Upload DogeForCausalLM
Browse files- config.json +8 -10
- configuration_doge.py +16 -24
- model.safetensors +2 -2
- modeling_doge.py +254 -295
config.json
CHANGED
@@ -1,33 +1,31 @@
|
|
1 |
{
|
2 |
-
"_name_or_path": "./
|
3 |
"architectures": [
|
4 |
"DogeForCausalLM"
|
5 |
],
|
|
|
6 |
"auto_map": {
|
7 |
"AutoConfig": "configuration_doge.DogeConfig",
|
8 |
"AutoModelForCausalLM": "modeling_doge.DogeForCausalLM"
|
9 |
},
|
10 |
"bos_token_id": 1,
|
11 |
"eos_token_id": 2,
|
|
|
12 |
"hidden_act": "silu",
|
13 |
"hidden_bias": false,
|
14 |
"hidden_dropout": 0.0,
|
15 |
"hidden_size": 512,
|
16 |
"initializer_range": 0.02,
|
17 |
-
"inner_values_retrieval_size": 128,
|
18 |
"intermediate_size": 2048,
|
19 |
-
"
|
|
|
20 |
"model_type": "doge",
|
21 |
"num_attention_heads": 4,
|
22 |
-
"num_cdmmoe_experts":
|
23 |
-
"num_cdmmoe_experts_per_head":
|
24 |
-
"num_cdmmoe_heads":
|
25 |
"num_hidden_layers": 8,
|
26 |
-
"num_inner_value_heads": 2,
|
27 |
-
"num_inner_values": 4,
|
28 |
-
"num_value_per_head": 2,
|
29 |
"pad_token_id": 0,
|
30 |
-
"private_expert_retrieval_size": 256,
|
31 |
"rms_norm_eps": 1e-06,
|
32 |
"rope_scaling": null,
|
33 |
"rope_theta": 10000.0,
|
|
|
1 |
{
|
2 |
+
"_name_or_path": "./results/Doge-60M",
|
3 |
"architectures": [
|
4 |
"DogeForCausalLM"
|
5 |
],
|
6 |
+
"attention_dropout": 0.0,
|
7 |
"auto_map": {
|
8 |
"AutoConfig": "configuration_doge.DogeConfig",
|
9 |
"AutoModelForCausalLM": "modeling_doge.DogeForCausalLM"
|
10 |
},
|
11 |
"bos_token_id": 1,
|
12 |
"eos_token_id": 2,
|
13 |
+
"expert_retrieval_size": 256,
|
14 |
"hidden_act": "silu",
|
15 |
"hidden_bias": false,
|
16 |
"hidden_dropout": 0.0,
|
17 |
"hidden_size": 512,
|
18 |
"initializer_range": 0.02,
|
|
|
19 |
"intermediate_size": 2048,
|
20 |
+
"is_moe": false,
|
21 |
+
"max_position_embeddings": 2048,
|
22 |
"model_type": "doge",
|
23 |
"num_attention_heads": 4,
|
24 |
+
"num_cdmmoe_experts": 4096,
|
25 |
+
"num_cdmmoe_experts_per_head": 8,
|
26 |
+
"num_cdmmoe_heads": 4,
|
27 |
"num_hidden_layers": 8,
|
|
|
|
|
|
|
28 |
"pad_token_id": 0,
|
|
|
29 |
"rms_norm_eps": 1e-06,
|
30 |
"rope_scaling": null,
|
31 |
"rope_theta": 10000.0,
|
configuration_doge.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
# coding=utf-8
|
2 |
-
# Copyright 2024 Jingze Shi and the HuggingFace Inc. team.
|
3 |
#
|
4 |
# This code is based on the Wonderful Matrices paper implementation.
|
5 |
#
|
6 |
-
# https://arxiv.org/abs/
|
7 |
#
|
8 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
9 |
# you may not use this file except in compliance with the License.
|
@@ -46,7 +46,7 @@ class DogeConfig(PretrainedConfig):
|
|
46 |
Dropout probability for each sequence transformation and state transformation module.
|
47 |
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
48 |
The non-linear activation function (function or string) in the decoder.
|
49 |
-
max_position_embeddings (`int`, *optional*, defaults to
|
50 |
The maximum sequence length that this model might ever be used with.
|
51 |
rope_theta (`float`, *optional*, defaults to 10000.0):
|
52 |
The base period of the RoPE embeddings.
|
@@ -104,22 +104,18 @@ class DogeConfig(PretrainedConfig):
|
|
104 |
Whether to tie weight embeddings
|
105 |
num_attention_heads (`int`, *optional*, defaults to 8):
|
106 |
Number of attention heads for each attention layer in the Transformer decoder.
|
107 |
-
|
108 |
-
|
109 |
-
|
110 |
-
|
111 |
-
num_value_per_head (`int`, *optional*, defaults to 4):
|
112 |
-
Number of values per head, can't be greater than `num_inner_values`.
|
113 |
-
inner_values_retrieval_size (`int`, *optional*, defaults to 128):
|
114 |
-
Dimension of the inner values retrieval states for each attention layer in the Transformer decoder
|
115 |
-
private_expert_retrieval_size (`int`, *optional*, defaults to 256):
|
116 |
-
Dimension of the Private Expert retrieval states for the Cross Domain Mixture of Experts.
|
117 |
num_cdmmoe_experts (`int`, *optional*, defaults to 4096):
|
118 |
Number of Private Experts for the Cross Domain Mixture of Experts.
|
119 |
num_cdmmoe_heads (`int`, *optional*, defaults to 4):
|
120 |
Number of heads of Private Experts for the Cross Domain Mixture of Experts.
|
121 |
num_cdmmoe_experts_per_head (`int`, *optional*, defaults to 8):
|
122 |
Number of Private Experts per head for the Cross Domain Mixture of Experts.
|
|
|
|
|
123 |
"""
|
124 |
|
125 |
model_type = "doge"
|
@@ -134,7 +130,7 @@ class DogeConfig(PretrainedConfig):
|
|
134 |
hidden_bias=False,
|
135 |
hidden_dropout=0.0,
|
136 |
hidden_act="silu",
|
137 |
-
max_position_embeddings=
|
138 |
rope_theta=10000.0,
|
139 |
rope_scaling=None,
|
140 |
initializer_range=0.02,
|
@@ -145,14 +141,12 @@ class DogeConfig(PretrainedConfig):
|
|
145 |
eos_token_id=2,
|
146 |
tie_word_embeddings=False,
|
147 |
num_attention_heads=8,
|
148 |
-
|
149 |
-
|
150 |
-
num_value_per_head=4,
|
151 |
-
inner_values_retrieval_size=128,
|
152 |
-
private_expert_retrieval_size=256,
|
153 |
num_cdmmoe_experts=4096,
|
154 |
num_cdmmoe_heads=4,
|
155 |
num_cdmmoe_experts_per_head=8,
|
|
|
156 |
**kwargs,
|
157 |
):
|
158 |
self.vocab_size = vocab_size
|
@@ -173,14 +167,12 @@ class DogeConfig(PretrainedConfig):
|
|
173 |
self.eos_token_id = eos_token_id
|
174 |
self.tie_word_embeddings = tie_word_embeddings
|
175 |
self.num_attention_heads = num_attention_heads
|
176 |
-
self.
