Update modeling_timer.py
Browse files- modeling_timer.py +572 -565
modeling_timer.py
CHANGED
@@ -1,565 +1,572 @@
|
|
1 |
-
from typing import Optional, Tuple, List, Union
|
2 |
-
import torch
|
3 |
-
from torch import nn
|
4 |
-
import torch.nn.functional as F
|
5 |
-
from transformers import PreTrainedModel, Cache, DynamicCache
|
6 |
-
from transformers.activations import ACT2FN
|
7 |
-
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
|
8 |
-
from transformers.modeling_outputs import MoeModelOutputWithPast, MoeCausalLMOutputWithPast
|
9 |
-
from .configuration_timer import TimerConfig
|
10 |
-
from .ts_generation_mixin import TSGenerationMixin
|
11 |
-
|
12 |
-
|
13 |
-
def rotate_half(x):
|
14 |
-
x1 = x[..., : x.shape[-1] // 2]
|
15 |
-
x2 = x[..., x.shape[-1] // 2:]
|
16 |
-
return torch.cat((-x2, x1), dim=-1)
|
17 |
-
|
18 |
-
|
19 |
-
def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
|
20 |
-
cos = cos[position_ids].unsqueeze(unsqueeze_dim)
|
21 |
-
sin = sin[position_ids].unsqueeze(unsqueeze_dim)
|
22 |
-
q_embed = (q * cos) + (rotate_half(q) * sin)
|
23 |
-
k_embed = (k * cos) + (rotate_half(k) * sin)
|
24 |
-
return q_embed, k_embed
|
25 |
-
|
26 |
-
|
27 |
-
class TimerPatchEmbedding(nn.Module):
|
28 |
-
def __init__(self, config: TimerConfig):
|
29 |
-
super().__init__()
|
30 |
-
self.input_token_len = config.input_token_len
|
31 |
-
self.emb = nn.Linear(config.input_token_len,
|
32 |
-
config.hidden_size, bias=False)
|
33 |
-
|
34 |
-
def forward(self, hidden_state: torch.Tensor):
|
35 |
-
hidden_state = hidden_state.unfold(
|
36 |
-
dimension=-1, size=self.input_token_len, step=self.input_token_len)
|
37 |
-
return self.emb(hidden_state)
|
38 |
-
|
39 |
-
|
40 |
-
class TimerPointEmbedding(nn.Module):
|
41 |
-
def __init__(self, config: TimerConfig):
|
42 |
-
super().__init__()
|
43 |
-
self.emb_layer = nn.Linear(
|
44 |
-
config.input_token_len, config.hidden_size, bias=False)
|
45 |
-
self.gate_layer = nn.Linear(
|
46 |
-
config.input_token_len, config.hidden_size, bias=False)
|
47 |
-
self.act_fn = ACT2FN[config.hidden_act]
|
48 |
-
|
49 |
-
def forward(self, x):
|
50 |
-
emb = self.act_fn(self.gate_layer(x)) * self.emb_layer(x)
|
51 |
-
return emb
|
52 |
-
|
53 |
-
|
54 |
-
class TimeMoeRotaryEmbedding(torch.nn.Module):
|
55 |
-
def __init__(self, dim, max_position_embeddings=10000, base=10000, device=None):
|
56 |
-
super().__init__()
|
57 |
-
self.dim = dim
|
58 |
-
self.max_position_embeddings = max_position_embeddings
|
59 |
-
self.base = base
|
60 |
-
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim,
|
61 |
-
2, dtype=torch.int64).float().to(device) / self.dim))
