return_dict: Optional[bool] = None,
1248
+ cache_position: Optional[torch.LongTensor] = None,
1249
+ ) -> Union[Tuple, MoeModelOutputWithPast]:
1250
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1251
+ output_router_logits = (
1252
+ output_router_logits if output_router_logits is not None else self.config.output_router_logits
1253
+ )
1254
+ output_hidden_states = (
1255
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1256
+ )
1257
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
1258
+
1259
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1260
+
1261
+ if (input_ids is None) ^ (inputs_embeds is not None):
1262
+ raise ValueError(
1263
+ "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one"
1264
+ )
1265
+
1266
+ if self.gradient_checkpointing and self.training:
1267
+ if use_cache:
1268
+ logger.warning_once(
1269
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
1270
+ )
1271
+ use_cache = False
1272
+
1273
+ use_legacy_cache = False
1274
+ if use_cache and not isinstance(past_key_values, Cache):
1275
+ use_legacy_cache = True
1276
+ past_key_values = DynamicCache.from_legacy_cache(past_key_values)
1277
+ logger.warning_once(
1278
+ "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. "
1279
+ "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)"
1280
+ )
1281
+
1282
+ if inputs_embeds is None:
1283
+ inputs_embeds = self.embed_tokens(input_ids)
1284
+
1285
+ if cache_position is None:
1286
+ past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
1287
+ cache_position = torch.arange(
1288
+ past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
1289
+ )
1290
+ if position_ids is None:
1291
+ position_ids = cache_position.unsqueeze(0)
1292
+
1293
+ causal_mask = self._update_causal_mask(
1294
+ attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
1295
+ )
1296
+
1297
+ hidden_states = inputs_embeds
1298
+
1299
+ # decoder layers
1300
+ all_hidden_states = () if output_hidden_states else None
1301
+ all_self_attns = () if output_attentions else None
1302
+ all_router_logits = () if output_router_logits else None
1303
+ next_decoder_cache = None
1304
+
1305
+ for decoder_layer in self.layers:
1306
+ if output_hidden_states:
1307
+ all_hidden_states += (hidden_states,)
1308
+
1309
+ if self.gradient_checkpointing and self.training:
1310
+ layer_outputs = self._gradient_checkpointing_func(
1311
+ decoder_layer.__call__,
1312
+ hidden_states,
1313
+ language_ids,
1314
+ causal_mask,
1315
+ position_ids,
1316
+ past_key_values,
1317
+ output_attentions,
1318
+ output_router_logits,
1319
+ use_cache,
1320
+ cache_position,
1321
+ )
1322
+ else:
1323
+ layer_outputs = decoder_layer(
1324
+ hidden_states,
1325
+ language_ids,
1326
+ attention_mask=causal_mask,
1327
+ position_ids=position_ids,
1328
+ past_key_value=past_key_values,
1329
+ output_attentions=output_attentions,
1330
+ output_router_logits=output_router_logits,
1331
+ use_cache=use_cache,
1332
+ cache_position=cache_position,
1333
+ )
1334
+
1335
+ hidden_states = layer_outputs[0]
1336
+
1337
+ if use_cache:
1338
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
1339
+
1340
+ if output_attentions:
1341
+ all_self_attns += (layer_outputs[1],)
1342
+
1343
+ if output_router_logits and layer_outputs[-1] is not None:
1344
+ all_router_logits += (layer_outputs[-1],)
1345
+
1346
+ hidden_states = self.norm(hidden_states)
1347
+
1348
+ # add hidden states from the last decoder layer
1349
+ if output_hidden_states:
1350
+ all_hidden_states += (hidden_states,)
1351
+
1352
+ next_cache = None
1353
+ if use_cache:
1354
+ next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
1355
+
1356
+ if not return_dict:
1357
+ return tuple(
1358
+ v
1359
+ for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_router_logits]
1360
+ if v is not None
1361
+ )
1362
+ return MoeModelOutputWithPast(
1363
+ last_hidden_state=hidden_states,
1364
+ past_key_values=next_cache,
1365
+ hidden_states=all_hidden_states,
1366
+ attentions=all_self_attns,
1367
+ router_logits=all_router_logits,
1368
+ )
1369
+
1370
+ # Copied from transformers.models.llama.modeling_llama.LlamaModel._update_causal_mask
1371
+ def _update_causal_mask(
1372
+ self,
1373
+ attention_mask: torch.Tensor,
1374
+ input_tensor: torch.Tensor,
1375
+ cache_position: torch.Tensor,
1376
+ past_key_values: Cache,
1377
+ output_attentions: bool,
1378
+ ):
1379
+ # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static
1380
+ # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes.