|
177 |
-
self.
|
178 |
-
self.num_value_per_head = num_value_per_head
|
179 |
-
self.inner_values_retrieval_size = inner_values_retrieval_size
|
180 |
-
self.private_expert_retrieval_size = private_expert_retrieval_size
|
181 |
self.num_cdmmoe_experts = num_cdmmoe_experts
|
182 |
self.num_cdmmoe_heads = num_cdmmoe_heads
|
183 |
self.num_cdmmoe_experts_per_head = num_cdmmoe_experts_per_head
|
|
|
184 |
|
185 |
# Validate the correctness of rotary position embeddings parameters
|
186 |
# BC: if there is a 'type' field, copy it it to 'rope_type'.
|
|
|
1 |
# coding=utf-8
|
2 |
+
# Copyright 2024 Jingze Shi and the HuggingFace Inc. team. All rights reserved.
|
3 |
#
|
4 |
# This code is based on the Wonderful Matrices paper implementation.
|
5 |
#
|
6 |
+
# https://arxiv.org/abs/2412.11834
|
7 |
#
|
8 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
9 |
# you may not use this file except in compliance with the License.
|
|
|
46 |
Dropout probability for each sequence transformation and state transformation module.
|
47 |
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
48 |
The non-linear activation function (function or string) in the decoder.
|
49 |
+
max_position_embeddings (`int`, *optional*, defaults to 2048):
|
50 |
The maximum sequence length that this model might ever be used with.
|
51 |
rope_theta (`float`, *optional*, defaults to 10000.0):
|
52 |
The base period of the RoPE embeddings.
|
|
|
104 |
Whether to tie weight embeddings
|
105 |
num_attention_heads (`int`, *optional*, defaults to 8):
|
106 |
Number of attention heads for each attention layer in the Transformer decoder.
|
107 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
108 |
+
The dropout ratio for the attention probabilities.
|
109 |
+
is_moe (`bool`, *optional*, defaults to `False`):
|
110 |
+
Whether to use the Cross Domain Mixture of Experts, if `True`, the MoE will inherit the MLP to initialize
|
|
|
|
|
|
|
|
|
|
|
|
|
111 |
num_cdmmoe_experts (`int`, *optional*, defaults to 4096):
|
112 |
Number of Private Experts for the Cross Domain Mixture of Experts.
|
113 |
num_cdmmoe_heads (`int`, *optional*, defaults to 4):
|
114 |
Number of heads of Private Experts for the Cross Domain Mixture of Experts.
|
115 |
num_cdmmoe_experts_per_head (`int`, *optional*, defaults to 8):
|
116 |
Number of Private Experts per head for the Cross Domain Mixture of Experts.
|
117 |
+
expert_retrieval_size (`int`, *optional*, defaults to 256):
|
118 |
+
Dimension of the Expert retrieval states for the Cross Domain Mixture of Experts.
|
119 |
"""
|
120 |
|
121 |
model_type = "doge"
|
|
|
130 |
hidden_bias=False,
|
131 |
hidden_dropout=0.0,
|
132 |
hidden_act="silu",
|
133 |
+
max_position_embeddings=2048,
|
134 |
rope_theta=10000.0,
|
135 |
rope_scaling=None,
|
136 |
initializer_range=0.02,
|
|
|
141 |
eos_token_id=2,
|
142 |
tie_word_embeddings=False,
|
143 |
num_attention_heads=8,
|
144 |
+
attention_dropout=0.0,
|
145 |
+
is_moe=False,
|
|
|
|
|
|
|
146 |
num_cdmmoe_experts=4096,
|
147 |
num_cdmmoe_heads=4,
|
148 |
num_cdmmoe_experts_per_head=8,
|
149 |
+
expert_retrieval_size=256,
|
150 |
**kwargs,
|
151 |
):
|
152 |
self.vocab_size = vocab_size
|
|
|
167 |
self.eos_token_id = eos_token_id
|
168 |
self.tie_word_embeddings = tie_word_embeddings
|
169 |
self.num_attention_heads = num_attention_heads
|
170 |
+
self.attention_dropout = attention_dropout
|
171 |
+
self.is_moe = is_moe
|
|
|
|
|
|
|
172 |
self.num_cdmmoe_experts = num_cdmmoe_experts
|
173 |
self.num_cdmmoe_heads = num_cdmmoe_heads
|
174 |
self.num_cdmmoe_experts_per_head = num_cdmmoe_experts_per_head
|
175 |
+
self.expert_retrieval_size = expert_retrieval_size
|
176 |
|
177 |
# Validate the correctness of rotary position embeddings parameters
|
178 |
# BC: if there is a 'type' field, copy it it to 'rope_type'.
|
model.safetensors
CHANGED
@@ -1,3 +1,3 @@
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
-
oid sha256:
|
3 |
-
size
|
|
|
1 |
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:26d80cdf90d4f053299b962b1ede76f0fe30ed31ebcb95e5dbd730ce23ffd36a