|
62 |
-
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
63 |
-
|
64 |
-
# Build here to make `torch.jit.trace` work.
|
65 |
-
self._set_cos_sin_cache(
|
66 |
-
seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
|
67 |
-
)
|
68 |
-
|
69 |
-
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
70 |
-
self.max_seq_len_cached = seq_len
|
71 |
-
t = torch.arange(self.max_seq_len_cached, device=device,
|
72 |
-
dtype=torch.int64).type_as(self.inv_freq)
|
73 |
-
|
74 |
-
freqs = torch.outer(t, self.inv_freq)
|
75 |
-
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
76 |
-
emb = torch.cat((freqs, freqs), dim=-1)
|
77 |
-
self.register_buffer(
|
78 |
-
"cos_cached", emb.cos().to(dtype), persistent=False)
|
79 |
-
self.register_buffer(
|
80 |
-
"sin_cached", emb.sin().to(dtype), persistent=False)
|
81 |
-
|
82 |
-
def forward(self, x, seq_len=None):
|
83 |
-
# x: [bs, num_attention_heads, seq_len, head_size]
|
84 |
-
if seq_len > self.max_seq_len_cached:
|
85 |
-
self._set_cos_sin_cache(
|
86 |
-
seq_len=seq_len, device=x.device, dtype=x.dtype)
|
87 |
-
|
88 |
-
return (
|
89 |
-
self.cos_cached[:seq_len].to(dtype=x.dtype),
|
90 |
-
self.sin_cached[:seq_len].to(dtype=x.dtype),
|
91 |
-
)
|
92 |
-
|
93 |
-
|
94 |
-
class TimerAttention(nn.Module):
|
95 |
-
def __init__(self, config: TimerConfig, layer_idx: Optional[int] = None):
|
96 |
-
super().__init__()
|
97 |
-
self.layer_idx = layer_idx
|
98 |
-
self.hidden_size = config.hidden_size
|
99 |
-
self.num_heads = config.num_attention_heads
|
100 |
-
self.head_dim = self.hidden_size // self.num_heads
|
101 |
-
self.attention_dropout = config.attention_dropout
|
102 |
-
self.q_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=True)
|
103 |
-
self.k_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=True)
|
104 |
-
self.v_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=True)
|
105 |
-
self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
|
106 |
-
self.rotary_emb = TimeMoeRotaryEmbedding(
|
107 |
-
self.head_dim, max_position_embeddings=config.max_position_embeddings)
|
108 |
-
|
109 |
-
def forward(
|
110 |
-
self,
|
111 |
-
hidden_states: torch.Tensor,
|
112 |
-
attention_mask: Optional[torch.Tensor] = None,
|
113 |
-
position_ids: Optional[torch.LongTensor] = None,
|
114 |
-
past_key_value: Optional[Cache] = None,
|
115 |
-
output_attentions: bool = False,
|
116 |
-
**kwargs,
|
117 |
-
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
118 |
-
bsz, q_len, _ = hidden_states.size()
|
119 |
-
|
120 |
-
query_states = self.q_proj(hidden_states)
|
121 |
-
key_states = self.k_proj(hidden_states)
|
122 |
-
value_states = self.v_proj(hidden_states)
|
123 |
-
|
124 |
-
query_states = query_states.view(
|
125 |
-
bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
126 |
-
key_states = key_states.view(
|
127 |
-
bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
128 |
-
value_states = value_states.view(
|
129 |
-
bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
130 |
-
|
131 |
-
kv_seq_len = key_states.shape[-2]
|
132 |
-
if past_key_value is not None:
|
133 |
-
kv_seq_len += past_key_value.get_usable_length(
|
134 |
-
kv_seq_len, self.layer_idx)
|
135 |
-
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
136 |
-
query_states, key_states = apply_rotary_pos_emb(
|
137 |
-
query_states, key_states, cos, sin, position_ids)
|
138 |
-
|
139 |
-
if past_key_value is not None:
|
140 |
-
key_states, value_states = past_key_value.update(
|
141 |
-
key_states, value_states, self.layer_idx)
|
142 |
-
|
143 |
-
attn_output = F.scaled_dot_product_attention(
|
144 |
-
query_states, key_states, value_states, attention_mask, dropout_p=self.attention_dropout)
|
145 |
-
|
146 |