1381
+ # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using
1382
+ # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114
1383
+
1384
+ if self.config._attn_implementation == "flash_attention_2":
1385
+ if attention_mask is not None and 0.0 in attention_mask:
1386
+ return attention_mask
1387
+ return None
1388
+
1389
+ # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
1390
+ # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
1391
+ # to infer the attention mask.
1392
+ past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
1393
+ using_static_cache = isinstance(past_key_values, StaticCache)
1394
+
1395
+ # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
1396
+ if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions:
1397
+ if AttentionMaskConverter._ignore_causal_mask_sdpa(
1398
+ attention_mask,
1399
+ inputs_embeds=input_tensor,
1400
+ past_key_values_length=past_seen_tokens,
1401
+ is_training=self.training,
1402
+ ):
1403
+ return None
1404
+
1405
+ dtype, device = input_tensor.dtype, input_tensor.device
1406
+ min_dtype = torch.finfo(dtype).min
1407
+ sequence_length = input_tensor.shape[1]
1408
+ if using_static_cache:
1409
+ target_length = past_key_values.get_max_length()
1410
+ else:
1411
+ target_length = (
1412
+ attention_mask.shape[-1]
1413
+ if isinstance(attention_mask, torch.Tensor)
1414
+ else past_seen_tokens + sequence_length + 1
1415
+ )
1416
+
1417
+ if attention_mask is not None and attention_mask.dim() == 4:
1418
+ # in this case we assume that the mask comes already in inverted form and requires no inversion or slicing
1419
+ if attention_mask.max() != 0:
1420
+ raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`")
1421
+ causal_mask = attention_mask
1422
+ else:
1423
+ causal_mask = torch.full(
1424
+ (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
1425
+ )
1426
+ if sequence_length != 1:
1427
+ causal_mask = torch.triu(causal_mask, diagonal=1)
1428
+ causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
1429
+ causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1)
1430
+ if attention_mask is not None:
1431
+ causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
1432
+ mask_length = attention_mask.shape[-1]
1433
+ padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
1434
+ padding_mask = padding_mask == 0
1435
+ causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
1436
+ padding_mask, min_dtype
1437
+ )
1438
+ if (
1439
+ self.config._attn_implementation == "sdpa"
1440
+ and attention_mask is not None
1441
+ and attention_mask.device.type == "cuda"
1442
+ and not output_attentions
1443
+ ):
1444
+ # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
1445
+ # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
1446
+ # Details: https://github.com/pytorch/pytorch/issues/110213
1447
+ causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
1448
+
1449
+ return causal_mask
1450
+
1451
+
1452
+ class UpcyclingQwen2MoeForCausalLM(UpcyclingQwen2MoePreTrainedModel):
1453
+ _tied_weights_keys = ["lm_head.weight"]
1454
+
1455
+ def __init__(self, config):
1456
+ super().__init__(config)
1457
+ self.model = UpcyclingQwen2MoeModel(config)
1458
+ self.vocab_size = config.vocab_size
1459
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
1460
+
1461
+ self.router_aux_loss_coef = config.router_aux_loss_coef
1462
+ self.num_experts = config.num_experts
1463
+ self.num_experts_per_tok = config.num_experts_per_tok
1464
+
1465
+ self.language_gate=config.language_gate
1466
+ # Initialize weights and apply final processing
1467
+ self.post_init()
1468
+
1469
+ def get_input_embeddings(self):
1470
+ return self.model.embed_tokens
1471
+
1472
+ def set_input_embeddings(self, value):
1473
+ self.model.embed_tokens = value
1474
+
1475