|
3 |
+
size 268580408
|
modeling_doge.py
CHANGED
@@ -1,9 +1,9 @@
|
|
1 |
# coding=utf-8
|
2 |
-
# Copyright 2024 Jingze Shi and the HuggingFace Inc. team.
|
3 |
#
|
4 |
# This code is based on the Wonderful Matrices paper implementation.
|
5 |
#
|
6 |
-
# https://arxiv.org/abs/
|
7 |
#
|
8 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
9 |
# you may not use this file except in compliance with the License.
|
@@ -39,16 +39,15 @@ from transformers.modeling_utils import PreTrainedModel
|
|
39 |
from transformers.utils import (
|
40 |
add_start_docstrings,
|
41 |
add_start_docstrings_to_model_forward,
|
42 |
-
# is_einx_available,
|
43 |
logging,
|
44 |
replace_return_docstrings,
|
45 |
)
|
46 |
from .configuration_doge import DogeConfig
|
47 |
|
48 |
-
|
49 |
-
|
50 |
-
|
51 |
-
|
52 |
|
53 |
|
54 |
logger = logging.get_logger(__name__)
|
@@ -76,6 +75,18 @@ class RMSNorm(nn.Module):
|
|
76 |
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
77 |
|
78 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
class RotaryEmbedding(nn.Module):
|
80 |
def __init__(self, config: Optional[DogeConfig] = None):
|
81 |
super().__init__()
|
@@ -172,8 +183,8 @@ def apply_QK_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
|
172 |
return q_embed, k_embed
|
173 |
|
174 |
|
175 |
-
class
|
176 |
-
"""
|
177 |
|
178 |
def __init__(self, config: DogeConfig, layer_idx: Optional[int] = None):
|
179 |
super().__init__()
|
@@ -189,15 +200,10 @@ class DogeInnerFuncAttn(nn.Module):
|
|
189 |
|
190 |
self.hidden_dim = config.hidden_size
|
191 |
self.num_attention_heads = config.num_attention_heads
|
192 |
-
|
193 |
-
# for accuracy of attention scores, we do not use GQA
|
194 |
self.attention_head_dim = self.hidden_dim // self.num_attention_heads
|
195 |
-
self.num_inner_values = config.num_inner_values
|
196 |
-
self.num_inner_value_heads = config.num_inner_value_heads
|
197 |
-
self.num_value_per_head = config.num_value_per_head
|
198 |
-
self.inner_values_retrieval_dim = config.inner_values_retrieval_size
|
199 |
|
200 |
-
# Q
|
201 |
self.q_proj = nn.Linear(
|
202 |
self.hidden_dim,
|
203 |
self.num_attention_heads * self.attention_head_dim,
|
@@ -208,157 +214,26 @@ class DogeInnerFuncAttn(nn.Module):
|
|
208 |
self.num_attention_heads * self.attention_head_dim,
|
209 |
bias=config.hidden_bias,
|
210 |
)
|
211 |
-
|
212 |
# dynamic mask for the QK^T attention score matrix
|
213 |
-
self.
|
214 |
-
torch.
|
215 |
)
|
216 |
-
|
217 |
-
# queries and keys for retrieval V
|
218 |
-
self.v_queries = nn.Linear(
|
219 |
self.hidden_dim,
|
220 |
-
self.
|
221 |
bias=config.hidden_bias,
|
222 |
)
|
223 |
-
self.
|
224 |
-
torch.zeros(
|
225 |
-
self.num_inner_value_heads,
|
226 |
-
self.inner_values_retrieval_dim,
|
227 |
-
self.num_inner_values,
|
228 |
-
)
|
229 |
-
)
|
230 |
-
|
231 |
-
# V for inner function
|
232 |
-
self.v_embed = nn.Embedding(
|
233 |
-
self.num_inner_values,
|
234 |
self.hidden_dim,
|
|
|
|
|
235 |
)
|
236 |
-
|
237 |
self.o_proj = nn.Linear(
|
238 |
self.hidden_dim,
|
239 |
self.hidden_dim,
|
240 |
bias=config.hidden_bias,
|
241 |
)
|
242 |
|
243 |
-
def _update_causal_mask(
|
244 |
-
self,
|
245 |
-
attention_mask: torch.Tensor = None,
|
246 |
-
input_tensor: torch.Tensor = None,
|
247 |
-
cache_position: torch.Tensor = None,
|
248 |
-
past_key_values: Cache = None,
|
249 |
-
output_attentions: bool = False,
|
250 |
-
):
|
251 |
-
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
252 |
-
using_static_cache = isinstance(past_key_values, StaticCache)
|
253 |
-
|
254 |
-
dtype, device = input_tensor.dtype, input_tensor.device
|
255 |
-
sequence_length = input_tensor.shape[1]
|
256 |
-
if using_static_cache:
|
257 |
-
target_length = past_key_values.get_max_cache_shape()
|
258 |
-
else:
|
259 |
-
target_length = (
|
260 |
-
attention_mask.shape[-1]
|
261 |
-
if isinstance(attention_mask, torch.Tensor)
|
262 |
-
else past_seen_tokens + sequence_length + 1
|
263 |
-
)
|
264 |
-
|
265 |
-
# in case the provided `attention` mask is 2D, we generate a causal mask here (4D).
|
266 |
-
causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position_and_dynamic_mask(
|
267 |
-
attention_mask=attention_mask,
|
268 |
-
dynamic_mask=self.dynamic_mask,
|
269 |
-
sequence_length=sequence_length,
|
270 |
-
target_length=target_length,
|
271 |
-
dtype=dtype,
|
272 |
-
device=device,
|
273 |
-
cache_position=cache_position,
|
274 |
-
batch_size=input_tensor.shape[0],
|
275 |
-
)
|
276 |
-
|
277 |
-
return causal_mask
|
278 |
-
|
279 |
-
@staticmethod
|
280 |
-
def _prepare_4d_causal_attention_mask_with_cache_position_and_dynamic_mask(
|
281 |
-
attention_mask: torch.Tensor = None,
|
282 |
-
dynamic_mask: torch.Tensor = None,
|
283 |
-
sequence_length: int = None,
|
284 |
-
target_length: int = None,
|
285 |
-
dtype: torch.dtype = None,
|
286 |
-
device: torch.device = None,
|
287 |
-
cache_position: torch.Tensor = None,
|
288 |
-
batch_size: int = None,
|
289 |
-
**kwargs,
|
290 |
-
):
|
291 |
-
"""
|
292 |
-
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
|
293 |
-
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
|
294 |
-
|
295 |
-
Args:
|
296 |
-
attention_mask (`torch.Tensor`):
|
297 |
-
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
|
298 |
-
`(batch_size, 1, query_length, key_value_length)`.
|
299 |
-
dynamic_mask (`torch.Tensor`):
|
300 |
-
A 2D dynamic mask of shape `(num_heads, max_position_embeddings)`.
|
301 |
-
sequence_length (`int`):
|
302 |
-
The sequence length being processed.
|
303 |
-
target_length (`int`):
|
304 |
-
The target length: when generating with static cache, the mask should be as long as the static cache,
|
305 |
-
to account for the 0 padding, the part of the cache that is not filled yet.
|
306 |
-
dtype (`torch.dtype`):
|
307 |
-
The dtype to use for the 4D attention mask.
|
308 |
-
device (`torch.device`):
|
309 |
-
The device to plcae the 4D attention mask on.
|
310 |
-
cache_position (`torch.Tensor`):
|
311 |
-
Indices depicting the position of the input sequence tokens in the sequence.
|
312 |
-
batch_size (`torch.Tensor`):
|
313 |
-
Batch size.