-
attn_output = attn_output.transpose(1, 2).contiguous()
|
147 |
-
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
148 |
-
attn_output = self.o_proj(attn_output)
|
149 |
-
|
150 |
-
if not output_attentions:
|
151 |
-
attn_weights = None
|
152 |
-
|
153 |
-
return attn_output, attn_weights, past_key_value
|
154 |
-
|
155 |
-
|
156 |
-
class TimerMLP(nn.Module):
|
157 |
-
def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
|
158 |
-
super().__init__()
|
159 |
-
self.hidden_size = hidden_size
|
160 |
-
self.intermediate_size = intermediate_size
|
161 |
-
self.gate_proj = nn.Linear(
|
162 |
-
self.hidden_size, self.intermediate_size, bias=False)
|
163 |
-
self.up_proj = nn.Linear(
|
164 |
-
self.hidden_size, self.intermediate_size, bias=False)
|
165 |
-
self.down_proj = nn.Linear(
|
166 |
-
self.intermediate_size, self.hidden_size, bias=False)
|
167 |
-
self.act_fn = ACT2FN[hidden_act]
|
168 |
-
|
169 |
-
def forward(self, hidden_state):
|
170 |
-
return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
|
171 |
-
|
172 |
-
|
173 |
-
class TimerDecoderLayer(nn.Module):
|
174 |
-
def __init__(self, config: TimerConfig, layer_idx: int):
|
175 |
-
super().__init__()
|
176 |
-
self.self_attn = TimerAttention(config, layer_idx)
|
177 |
-
|
178 |
-
self.ffn_layer = TimerMLP(
|
179 |
-
hidden_size=config.hidden_size,
|
180 |
-
intermediate_size=config.intermediate_size,
|
181 |
-
hidden_act=config.hidden_act,
|
182 |
-
)
|
183 |
-
self.norm1 = torch.nn.LayerNorm(config.hidden_size)
|
184 |
-
self.norm2 = torch.nn.LayerNorm(config.hidden_size)
|
185 |
-
|
186 |
-
def forward(
|
187 |
-
self,
|
188 |
-
hidden_states: torch.Tensor,
|
189 |
-
attention_mask: Optional[torch.Tensor] = None,
|
190 |
-
position_ids: Optional[torch.LongTensor] = None,
|
191 |
-
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
192 |
-
output_attentions: Optional[bool] = False,
|
193 |
-
use_cache: Optional[bool] = False,
|
194 |
-
**kwargs,
|
195 |
-
) -> Tuple[torch.FloatTensor, torch.FloatTensor, Optional[torch.FloatTensor], Optional[torch.FloatTensor]]:
|
196 |
-
residual = hidden_states
|
197 |
-
|
198 |
-
# Self Attention
|
199 |
-
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
200 |
-
hidden_states=hidden_states,
|
201 |
-
attention_mask=attention_mask,
|
202 |
-
position_ids=position_ids,
|
203 |
-
past_key_value=past_key_value,
|
204 |
-
output_attentions=output_attentions,
|
205 |
-
use_cache=use_cache,
|
206 |
-
)
|
207 |
-
hidden_states = residual + hidden_states
|
208 |
-
hidden_states = self.norm1(hidden_states)
|
209 |
-
|
210 |
-
# Fully Connected
|
211 |
-
residual = hidden_states
|
212 |
-
hidden_states = self.ffn_layer(hidden_states)
|
213 |
-
hidden_states = residual + hidden_states
|
214 |
-
hidden_states = self.norm2(hidden_states)
|
215 |
-
|
216 |
-
if not output_attentions:
|
217 |
-
self_attn_weights = None
|
218 |
-
|
219 |
-
if not use_cache:
|
220 |
-
present_key_value = None
|
221 |
-
return hidden_states, self_attn_weights, present_key_value
|
222 |
-
|
223 |
-
|
224 |
-
class TimerPreTrainedModel(PreTrainedModel):
|
225 |
-
config_class = TimerConfig
|
226 |
-
base_model_prefix = "model"
|
227 |
-
supports_gradient_checkpointing = True
|
228 |
-
_no_split_modules = ["TimeMoeDecoderLayer"]
|
229 |
-
_skip_keys_device_placement = "past_key_values"
|
230 |
-
_supports_flash_attn_2 = True
|
231 |
-
_supports_sdpa = False
|
232 |
-
_supports_cache_class = True
|
233 |
-
|
234 |
-
def _init_weights(self, module):
|
235 |
-
std = self.config.initializer_range
|
236 |
-
if isinstance(module, torch.nn.Linear):
|
237 |
-
module.weight.data.normal_(mean=0.0, std=std)
|
238 |
-
if module.bias is not None:
|
239 |
-
module.bias.data.zero_()