+ def get_output_embeddings(self):
1476
+ return self.lm_head
1477
+
1478
+ def set_output_embeddings(self, new_embeddings):
1479
+ self.lm_head = new_embeddings
1480
+
1481
+ def set_decoder(self, decoder):
1482
+ self.model = decoder
1483
+
1484
+ def get_decoder(self):
1485
+ return self.model
1486
+
1487
+ # @replace_return_docstrings(output_type=MoeCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
1488
+ def forward(
1489
+ self,
1490
+ input_ids: torch.LongTensor = None,
1491
+ language_ids: Optional[torch.LongTensor] = None,
1492
+ attention_mask: Optional[torch.Tensor] = None,
1493
+ position_ids: Optional[torch.LongTensor] = None,
1494
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
1495
+ inputs_embeds: Optional[torch.FloatTensor] = None,
1496
+ labels: Optional[torch.LongTensor] = None,
1497
+ use_cache: Optional[bool] = None,
1498
+ output_attentions: Optional[bool] = None,
1499
+ output_hidden_states: Optional[bool] = None,
1500
+ output_router_logits: Optional[bool] = None,
1501
+ return_dict: Optional[bool] = None,
1502
+ cache_position: Optional[torch.LongTensor] = None,
1503
+ ) -> Union[Tuple, MoeCausalLMOutputWithPast]:
1504
+
1505
+ output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
1506
+ output_router_logits = (
1507
+ output_router_logits if output_router_logits is not None else self.config.output_router_logits
1508
+ )
1509
+ output_hidden_states = (
1510
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
1511
+ )
1512
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
1513
+
1514
+ # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
1515
+ outputs = self.model(
1516
+ input_ids=input_ids,
1517
+ language_ids=language_ids,
1518
+ attention_mask=attention_mask,
1519
+ position_ids=position_ids,
1520
+ past_key_values=past_key_values,
1521
+ inputs_embeds=inputs_embeds,
1522
+ use_cache=use_cache,
1523
+ output_attentions=output_attentions,
1524
+ output_hidden_states=output_hidden_states,
1525
+ output_router_logits=output_router_logits,
1526
+ return_dict=return_dict,
1527
+ cache_position=cache_position,
1528
+ )
1529
+
1530
+ hidden_states = outputs[0]
1531
+ logits = self.lm_head(hidden_states)
1532
+ logits = logits.float()
1533
+
1534
+ loss = None
1535
+ if labels is not None:
1536
+ # Shift so that tokens < n predict n
1537
+ shift_logits = logits[..., :-1, :].contiguous()
1538
+ shift_labels = labels[..., 1:].contiguous()
1539
+ # Flatten the tokens
1540
+ loss_fct = CrossEntropyLoss()
1541
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
1542
+ shift_labels = shift_labels.view(-1)
1543
+ # Enable model parallelism
1544
+ shift_labels = shift_labels.to(shift_logits.device)
1545
+ loss = loss_fct(shift_logits, shift_labels)
1546
+
1547
+ aux_loss = None
1548
+ if output_router_logits:
1549
+ aux_loss = load_balancing_loss_func(
1550
+ outputs.router_logits if return_dict else outputs[-1],
1551
+ self.num_experts,
1552
+ self.num_experts_per_tok,
1553
+ attention_mask,
1554
+ )
1555
+ if labels is not None:
1556
+ loss += self.router_aux_loss_coef * aux_loss.to(loss.device) # make sure to reside in the same device
1557
+
1558
+ if not return_dict:
1559
+ output = (logits,) + outputs[1:]
1560
+ if output_router_logits:
1561
+ output = (aux_loss,) + output
1562
+ return (loss,) + output if loss is not None else output
1563
+
1564
+ return MoeCausalLMOutputWithPast(
1565
+ loss=loss,
1566
+ aux_loss=aux_loss,
1567
+ logits=logits,
1568
+ past_key_values=outputs.past_key_values,
1569
+ hidden_states=outputs.hidden_states,
1570
+ attentions=outputs.attentions,
1571
+ router_logits=outputs.router_logits,
1572
+ )
1573
+
1574
+ def prepare_inputs_for_generation(
1575
+ self,
1576
+ input_ids,
1577
+ past_key_values=None,
1578
+ attention_mask=None,
1579
+ inputs_embeds=None,