|
314 |
-
"""
|
315 |
-
if attention_mask is not None and attention_mask.dim() == 4:
|
316 |
-
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
|
317 |
-
causal_mask = attention_mask
|
318 |
-
else:
|
319 |
-
num_heads = 1 if dynamic_mask is None else dynamic_mask.size(0)
|
320 |
-
min_dtype = torch.finfo(dtype).min
|
321 |
-
causal_mask = torch.full(
|
322 |
-
(sequence_length, target_length),
|
323 |
-
fill_value=min_dtype,
|
324 |
-
dtype=dtype,
|
325 |
-
device=device,
|
326 |
-
)
|
327 |
-
if sequence_length != 1:
|
328 |
-
causal_mask = torch.triu(causal_mask, diagonal=1)
|
329 |
-
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
330 |
-
causal_mask = causal_mask[None, None, :, :].expand(batch_size, num_heads, -1, -1)
|
331 |
-
if attention_mask is not None:
|
332 |
-
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
333 |
-
mask_length = attention_mask.shape[-1]
|
334 |
-
attention_mask = attention_mask[:, None, None, :].expand(-1, num_heads, 1, -1)
|
335 |
-
if dynamic_mask is not None:
|
336 |
-
dynamic_mask = dynamic_mask[None, :, None, :mask_length].expand(batch_size, -1, 1, -1)
|
337 |
-
attention_mask = attention_mask.clone() * dynamic_mask
|
338 |
-
|
339 |
-
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask
|
340 |
-
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
341 |
-
padding_mask == 0, min_dtype
|
342 |
-
)
|
343 |
-
|
344 |
-
return causal_mask
|
345 |
-
|
346 |
-
def inner_func(
|
347 |
-
self,
|
348 |
-
hidden_states: torch.Tensor,
|
349 |
-
) -> torch.Tensor:
|
350 |
-
"""
|
351 |
-
Each value can share weights with other values to increase the expressive power
|
352 |
-
"""
|
353 |
-
bsz, seq_len, _ = hidden_states.shape
|
354 |
-
|
355 |
-
v_queries = self.v_queries(hidden_states)
|
356 |
-
v_queries = v_queries.view(bsz, seq_len, self.num_inner_value_heads, -1).transpose(1, 2)
|
357 |
-
sim = torch.matmul(v_queries, self.v_keys).transpose(1, 2)
|
358 |
-
v_embed = self.v_embed(sim.topk(k=self.num_value_per_head, dim=-1).indices)
|
359 |
-
v = hidden_states * v_embed.sum(dim=-2).sum(dim=-2)
|
360 |
-
return v
|
361 |
-
|
362 |
def forward(
|
363 |
self,
|
364 |
hidden_states: torch.Tensor,
|
@@ -369,24 +244,24 @@ class DogeInnerFuncAttn(nn.Module):
|
|
369 |
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
370 |
**kwargs,
|
371 |
) -> Tuple[torch.Tensor, Optional[Cache]]:
|
372 |
-
bsz,
|
373 |
|
374 |
query_states = self.q_proj(hidden_states)
|
375 |
key_states = self.k_proj(hidden_states)
|
376 |
-
value_states = self.
|
377 |
|
378 |
-
query_states = query_states.view(bsz,
|
379 |
1, 2
|
380 |
)
|
381 |
-
key_states = key_states.view(bsz,
|
382 |
1, 2
|
383 |
)
|
384 |
-
value_states = value_states.view(bsz,
|
385 |
1, 2
|
386 |
)
|
387 |
|
388 |
cos, sin = position_embeddings
|
389 |
-
query_states,
|
390 |
|
391 |
if past_key_value is not None:
|
392 |
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
@@ -397,38 +272,101 @@ class DogeInnerFuncAttn(nn.Module):
|
|
397 |
attn_weights = torch.matmul(query_states, key_states.transpose(-1, -2)) / math.sqrt(self.attention_head_dim)
|
398 |
|
399 |
# add mask to attention scores
|
400 |
-
|
401 |
-
|
402 |
-
|
|
|
|
|
|
|
403 |
|
404 |
# upcast attention scores to fp32
|
405 |
attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
|
|
406 |
|
407 |
# apply attention scores to value states
|
408 |
attn_output = torch.matmul(attn_weights, value_states)
|
409 |
|
410 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
411 |
-
attn_output = attn_output.reshape(bsz,
|
412 |
attn_output = self.o_proj(attn_output)
|
413 |
|
414 |
return attn_output, past_key_value
|
415 |
|
416 |
|
417 |
-
class
|
418 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
419 |
|
420 |
def __init__(self, config: DogeConfig):
|
421 |
super().__init__()
|
422 |
self.hidden_dim = config.hidden_size
|
423 |
-
self.act_fn = ACT2FN[config.hidden_act]
|
424 |
self.intermediate_dim = config.intermediate_size
|
|
|
425 |
|
426 |
-
self.
|
427 |
-
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
# cross domain
|
432 |
self.up_proj = nn.Linear(
|
433 |
self.hidden_dim,
|
434 |
self.intermediate_dim,
|
@@ -440,24 +378,46 @@ class DogeCDMoE(nn.Module):
|
|
440 |
bias=config.hidden_bias,
|
441 |
)
|
442 |
|
443 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
444 |
self.queries = nn.Linear(
|
445 |
self.hidden_dim,
|
446 |
-
self.num_cdmmoe_heads * self.
|
447 |
bias=False,
|
448 |
)
|
449 |
-
self.num_keys = int(math.sqrt(self.num_cdmmoe_experts))
|
450 |
self.keys = nn.Parameter(
|
451 |
torch.zeros(
|
452 |
self.num_cdmmoe_heads,
|
453 |
self.num_keys,
|
454 |
2,
|
455 |
-
self.
|
456 |
)
|
457 |
)
|
458 |
|
459 |
-
#
|
460 |
-
self.down_embed
|
461 |
self.num_cdmmoe_experts,
|
462 |
self.hidden_dim,
|
463 |
)
|
@@ -471,7 +431,7 @@ class DogeCDMoE(nn.Module):
|
|
471 |
self,
|
472 |
hidden_states: torch.Tensor,
|
473 |
**kwargs,
|
474 |
-
) ->
|
475 |
bsz, seq_len, _ = hidden_states.shape
|
476 |
|
477 |
# get similarity with queries and keys
|
@@ -479,7 +439,7 @@ class DogeCDMoE(nn.Module):
|
|
479 |
queries = queries.view(bsz, seq_len, 2, self.num_cdmmoe_heads, -1).permute(2, 0, 1, 3, 4)
|
480 |
sim = torch.einsum("p b t h n, h k p n -> p b t h k", queries, self.keys)
|
481 |
|
482 |
-
# get
|
483 |
(scores_x, scores_y), (indices_x, indices_y) = sim.topk(self.num_cdmmoe_experts_per_head, dim=-1)
|
484 |
if einx_add is not None:
|
485 |
all_scores = einx_add("... i, ... j -> ... (i j)", scores_x, scores_y)
|
@@ -491,17 +451,14 @@ class DogeCDMoE(nn.Module):
|
|
491 |
all_indices = all_indices.view(*indices_x.shape[:-1], -1)
|
492 |
scores, pk_indices = all_scores.topk(self.num_cdmmoe_experts_per_head, dim=-1)
|
493 |
indices = all_indices.gather(-1, pk_indices)
|
494 |
-
|
495 |
-
# get related expert embeddings based on indices
|
496 |
down_embed = self.down_embed(indices)
|
497 |
up_embed = self.up_embed(indices)
|
498 |
|
499 |
-
#
|
500 |
-
experts_weights =
|
|
|
501 |
experts_states = torch.einsum("b t h k, b t h k d -> b t d", experts_weights, up_embed)
|
502 |
-
|
503 |
-
# mix with shared parameters of cross domain
|
504 |
-
hidden_states = self.down_proj(self.act_fn(self.up_proj(hidden_states)))
|
505 |
hidden_states = hidden_states + experts_states
|
506 |
return hidden_states
|
507 |
|
@@ -511,10 +468,13 @@ class DogeDecoderLayer(nn.Module):
|
|
511 |
super().__init__()
|
512 |
self.hidden_dropout = config.hidden_dropout
|
513 |
|
514 |
-
self.