|
240 |
-
elif isinstance(module, torch.nn.Embedding):
|
241 |
-
module.weight.data.normal_(mean=0.0, std=std)
|
242 |
-
if module.padding_idx is not None:
|
243 |
-
module.weight.data[module.padding_idx].zero_()
|
244 |
-
|
245 |
-
|
246 |
-
class TimerModel(TimerPreTrainedModel):
|
247 |
-
def __init__(self, config: TimerConfig):
|
248 |
-
super().__init__(config)
|
249 |
-
self.embed_layer = TimerPatchEmbedding(config)
|
250 |
-
self.layers = nn.ModuleList(
|
251 |
-
[TimerDecoderLayer(config, layer_idx)
|
252 |
-
for layer_idx in range(config.num_hidden_layers)]
|
253 |
-
)
|
254 |
-
self.norm = torch.nn.LayerNorm(config.hidden_size)
|
255 |
-
self.gradient_checkpointing = False
|
256 |
-
|
257 |
-
def forward(
|
258 |
-
self,
|
259 |
-
input_ids: torch.FloatTensor = None,
|
260 |
-
attention_mask: Optional[torch.Tensor] = None,
|
261 |
-
position_ids: Optional[torch.LongTensor] = None,
|
262 |
-
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
263 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
264 |
-
use_cache: Optional[bool] = None,
|
265 |
-
output_attentions: Optional[bool] = None,
|
266 |
-
output_hidden_states: Optional[bool] = None,
|
267 |
-
return_dict: Optional[bool] = None,
|
268 |
-
) -> Union[Tuple, MoeModelOutputWithPast]:
|
269 |
-
# input_ids is the input of time series, its shape is [batch_size, seq_len]
|
270 |
-
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
271 |
-
output_hidden_states = (
|
272 |
-
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
273 |
-
)
|
274 |
-
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
275 |
-
|
276 |
-
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
277 |
-
|
278 |
-
# retrieve input_ids and inputs_embeds
|
279 |
-
if input_ids is not None and inputs_embeds is not None:
|
280 |
-
raise ValueError(
|
281 |
-
"You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
|
282 |
-
elif input_ids is not None:
|
283 |
-
batch_size, seq_length = input_ids.shape
|
284 |
-
elif inputs_embeds is not None:
|
285 |
-
batch_size, seq_length, _ = inputs_embeds.shape
|
286 |
-
else:
|
287 |
-
raise ValueError(
|
288 |
-
"You have to specify either decoder_input_ids or decoder_inputs_embeds")
|
289 |
-
|
290 |
-
if inputs_embeds is None:
|
291 |
-
inputs_embeds = self.embed_layer(input_ids)
|
292 |
-
seq_length = inputs_embeds.shape[1]
|
293 |
-
|
294 |
-
if self.gradient_checkpointing and self.training:
|
295 |
-
if use_cache:
|
296 |
-
use_cache = False
|
297 |
-
|
298 |
-
past_key_values_length = 0
|
299 |
-
|
300 |
-
if use_cache:
|
301 |
-
use_legacy_cache = not isinstance(past_key_values, Cache)
|
302 |
-
if use_legacy_cache:
|
303 |
-
past_key_values = DynamicCache.from_legacy_cache(
|
304 |
-
past_key_values)
|
305 |
-
past_key_values_length = past_key_values.get_usable_length(
|
306 |
-
seq_length)
|
307 |
-
|
308 |
-
if position_ids is None:
|
309 |
-
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
310 |
-
position_ids = torch.arange(
|
311 |
-
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
|
312 |
-
)
|
313 |
-
# position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
|
314 |
-
position_ids = position_ids.view(-1, seq_length)
|
315 |
-
else:
|
316 |
-
position_ids = position_ids.view(-1, seq_length).long()
|
317 |
-
|
318 |
-
# 4d mask is passed through the layers
|
319 |
-
attention_mask = _prepare_4d_causal_attention_mask(
|
320 |
-
attention_mask,
|
321 |
-
(batch_size, seq_length),
|
322 |
-
inputs_embeds,
|
323 |
-
past_key_values_length,
|
324 |
-
sliding_window=None,
|
325 |
-
)
|
326 |
-
|
327 |
-
hidden_states = inputs_embeds
|
328 |
-
|
329 |
-
# decoder layers
|
330 |
-
all_hidden_states = () if output_hidden_states else None