1580
+ cache_position=None,
1581
+ use_cache=True,
1582
+ **kwargs,
1583
+ ):
1584
+ past_length = 0
1585
+
1586
+ # ##### by own
1587
+ if past_key_values is not None:
1588
+ if isinstance(past_key_values,Cache):
1589
+ # Past key values are always initialized with a `Cache` object -> no need for if-else anymore
1590
+ past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length()
1591
+ max_cache_length = (
1592
+ torch.tensor(past_key_values.get_max_length(), device=input_ids.device)
1593
+ if past_key_values.get_max_length() is not None
1594
+ else None
1595
+ )
1596
+ cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length)
1597
+ else:
1598
+ cache_length=past_length=past_key_values[0][0].shape[2]
1599
+ max_cache_length=None
1600
+ # # #####
1601
+ # Omit tokens covered by past_key_values
1602
+ # if past_key_values is not None:
1603
+ # # Past key values are always initialized with a `Cache` object -> no need for if-else anymore
1604
+ # past_length = cache_position[0] if cache_position is not None else past_key_values.get_seq_length()
1605
+ # max_cache_length = (
1606
+ # torch.tensor(past_key_values.get_max_length(), device=input_ids.device)
1607
+ # if past_key_values.get_max_length() is not None
1608
+ # else None
1609
+ # )
1610
+ # cache_length = past_length if max_cache_length is None else torch.min(max_cache_length, past_length)
1611
+
1612
+ # Keep only the unprocessed tokens:
1613
+ # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
1614
+ # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
1615
+ # input)
1616
+ if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
1617
+ input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
1618
+ # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
1619
+ # input_ids based on the past_length.
1620
+ elif past_length < input_ids.shape[1]:
1621
+ input_ids = input_ids[:, past_length:]
1622
+ # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
1623
+
1624
+ # If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
1625
+ if (
1626
+ max_cache_length is not None
1627
+ and attention_mask is not None
1628
+ and cache_length + input_ids.shape[1] > max_cache_length
1629
+ ):
1630
+ attention_mask = attention_mask[:, -max_cache_length:]
1631
+
1632
+ position_ids = kwargs.get("position_ids", None)
1633
+ if attention_mask is not None and position_ids is None:
1634
+ # create position_ids on the fly for batch generation
1635
+ position_ids = attention_mask.long().cumsum(-1) - 1
1636
+ position_ids.masked_fill_(attention_mask == 0, 1)
1637
+ if past_key_values:
1638
+ position_ids = position_ids[:, -input_ids.shape[1] :]
1639
+
1640
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
1641
+ if inputs_embeds is not None and past_length == 0:
1642
+ model_inputs = {"inputs_embeds": inputs_embeds}
1643
+ else:
1644
+ model_inputs = {"input_ids": input_ids}
1645
+
1646
+ input_length = position_ids.shape[-1] if position_ids is not None else input_ids.shape[-1]
1647
+ if cache_position is None:
1648
+ cache_position = torch.arange(past_length, past_length + input_length, device=input_ids.device)
1649
+ elif use_cache:
1650
+ cache_position = cache_position[-input_length:]
1651
+
1652
+ model_inputs.update(
1653
+ {
1654
+ "position_ids": position_ids,
1655
+ "past_key_values": past_key_values,
1656
+ "use_cache": use_cache,
1657
+ "attention_mask": attention_mask,
1658
+ "cache_position": cache_position,
1659
+ }
1660
+ )
1661
+ return model_inputs
1662
+
1663
+ @staticmethod
1664
+ def _reorder_cache(past_key_values, beam_idx):
1665
+ reordered_past = ()
1666
+ for layer_past in past_key_values:
1667
+ reordered_past += (
1668
+ tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past),
1669
+ )
1670
+ return reordered_past
1671
+
1672
+
1673
+