|
515 |
-
self.attn =
|
516 |
-
self.
|
517 |
-
|
|
|
|
|
|
|
518 |
|
519 |
def forward(
|
520 |
self,
|
@@ -553,7 +513,7 @@ class DogeDecoderLayer(nn.Module):
|
|
553 |
|
554 |
# sequence transformation
|
555 |
residual = hidden_states
|
556 |
-
hidden_states = self.
|
557 |
hidden_states, present_key_value = self.attn(
|
558 |
hidden_states=hidden_states,
|
559 |
attention_mask=attention_mask,
|
@@ -565,14 +525,14 @@ class DogeDecoderLayer(nn.Module):
|
|
565 |
)
|
566 |
self_attn_weights = None
|
567 |
hidden_states = F.dropout(hidden_states, p=self.hidden_dropout, training=self.training)
|
568 |
-
hidden_states = residual
|
569 |
|
570 |
# state transformation
|
571 |
residual = hidden_states
|
572 |
-
hidden_states = self.
|
573 |
hidden_states = self.feed_forward(hidden_states)
|
574 |
hidden_states = F.dropout(hidden_states, p=self.hidden_dropout, training=self.training)
|
575 |
-
hidden_states = residual
|
576 |
|
577 |
outputs = (hidden_states,)
|
578 |
|
@@ -592,6 +552,7 @@ class DogePreTrainedModel(PreTrainedModel):
|
|
592 |
supports_gradient_checkpointing = True
|
593 |
_no_split_modules = ["DogeDecoderLayer"]
|
594 |
_skip_keys_device_placement = ["past_key_values"]
|
|
|
595 |
_supports_cache_class = True
|
596 |
_supports_quantized_cache = True
|
597 |
_supports_static_cache = True
|
@@ -765,9 +726,9 @@ class DogeModel(DogePreTrainedModel):
|
|
765 |
if position_ids is None:
|
766 |
position_ids = cache_position.unsqueeze(0)
|
767 |
|
768 |
-
|
769 |
-
|
770 |
-
|
771 |
hidden_states = inputs_embeds
|
772 |
|
773 |
# create position embeddings to be shared across the decoder layers
|
@@ -776,6 +737,7 @@ class DogeModel(DogePreTrainedModel):
|
|
776 |
# decoder layers
|
777 |
all_hidden_states = () if output_hidden_states else None
|
778 |
all_self_attns = () if output_attentions else None
|
|
|
779 |
|
780 |
for decoder_layer in self.layers:
|
781 |
if output_hidden_states:
|
@@ -785,7 +747,7 @@ class DogeModel(DogePreTrainedModel):
|
|
785 |
layer_outputs = self._gradient_checkpointing_func(
|
786 |
decoder_layer.__call__,
|
787 |
hidden_states,
|
788 |
-
|
789 |
position_ids,
|
790 |
past_key_values,
|
791 |
output_attentions,
|
@@ -796,7 +758,7 @@ class DogeModel(DogePreTrainedModel):
|
|
796 |
else:
|
797 |
layer_outputs = decoder_layer(
|
798 |
hidden_states,
|
799 |
-
attention_mask=
|
800 |
position_ids=position_ids,
|
801 |
past_key_value=past_key_values,
|
802 |
output_attentions=output_attentions,
|
@@ -833,100 +795,97 @@ class DogeModel(DogePreTrainedModel):
|
|
833 |
attentions=all_self_attns,
|
834 |
)
|
835 |
|
836 |
-
|
837 |
-
|
838 |
-
|
839 |
-
|
840 |
-
|
841 |
-
|
842 |
-
|
843 |
-
|
844 |
-
|
845 |
-
|
846 |
-
|
847 |
-
|
848 |
-
|
849 |
-
|
850 |
-
|
851 |
-
|
852 |
-
|
853 |
-
|
854 |
-
|
855 |
-
|
856 |
-
|
857 |
-
|
858 |
-
|
859 |
-
|
860 |
-
|
861 |
-
|
862 |
-
|
863 |
-
|
864 |
-
|
865 |
-
|
866 |
-
|
867 |
-
|
868 |
-
|
869 |
-
|
870 |
-
|
871 |
-
|
872 |
-
|
873 |
-
|
874 |
-
|
875 |
-
|
876 |
-
|
877 |
-
|
878 |
-
|
879 |
-
|
880 |
-
|
881 |
-
|
882 |
-
|
883 |
-
|
884 |
-
|
885 |
-
|
886 |
-
|
887 |
-
|
888 |
-
|
889 |
-
|
890 |
-
|
891 |
-
|
892 |
-
|
893 |
-
|
894 |
-
|
895 |
-
|
896 |
-
|
897 |
-
|
898 |
-
|
899 |
-
|
900 |
-
|
901 |
-
|
902 |
-
|
903 |
-
|
904 |
-
|
905 |
-
|
906 |
-
|
907 |
-
|
908 |
-
|
909 |
-
|
910 |
-
|
911 |
-
|
912 |
-
|
913 |
-
|
914 |
-
|
915 |
-
|
916 |
-
|
917 |
-
|
918 |
-
|
919 |
-
|
920 |
-
|
921 |
-
|
922 |
-
|
923 |
-
|
924 |
-
|
925 |
-
|
926 |
-
|
927 |
-
# )
|
928 |
-
|
929 |
-
# return causal_mask
|
930 |
|
931 |
|
932 |
class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
|
|
|
1 |
# coding=utf-8
|
2 |
+
# Copyright 2024 Jingze Shi and the HuggingFace Inc. team. All rights reserved.
|
3 |
#
|
4 |
# This code is based on the Wonderful Matrices paper implementation.
|
5 |
#
|
6 |
+
# https://arxiv.org/abs/2412.11834
|
7 |
#
|
8 |
# Licensed under the Apache License, Version 2.0 (the "License");
|
9 |
# you may not use this file except in compliance with the License.
|
|
|
39 |
from transformers.utils import (
|
40 |
add_start_docstrings,
|
41 |
add_start_docstrings_to_model_forward,
|
|
|
42 |
logging,
|
43 |
replace_return_docstrings,
|
44 |
)
|
45 |
from .configuration_doge import DogeConfig
|
46 |
|
47 |
+
try:
|
48 |
+
from einx import add as einx_add
|
49 |
+
except ImportError:
|
50 |
+
einx_add = None
|
51 |
|
52 |
|
53 |
logger = logging.get_logger(__name__)
|
|
|
75 |
return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
|
76 |
|
77 |
|
78 |
+
class Residual(nn.Module):
|
79 |
+
def __init__(self, hidden_size):
|
80 |
+
super().__init__()
|
81 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
82 |
+
|
83 |
+
def forward(self, residual_states, hidden_states):
|
84 |
+
return self.weight * residual_states + hidden_states
|
85 |
+
|
86 |
+
def extra_repr(self):
|
87 |
+
return f"{tuple(self.weight.shape)}"
|
88 |
+
|
89 |
+
|
90 |
class RotaryEmbedding(nn.Module):
|
91 |
def __init__(self, config: Optional[DogeConfig] = None):
|
92 |
super().__init__()
|
|
|
183 |
return q_embed, k_embed
|
184 |
|
185 |
|
186 |
+
class DogeDynamicMaskAttention(nn.Module):
|
187 |
+
"""Dynamic Mask Attention from 'Wonderful Matrices' paper."""