|
331 |
-
all_self_attns = () if output_attentions else None
|
332 |
-
next_decoder_cache = None
|
333 |
-
|
334 |
-
for decoder_layer in self.layers:
|
335 |
-
if output_hidden_states:
|
336 |
-
all_hidden_states += (hidden_states,)
|
337 |
-
|
338 |
-
if self.gradient_checkpointing and self.training:
|
339 |
-
layer_outputs = self._gradient_checkpointing_func(
|
340 |
-
decoder_layer.__call__,
|
341 |
-
hidden_states,
|
342 |
-
attention_mask,
|
343 |
-
position_ids,
|
344 |
-
past_key_values,
|
345 |
-
output_attentions,
|
346 |
-
use_cache,
|
347 |
-
)
|
348 |
-
else:
|
349 |
-
layer_outputs = decoder_layer(
|
350 |
-
hidden_states,
|
351 |
-
attention_mask=attention_mask,
|
352 |
-
position_ids=position_ids,
|
353 |
-
past_key_value=past_key_values,
|
354 |
-
output_attentions=output_attentions,
|
355 |
-
use_cache=use_cache,
|
356 |
-
)
|
357 |
-
|
358 |
-
hidden_states = layer_outputs[0]
|
359 |
-
|
360 |
-
if output_attentions:
|
361 |
-
all_self_attns += (layer_outputs[1],)
|
362 |
-
|
363 |
-
if use_cache:
|
364 |
-
next_decoder_cache = layer_outputs[2]
|
365 |
-
|
366 |
-
hidden_states = self.norm(hidden_states)
|
367 |
-
# add hidden states from the last decoder layer
|
368 |
-
if output_hidden_states:
|
369 |
-
all_hidden_states += (hidden_states,)
|
370 |
-
|
371 |
-
next_cache = None
|
372 |
-
if use_cache:
|
373 |
-
next_cache = next_decoder_cache.to_legacy_cache(
|
374 |
-
) if use_legacy_cache else next_decoder_cache
|
375 |
-
|
376 |
-
if not return_dict:
|
377 |
-
return tuple(
|
378 |
-
v
|
379 |
-
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
|
380 |
-
if v is not None
|
381 |
-
)
|
382 |
-
return MoeModelOutputWithPast(
|
383 |
-
last_hidden_state=hidden_states,
|
384 |
-
past_key_values=next_cache,
|
385 |
-
hidden_states=all_hidden_states,
|
386 |
-
attentions=all_self_attns,
|
387 |
-
)
|
388 |
-
|
389 |
-
|
390 |
-
class TimerForPrediction(TimerPreTrainedModel, TSGenerationMixin):
|
391 |
-
def __init__(self, config: TimerConfig):
|
392 |
-
super().__init__(config)
|
393 |
-
self.config = config
|
394 |
-
self.model = TimerModel(self.config)
|
395 |
-
lm_head_list = []
|
396 |
-
self.output_token_len_map = {}
|
397 |
-
for i, output_token_len in enumerate(self.config.output_token_lens):
|
398 |
-
lm_head_list.append(
|
399 |
-
nn.Linear(self.config.hidden_size, output_token_len, bias=False))
|
400 |
-
self.output_token_len_map[output_token_len] = i
|
401 |
-
self.lm_heads = nn.ModuleList(lm_head_list)
|
402 |
-
self.loss_function = torch.nn.MSELoss(reduction='none')
|
403 |
-
self.post_init()
|
404 |
-
|
405 |
-
def set_decoder(self, decoder):
|
406 |
-
self.model = decoder
|
407 |
-
|
408 |
-
def get_decoder(self):
|
409 |
-
return self.model
|
410 |
-
|
411 |
-
def forward(
|
412 |
-
self,
|
413 |
-
input_ids: torch.FloatTensor = None,
|
414 |
-
attention_mask: Optional[torch.Tensor] = None,
|
415 |
-
position_ids: Optional[torch.LongTensor] = None,
|
416 |
-
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
417 |
-
inputs_embeds: Optional[torch.FloatTensor] = None,
|
418 |
-
labels: Optional[torch.FloatTensor] = None,
|
419 |
-
loss_masks: Optional[torch.FloatTensor] = None,
|
420 |
-
use_cache: Optional[bool] = None,
|
421 |
-
output_attentions: Optional[bool] = None,
|
422 |
-
output_hidden_states: Optional[bool] = None,
|
423 |
-
return_dict: Optional[bool] = None,
|
424 |
-
max_output_length: Optional[int] = None,
|
425 |
-
|
426 |
-
|
427 |
-
|
428 |
-
|
429 |
-
|
430 |
-
|
431 |
-
|
432 |
-
|
433 |
-
|
434 |
-
|
435 |
-
|
436 |
-
|
437 |
-
|
438 |
-
|
439 |
-
|
440 |
-
|
441 |
-
|
442 |
-
|
443 |
-
|
444 |
-
|
445 |
-
|
446 |
-
|
447 |
-
|
448 |
-
|
449 |
-
|
450 |
-
|
451 |
-
|
452 |
-