|
188 |
|
189 |
def __init__(self, config: DogeConfig, layer_idx: Optional[int] = None):
|
190 |
super().__init__()
|
|
|
200 |
|
201 |
self.hidden_dim = config.hidden_size
|
202 |
self.num_attention_heads = config.num_attention_heads
|
203 |
+
self.attention_dropout = config.attention_dropout
|
|
|
204 |
self.attention_head_dim = self.hidden_dim // self.num_attention_heads
|
|
|
|
|
|
|
|
|
205 |
|
206 |
+
# Q K V O projections
|
207 |
self.q_proj = nn.Linear(
|
208 |
self.hidden_dim,
|
209 |
self.num_attention_heads * self.attention_head_dim,
|
|
|
214 |
self.num_attention_heads * self.attention_head_dim,
|
215 |
bias=config.hidden_bias,
|
216 |
)
|
|
|
217 |
# dynamic mask for the QK^T attention score matrix
|
218 |
+
self.A = nn.Parameter(
|
219 |
+
torch.ones(self.num_attention_heads)
|
220 |
)
|
221 |
+
self.dt_proj = nn.Linear(
|
|
|
|
|
222 |
self.hidden_dim,
|
223 |
+
self.num_attention_heads,
|
224 |
bias=config.hidden_bias,
|
225 |
)
|
226 |
+
self.v_proj = nn.Linear(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
227 |
self.hidden_dim,
|
228 |
+
self.num_attention_heads * self.attention_head_dim,
|
229 |
+
bias=config.hidden_bias,
|
230 |
)
|
|
|
231 |
self.o_proj = nn.Linear(
|
232 |
self.hidden_dim,
|
233 |
self.hidden_dim,
|
234 |
bias=config.hidden_bias,
|
235 |
)
|
236 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
237 |
def forward(
|
238 |
self,
|
239 |
hidden_states: torch.Tensor,
|
|
|
244 |
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
245 |
**kwargs,
|
246 |
) -> Tuple[torch.Tensor, Optional[Cache]]:
|
247 |
+
bsz, q_len, _ = hidden_states.shape
|
248 |
|
249 |
query_states = self.q_proj(hidden_states)
|
250 |
key_states = self.k_proj(hidden_states)
|
251 |
+
value_states = self.v_proj(hidden_states)
|
252 |
|
253 |
+
query_states = query_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(
|
254 |
1, 2
|
255 |
)
|
256 |
+
key_states = key_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(
|
257 |
1, 2
|
258 |
)
|
259 |
+
value_states = value_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(
|
260 |
1, 2
|
261 |
)
|
262 |
|
263 |
cos, sin = position_embeddings
|
264 |
+
query_states, key_states = apply_QK_rotary_pos_emb(query_states, key_states, cos, sin)
|
265 |
|
266 |
if past_key_value is not None:
|
267 |
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
|
|
272 |
attn_weights = torch.matmul(query_states, key_states.transpose(-1, -2)) / math.sqrt(self.attention_head_dim)
|
273 |
|
274 |
# add mask to attention scores
|
275 |
+
if attention_mask is not None:
|
276 |
+
dt_states = self.dt_proj(value_states.transpose(1, 2).reshape(bsz, value_states.shape[-2], -1))
|
277 |
+
dynamic_mask = torch.exp(self.A * F.softplus(dt_states)).transpose(-1, -2)
|
278 |
+
dynamic_mask = dynamic_mask < 1.0
|
279 |
+
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]].masked_fill(dynamic_mask[:, :, None, :], torch.finfo(hidden_states.dtype).min)
|
280 |
+
attn_weights = attn_weights + causal_mask
|
281 |
|
282 |
# upcast attention scores to fp32
|
283 |
attn_weights = F.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
284 |
+
attn_weights = F.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
285 |
|
286 |
# apply attention scores to value states
|
287 |
attn_output = torch.matmul(attn_weights, value_states)
|
288 |
|
289 |
attn_output = attn_output.transpose(1, 2).contiguous()
|
290 |
+
attn_output = attn_output.reshape(bsz, q_len, -1)
|
291 |
attn_output = self.o_proj(attn_output)
|
292 |
|
293 |
return attn_output, past_key_value
|
294 |
|
295 |
|
296 |
+
class DogeSdpaDynamicMaskAttn(DogeDynamicMaskAttention):
|
297 |
+
|
298 |
+
def forward(
|
299 |
+
self,
|
300 |
+
hidden_states: torch.Tensor,
|
301 |
+
attention_mask: Optional[torch.Tensor] = None,
|
302 |
+
position_ids: Optional[torch.LongTensor] = None,
|
303 |
+
past_key_value: Optional[Cache] = None,
|
304 |
+
cache_position: Optional[torch.LongTensor] = None,
|
305 |
+
position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
|
306 |
+
**kwargs,
|
307 |
+
) -> Tuple[torch.Tensor, Optional[Cache]]:
|
308 |
+
bsz, q_len, _ = hidden_states.shape
|
309 |
+
|
310 |
+
query_states = self.q_proj(hidden_states)
|
311 |
+
key_states = self.k_proj(hidden_states)
|
312 |
+
value_states = self.v_proj(hidden_states)
|
313 |
+
|
314 |
+
query_states = query_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(1, 2)
|
315 |
+
key_states = key_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(1, 2)
|
316 |
+
value_states = value_states.view(bsz, q_len, self.num_attention_heads, self.attention_head_dim).transpose(1, 2)
|
317 |
+
|
318 |
+
cos, sin = position_embeddings
|
319 |
+
query_states, key_states = apply_QK_rotary_pos_emb(query_states, key_states, cos, sin)
|
320 |
+
|
321 |
+
if past_key_value is not None:
|
322 |
+
# sin and cos are specific to RoPE models; cache_position needed for the static cache
|
323 |
+
cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
|
324 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
325 |
+
|
326 |
+
if attention_mask is not None:
|
327 |
+
dt_states = self.dt_proj(value_states.transpose(1, 2).reshape(bsz, value_states.shape[-2], -1))
|
328 |
+
dynamic_mask = torch.exp(self.A * F.softplus(dt_states)).transpose(-1, -2)
|
329 |
+
dynamic_mask = dynamic_mask < 1.0
|
330 |
+
causal_mask = attention_mask[:, :, :, : key_states.shape[-2]].masked_fill(dynamic_mask[:, :, None, :], torch.finfo(hidden_states.dtype).min)
|
331 |
+
|
332 |
+
query_states = query_states.contiguous()
|
333 |
+
key_states = key_states.contiguous()