|
453 |
-
|
454 |
-
|
455 |
-
|
456 |
-
|
457 |
-
|
458 |
-
|
459 |
-
|
460 |
-
|
461 |
-
|
462 |
-
|
463 |
-
|
464 |
-
|
465 |
-
|
466 |
-
|
467 |
-
|
468 |
-
|
469 |
-
|
470 |
-
|
471 |
-
|
472 |
-
|
473 |
-
|
474 |
-
|
475 |
-
|
476 |
-
|
477 |
-
|
478 |
-
|
479 |
-
|
480 |
-
|
481 |
-
|
482 |
-
|
483 |
-
|
484 |
-
|
485 |
-
|
486 |
-
|
487 |
-
|
488 |
-
|
489 |
-
|
490 |
-
|
491 |
-
|
492 |
-
|
493 |
-
|
494 |
-
|
495 |
-
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
|
500 |
-
|
501 |
-
|
502 |
-
|
503 |
-
|
504 |
-
|
505 |
-
|
506 |
-
|
507 |
-
|
508 |
-
|
509 |
-
|
510 |
-
|
511 |
-
|
512 |
-
|
513 |
-
|
514 |
-
|
515 |
-
|
516 |
-
|
517 |
-
|
518 |
-
|
519 |
-
|
520 |
-
|
521 |
-
|
522 |
-
|
523 |
-
|
524 |
-
|
525 |
-
|
526 |
-
|
527 |
-
#
|
528 |
-
#
|
529 |
-
|
530 |
-
|
531 |
-
|
532 |
-
|
533 |
-
|
534 |
-
#
|
535 |
-
|
536 |
-
|
537 |
-
|
538 |
-
|
539 |
-
|
540 |
-
|
541 |
-
|
542 |
-
|
543 |
-
|
544 |
-
|
545 |
-
|
546 |
-
|
547 |
-
|
548 |
-
|
549 |
-
|
550 |
-
|
551 |
-
|
552 |
-
|
553 |
-
|
554 |
-
|
555 |
-
|
556 |
-
|
557 |
-
|
558 |
-
|
559 |
-
|
560 |
-
|
561 |
-
|
562 |
-
|
563 |
-
|
564 |
-
|
565 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Optional, Tuple, List, Union
|
2 |
+
import torch
|
3 |
+
from torch import nn
|
4 |
+
import torch.nn.functional as F
|
5 |
+
from transformers import PreTrainedModel, Cache, DynamicCache
|
6 |
+
from transformers.activations import ACT2FN
|
7 |
+
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
|
8 |
+
from transformers.modeling_outputs import MoeModelOutputWithPast, MoeCausalLMOutputWithPast
|
9 |
+
from .configuration_timer import TimerConfig
|
10 |
+
from .ts_generation_mixin import TSGenerationMixin
|
11 |
+
|
12 |
+
|
13 |
+
def rotate_half(x):
|
14 |
+
x1 = x[..., : x.shape[-1] // 2]
|
15 |
+
x2 = x[..., x.shape[-1] // 2:]
|
16 |
+
return torch.cat((-x2, x1), dim=-1)
|
17 |
+
|
18 |
+
|
19 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
|
20 |
+
cos = cos[position_ids].unsqueeze(unsqueeze_dim)
|
21 |
+
sin = sin[position_ids].unsqueeze(unsqueeze_dim)
|
22 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
23 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
24 |
+
return q_embed, k_embed
|
25 |
+
|
26 |
+
|
27 |
+
class TimerPatchEmbedding(nn.Module):
|
28 |
+
def __init__(self, config: TimerConfig):
|
29 |
+
super().__init__()
|
30 |
+
self.input_token_len = config.input_token_len
|
31 |
+
self.emb = nn.Linear(config.input_token_len,
|
32 |
+
config.hidden_size, bias=False)
|
33 |
+
|
34 |
+
def forward(self, hidden_state: torch.Tensor):
|
35 |
+
hidden_state = hidden_state.unfold(
|
36 |
+
dimension=-1, size=self.input_token_len, step=self.input_token_len)
|
37 |
+
return self.emb(hidden_state)
|
38 |
+
|
39 |
+
|
40 |
+
class TimerPointEmbedding(nn.Module):
|
41 |
+
def __init__(self, config: TimerConfig):
|
42 |
+
super().__init__()
|
43 |
+
self.emb_layer = nn.Linear(
|
44 |
+
config.input_token_len, config.hidden_size, bias=False)
|
45 |
+
self.gate_layer = nn.Linear(
|
46 |
+
config.input_token_len, config.hidden_size, bias=False)
|
47 |
+
self.act_fn = ACT2FN[config.hidden_act]
|
48 |
+
|
49 |
+
def forward(self, x):
|
50 |
+
emb = self.act_fn(self.gate_layer(x)) * self.emb_layer(x)
|
51 |
+
return emb
|
52 |
+
|
53 |
+
|
54 |
+
class TimeMoeRotaryEmbedding(torch.nn.Module):
|
55 |
+
def __init__(self, dim, max_position_embeddings=10000, base=10000, device=None):
|
56 |
+
super().__init__()
|
57 |
+
self.dim = dim
|
58 |
+
self.max_position_embeddings = max_position_embeddings
|
59 |
+
self.base = base
|
60 |
+
inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim,
|
61 |
+
2, dtype=torch.int64).float().to(device) / self.dim))
|
62 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
63 |
+
|
64 |
+
# Build here to make `torch.jit.trace` work.
|
65 |
+
self._set_cos_sin_cache(
|
66 |
+
seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype()
|
67 |
+
)
|
68 |
+
|
69 |
+
def _set_cos_sin_cache(self, seq_len, device, dtype):
|
70 |
+
self.max_seq_len_cached = seq_len
|
71 |
+
t = torch.arange(self.max_seq_len_cached, device=device,
|
72 |
+
dtype=torch.int64).type_as(self.inv_freq)
|
73 |
+
|
74 |
+
freqs = torch.outer(t, self.inv_freq)
|
75 |
+
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
76 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
77 |
+
self.register_buffer(
|
78 |
+
"cos_cached", emb.cos().to(dtype), persistent=False)
|
79 |
+
self.register_buffer(
|
80 |
+
"sin_cached", emb.sin().to(dtype), persistent=False)
|
81 |
+
|
82 |
+
def forward(self, x, seq_len=None):
|
83 |
+
# x: [bs, num_attention_heads, seq_len, head_size]
|
84 |
+
if seq_len > self.max_seq_len_cached:
|
85 |
+
self._set_cos_sin_cache(
|
86 |
+
seq_len=seq_len, device=x.device, dtype=x.dtype)
|
87 |
+
|
88 |
+
return (
|
89 |
+
self.cos_cached[:seq_len].to(dtype=x.dtype),
|
90 |
+
self.sin_cached[:seq_len].to(dtype=x.dtype),
|
91 |
+
)
|
92 |
+
|
93 |
+
|
94 |
+
class TimerAttention(nn.Module):
|
95 |
+
def __init__(self, config: TimerConfig, layer_idx: Optional[int] = None):
|
96 |
+
super().__init__()
|
97 |
+
self.layer_idx = layer_idx
|
98 |
+
self.hidden_size = config.hidden_size
|
99 |
+
self.num_heads = config.num_attention_heads
|
100 |
+
self.head_dim = self.hidden_size // self.num_heads
|
101 |
+
self.attention_dropout = config.attention_dropout
|
102 |
+
self.q_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=True)
|
103 |
+
self.k_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=True)
|
104 |
+
self.v_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=True)
|
105 |
+
self.o_proj = nn.Linear(self.hidden_size, self.hidden_size, bias=False)
|
106 |
+
self.rotary_emb = TimeMoeRotaryEmbedding(
|
107 |
+
self.head_dim, max_position_embeddings=config.max_position_embeddings)
|
108 |
+
|
109 |
+
def forward(
|
110 |
+
self,
|
111 |
+
hidden_states: torch.Tensor,
|
112 |
+
attention_mask: Optional[torch.Tensor] = None,
|
113 |
+
position_ids: Optional[torch.LongTensor] = None,
|
114 |
+
past_key_value: Optional[Cache] = None,
|
115 |
+
output_attentions: bool = False,
|
116 |
+
**kwargs,
|
117 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
118 |
+
bsz, q_len, _ = hidden_states.size()
|
119 |
+
|
120 |
+
query_states = self.q_proj(hidden_states)
|
121 |
+
key_states = self.k_proj(hidden_states)
|
122 |
+
value_states = self.v_proj(hidden_states)
|
123 |
+
|
124 |
+
query_states = query_states.view(
|
125 |
+
bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
126 |
+
key_states = key_states.view(
|
127 |
+
bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
128 |
+
value_states = value_states.view(
|
129 |
+
bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
130 |
+
|
131 |
+
kv_seq_len = key_states.shape[-2]
|
132 |
+
if past_key_value is not None:
|
133 |
+
kv_seq_len += past_key_value.get_usable_length(
|
134 |
+
kv_seq_len, self.layer_idx)
|
135 |
+
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
136 |
+
query_states, key_states = apply_rotary_pos_emb(
|
137 |
+
query_states, key_states, cos, sin, position_ids)
|
138 |
+
|
139 |
+
if past_key_value is not None:
|
140 |
+
key_states, value_states = past_key_value.update(
|
141 |
+
key_states, value_states, self.layer_idx)
|
142 |
+
|
143 |
+
attn_output = F.scaled_dot_product_attention(
|
144 |
+
query_states, key_states, value_states, attention_mask, dropout_p=self.attention_dropout)
|
145 |
+
|
146 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
147 |
+
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
148 |
+
attn_output = self.o_proj(attn_output)
|
149 |
+
|
150 |
+
if not output_attentions:
|
151 |
+
attn_weights = None
|
152 |
+
|
153 |
+
return attn_output, attn_weights, past_key_value
|
154 |
+
|
155 |
+
|
156 |
+
class TimerMLP(nn.Module):
|
157 |
+
def __init__(self, hidden_size: int, intermediate_size: int, hidden_act: str):
|
158 |
+
super().__init__()
|
159 |
+
self.hidden_size = hidden_size
|
160 |
+
self.intermediate_size = intermediate_size
|
161 |
+
self.gate_proj = nn.Linear(
|
162 |
+
self.hidden_size, self.intermediate_size, bias=False)
|
163 |
+
self.up_proj = nn.Linear(
|
164 |