|
334 |
+
value_states = value_states.contiguous()
|
335 |
+
|
336 |
+
attn_output = F.scaled_dot_product_attention(
|
337 |
+
query_states,
|
338 |
+
key_states,
|
339 |
+
value_states,
|
340 |
+
attn_mask=causal_mask,
|
341 |
+
dropout_p=self.attention_dropout,
|
342 |
+
)
|
343 |
+
|
344 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
345 |
+
attn_output = attn_output.view(bsz, q_len, -1)
|
346 |
+
attn_output = self.o_proj(attn_output)
|
347 |
+
|
348 |
+
return attn_output, past_key_value
|
349 |
+
|
350 |
+
|
351 |
+
DOGE_ATTENTION_CLASSES = {
|
352 |
+
"eager": DogeDynamicMaskAttention,
|
353 |
+
"sdpa": DogeSdpaDynamicMaskAttn,
|
354 |
+
}
|
355 |
+
|
356 |
+
|
357 |
+
class DogeMLP(nn.Module):
|
358 |
|
359 |
def __init__(self, config: DogeConfig):
|
360 |
super().__init__()
|
361 |
self.hidden_dim = config.hidden_size
|
|
|
362 |
self.intermediate_dim = config.intermediate_size
|
363 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
364 |
|
365 |
+
self.gate_proj = nn.Linear(
|
366 |
+
self.hidden_dim,
|
367 |
+
self.intermediate_dim,
|
368 |
+
bias=config.hidden_bias,
|
369 |
+
)
|
|
|
370 |
self.up_proj = nn.Linear(
|
371 |
self.hidden_dim,
|
372 |
self.intermediate_dim,
|
|
|
378 |
bias=config.hidden_bias,
|
379 |
)
|
380 |
|
381 |
+
def forward(
|
382 |
+
self,
|
383 |
+
hidden_states: torch.Tensor,
|
384 |
+
**kwargs,
|
385 |
+
) -> torch.Tensor:
|
386 |
+
hidden_states = self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
|
387 |
+
return hidden_states
|
388 |
+
|
389 |
+
|
390 |
+
class DogeCDMoE(DogeMLP):
|
391 |
+
"""Cross Domain Mixture of Experts from 'Wonderful Matrices' paper."""
|
392 |
+
|
393 |
+
def __init__(self, config: DogeConfig):
|
394 |
+
super().__init__(config)
|
395 |
+
self.hidden_dim = config.hidden_size
|
396 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
397 |
+
|
398 |
+
self.expert_retrieval_dim = config.expert_retrieval_size
|
399 |
+
self.num_cdmmoe_experts = config.num_cdmmoe_experts
|
400 |
+
self.num_cdmmoe_heads = config.num_cdmmoe_heads
|
401 |
+
self.num_cdmmoe_experts_per_head = config.num_cdmmoe_experts_per_head
|
402 |
+
self.num_keys = int(math.sqrt(self.num_cdmmoe_experts))
|
403 |
+
|
404 |
+
# queries and keys for retrieval experts
|
405 |
self.queries = nn.Linear(
|
406 |
self.hidden_dim,
|
407 |
+
self.num_cdmmoe_heads * self.expert_retrieval_dim,
|
408 |
bias=False,
|
409 |
)
|
|
|
410 |
self.keys = nn.Parameter(
|
411 |
torch.zeros(
|
412 |
self.num_cdmmoe_heads,
|
413 |
self.num_keys,
|
414 |
2,
|
415 |
+
self.expert_retrieval_dim // 2,
|
416 |
)
|
417 |
)
|
418 |
|
419 |
+
# experts
|
420 |
+
self.down_embed = nn.Embedding(
|
421 |
self.num_cdmmoe_experts,
|
422 |
self.hidden_dim,
|
423 |
)
|
|
|
431 |
self,
|
432 |
hidden_states: torch.Tensor,
|
433 |
**kwargs,
|
434 |
+
) -> torch.Tensor:
|
435 |
bsz, seq_len, _ = hidden_states.shape
|
436 |
|
437 |
# get similarity with queries and keys
|
|
|
439 |
queries = queries.view(bsz, seq_len, 2, self.num_cdmmoe_heads, -1).permute(2, 0, 1, 3, 4)
|
440 |
sim = torch.einsum("p b t h n, h k p n -> p b t h k", queries, self.keys)
|
441 |
|
442 |
+
# get experts with the highest similarity
|
443 |
(scores_x, scores_y), (indices_x, indices_y) = sim.topk(self.num_cdmmoe_experts_per_head, dim=-1)
|
444 |
if einx_add is not None:
|
445 |
all_scores = einx_add("... i, ... j -> ... (i j)", scores_x, scores_y)
|
|
|
451 |
all_indices = all_indices.view(*indices_x.shape[:-1], -1)
|
452 |
scores, pk_indices = all_scores.topk(self.num_cdmmoe_experts_per_head, dim=-1)
|
453 |
indices = all_indices.gather(-1, pk_indices)
|
|
|
|
|
454 |
down_embed = self.down_embed(indices)
|
455 |
up_embed = self.up_embed(indices)
|
456 |
|
457 |
+
# mix experts states with cross domain states
|
458 |
+
experts_weights = torch.einsum("b t d, b t h k d -> b t h k", hidden_states, down_embed)
|
459 |
+
experts_weights = self.act_fn(experts_weights) * scores.softmax(dim=-1)
|
460 |
experts_states = torch.einsum("b t h k, b t h k d -> b t d", experts_weights, up_embed)
|
461 |
+
hidden_states = self.down_proj(self.act_fn(self.gate_proj(hidden_states)) * self.up_proj(hidden_states))
|
|
|
|
|
462 |
hidden_states = hidden_states + experts_states
|
463 |
return hidden_states
|
464 |
|
|
|
468 |
super().__init__()
|
469 |
self.hidden_dropout = config.hidden_dropout
|
470 |
|
471 |
+
self.pre_sequence_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
472 |
+
self.attn = DOGE_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
|
473 |
+
self.post_sequence_residual = Residual(config.hidden_size)
|
474 |
+
|
475 |
+
self.pre_state_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
476 |
+
self.feed_forward = DogeMLP(config) if config.is_moe == False else DogeCDMoE(config)
|
477 |
+
self.post_state_residual = Residual(config.hidden_size)
|
478 |
|
479 |
def forward(
|
480 |
self,
|
|
|
513 |
|
514 |
# sequence transformation
|
515 |
residual = hidden_states
|
516 |
+
hidden_states = self.pre_sequence_layernorm(hidden_states)
|
517 |
hidden_states, present_key_value = self.attn(
|
518 |
hidden_states=hidden_states,
|
519 |
attention_mask=attention_mask,
|
|
|
525 |
)
|
526 |
self_attn_weights = None
|
527 |
hidden_states = F.dropout(hidden_states, p=self.hidden_dropout, training=self.training)
|
528 |
+
hidden_states = self.post_sequence_residual(residual, hidden_states)
|
529 |
|
530 |
# state transformation
|
531 |
residual = hidden_states
|
532 |
+
hidden_states = self.pre_state_layernorm(hidden_states)
|
533 |
hidden_states = self.feed_forward(hidden_states)
|
534 |
hidden_states = F.dropout(hidden_states, p=self.hidden_dropout, training=self.training)
|
535 |
+
hidden_states = self.post_state_residual(residual, hidden_states)
|
536 |
|
537 |
outputs = (hidden_states,)
|
538 |
|
|
|
552 |
supports_gradient_checkpointing = True
|
553 |
_no_split_modules = ["DogeDecoderLayer"]
|
554 |
_skip_keys_device_placement = ["past_key_values"]
|
555 |
+
_supports_sdpa = True
|
556 |
_supports_cache_class = True
|
557 |
_supports_quantized_cache = True
|
558 |
_supports_static_cache = True
|
|
|
726 |
if position_ids is None:
|
727 |
position_ids = cache_position.unsqueeze(0)
|
728 |
|
729 |
+
causal_mask = self._update_causal_mask(
|
730 |
+
attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
|
731 |
+
)
|
732 |
hidden_states = inputs_embeds
|
733 |
|
734 |
# create position embeddings to be shared across the decoder layers
|
|
|
737 |
# decoder layers
|
738 |
all_hidden_states = () if output_hidden_states else None
|
739 |
all_self_attns = () if output_attentions else None
|
740 |
+
next_decoder_cache = None
|
741 |
|
742 |
for decoder_layer in self.layers:
|
743 |
if output_hidden_states:
|
|
|
747 |
layer_outputs = self._gradient_checkpointing_func(
|
748 |
decoder_layer.__call__,
|
749 |
hidden_states,
|
750 |
+
causal_mask,
|
751 |
position_ids,
|
752 |
past_key_values,
|
753 |
output_attentions,
|
|
|
758 |
else:
|
759 |
layer_outputs = decoder_layer(
|
760 |
hidden_states,
|
761 |
+
attention_mask=causal_mask,
|
762 |
position_ids=position_ids,
|
763 |
past_key_value=past_key_values,
|
764 |
output_attentions=output_attentions,
|
|
|
795 |
attentions=all_self_attns,
|
796 |
)
|
797 |
|
798 |
+
def _update_causal_mask(
|
799 |
+
self,
|
800 |
+
attention_mask: torch.Tensor = None,
|
801 |
+
input_tensor: torch.Tensor = None,
|
802 |
+
cache_position: torch.Tensor = None,
|
803 |
+
past_key_values: Cache = None,
|
804 |
+
output_attentions: bool = False,
|
805 |
+
):
|
806 |
+
past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
|
807 |
+
using_static_cache = isinstance(past_key_values, StaticCache)
|
808 |
+
|
809 |
+
dtype, device = input_tensor.dtype, input_tensor.device
|
810 |
+
sequence_length = input_tensor.shape[1]
|
811 |
+
if using_static_cache:
|
812 |
+
target_length = past_key_values.get_max_cache_shape()
|
813 |
+
else:
|
814 |
+
target_length = (
|
815 |
+
attention_mask.shape[-1]
|
816 |
+
if isinstance(attention_mask, torch.Tensor)
|
817 |
+
else past_seen_tokens + sequence_length + 1
|
818 |
+
)
|
819 |
+
|
820 |
+
# in case the provided `attention` mask is 2D, we generate a causal mask here (4D).
|
821 |
+
causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
|
822 |
+
attention_mask=attention_mask,
|
823 |
+
sequence_length=sequence_length,
|
824 |
+
target_length=target_length,
|
825 |
+
dtype=dtype,
|
826 |
+
device=device,
|
827 |
+
cache_position=cache_position,
|
828 |
+
batch_size=input_tensor.shape[0],
|
829 |
+
)
|
830 |
+
|
831 |
+
return causal_mask
|
832 |
+
|
833 |
+
@staticmethod
|
834 |
+
def _prepare_4d_causal_attention_mask_with_cache_position(
|
835 |
+
attention_mask: torch.Tensor = None,
|
836 |
+
sequence_length: int = None,
|
837 |
+
target_length: int = None,
|
838 |
+
dtype: torch.dtype = None,
|
839 |
+
device: torch.device = None,
|
840 |
+
cache_position: torch.Tensor = None,
|
841 |
+
batch_size: int = None,
|
842 |
+
**kwargs,
|
843 |
+
):
|
844 |
+
"""
|
845 |
+
Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
|
846 |
+
`(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
|
847 |
+
|
848 |
+
Args:
|
849 |
+
attention_mask (`torch.Tensor`):
|
850 |
+
A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
|
851 |
+
`(batch_size, 1, query_length, key_value_length)`.
|
852 |
+
sequence_length (`int`):
|
853 |
+
The sequence length being processed.
|
854 |
+
target_length (`int`):
|
855 |
+
The target length: when generating with static cache, the mask should be as long as the static cache,
|
856 |
+
to account for the 0 padding, the part of the cache that is not filled yet.
|
857 |
+
dtype (`torch.dtype`):
|
858 |
+
The dtype to use for the 4D attention mask.
|
859 |
+
device (`torch.device`):
|
860 |
+
The device to plcae the 4D attention mask on.
|
861 |
+
cache_position (`torch.Tensor`):
|
862 |
+
Indices depicting the position of the input sequence tokens in the sequence.
|
863 |
+
batch_size (`torch.Tensor`):
|
864 |
+
Batch size.
|
865 |
+
"""
|
866 |
+
if attention_mask is not None and attention_mask.dim() == 4:
|
867 |
+
# In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
|
868 |
+
causal_mask = attention_mask
|
869 |
+
else:
|
870 |
+
min_dtype = torch.finfo(dtype).min
|
871 |
+
causal_mask = torch.full(
|
872 |
+
(sequence_length, target_length),
|
873 |
+
fill_value=min_dtype, dtype=dtype, device=device,
|
874 |
+
)
|
875 |
+
if sequence_length != 1:
|
876 |
+
causal_mask = torch.triu(causal_mask, diagonal=1)
|
877 |
+
causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
|
878 |
+
causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
|
879 |
+
if attention_mask is not None:
|
880 |
+
causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
|
881 |
+
mask_length = attention_mask.shape[-1]
|
882 |
+
padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
|
883 |
+
padding_mask = padding_mask == 0
|
884 |
+
causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
|
885 |
+
padding_mask, min_dtype
|
886 |
+
)
|
887 |
+
|
888 |
+
return causal_mask
|
|
|
|
|
|
|
889 |
|
890 |
|
891 |
class DogeForCausalLM(DogePreTrainedModel, GenerationMixin):
|