+
self.hidden_size, self.intermediate_size, bias=False)
|
165 |
+
self.down_proj = nn.Linear(
|
166 |
+
self.intermediate_size, self.hidden_size, bias=False)
|
167 |
+
self.act_fn = ACT2FN[hidden_act]
|
168 |
+
|
169 |
+
def forward(self, hidden_state):
|
170 |
+
return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
|
171 |
+
|
172 |
+
|
173 |
+
class TimerDecoderLayer(nn.Module):
|
174 |
+
def __init__(self, config: TimerConfig, layer_idx: int):
|
175 |
+
super().__init__()
|
176 |
+
self.self_attn = TimerAttention(config, layer_idx)
|
177 |
+
|
178 |
+
self.ffn_layer = TimerMLP(
|
179 |
+
hidden_size=config.hidden_size,
|
180 |
+
intermediate_size=config.intermediate_size,
|
181 |
+
hidden_act=config.hidden_act,
|
182 |
+
)
|
183 |
+
self.norm1 = torch.nn.LayerNorm(config.hidden_size)
|
184 |
+
self.norm2 = torch.nn.LayerNorm(config.hidden_size)
|
185 |
+
|
186 |
+
def forward(
|
187 |
+
self,
|
188 |
+
hidden_states: torch.Tensor,
|
189 |
+
attention_mask: Optional[torch.Tensor] = None,
|
190 |
+
position_ids: Optional[torch.LongTensor] = None,
|
191 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
192 |
+
output_attentions: Optional[bool] = False,
|
193 |
+
use_cache: Optional[bool] = False,
|
194 |
+
**kwargs,
|
195 |
+
) -> Tuple[torch.FloatTensor, torch.FloatTensor, Optional[torch.FloatTensor], Optional[torch.FloatTensor]]:
|
196 |
+
residual = hidden_states
|
197 |
+
|
198 |
+
# Self Attention
|
199 |
+
hidden_states, self_attn_weights, present_key_value = self.self_attn(
|
200 |
+
hidden_states=hidden_states,
|
201 |
+
attention_mask=attention_mask,
|
202 |
+
position_ids=position_ids,
|
203 |
+
past_key_value=past_key_value,
|
204 |
+
output_attentions=output_attentions,
|
205 |
+
use_cache=use_cache,
|
206 |
+
)
|
207 |
+
hidden_states = residual + hidden_states
|
208 |
+
hidden_states = self.norm1(hidden_states)
|
209 |
+
|
210 |
+
# Fully Connected
|
211 |
+
residual = hidden_states
|
212 |
+
hidden_states = self.ffn_layer(hidden_states)
|
213 |
+
hidden_states = residual + hidden_states
|
214 |
+
hidden_states = self.norm2(hidden_states)
|
215 |
+
|
216 |
+
if not output_attentions:
|
217 |
+
self_attn_weights = None
|
218 |
+
|
219 |
+
if not use_cache:
|
220 |
+
present_key_value = None
|
221 |
+
return hidden_states, self_attn_weights, present_key_value
|
222 |
+
|
223 |
+
|
224 |
+
class TimerPreTrainedModel(PreTrainedModel):
|
225 |
+
config_class = TimerConfig
|
226 |
+
base_model_prefix = "model"
|
227 |
+
supports_gradient_checkpointing = True
|
228 |
+
_no_split_modules = ["TimeMoeDecoderLayer"]
|
229 |
+
_skip_keys_device_placement = "past_key_values"
|
230 |
+
_supports_flash_attn_2 = True
|
231 |
+
_supports_sdpa = False
|
232 |
+
_supports_cache_class = True
|
233 |
+
|
234 |
+
def _init_weights(self, module):
|
235 |
+
std = self.config.initializer_range
|
236 |
+
if isinstance(module, torch.nn.Linear):
|
237 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
238 |
+
if module.bias is not None:
|
239 |
+
module.bias.data.zero_()
|
240 |
+
elif isinstance(module, torch.nn.Embedding):
|
241 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
242 |
+
if module.padding_idx is not None:
|
243 |
+
module.weight.data[module.padding_idx].zero_()
|
244 |
+
|
245 |
+
|
246 |
+
class TimerModel(TimerPreTrainedModel):
|
247 |
+
def __init__(self, config: TimerConfig):
|
248 |
+
super().__init__(config)
|
249 |
+
self.embed_layer = TimerPatchEmbedding(config)
|
250 |
+
self.layers = nn.ModuleList(
|
251 |
+
[TimerDecoderLayer(config, layer_idx)
|
252 |
+
for layer_idx in range(config.num_hidden_layers)]
|
253 |
+
)
|
254 |
+
self.norm = torch.nn.LayerNorm(config.hidden_size)
|
255 |
+
self.gradient_checkpointing = False
|
256 |
+
|
257 |
+
def forward(
|
258 |
+
self,
|
259 |
+
input_ids: torch.FloatTensor = None,
|
260 |
+
attention_mask: Optional[torch.Tensor] = None,
|
261 |
+
position_ids: Optional[torch.LongTensor] = None,
|
262 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
263 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
264 |
+
use_cache: Optional[bool] = None,
|
265 |
+
output_attentions: Optional[bool] = None,
|
266 |
+
output_hidden_states: Optional[bool] = None,
|
267 |
+
return_dict: Optional[bool] = None,
|
268 |
+
) -> Union[Tuple, MoeModelOutputWithPast]:
|
269 |
+
# input_ids is the input of time series, its shape is [batch_size, seq_len]
|
270 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
271 |
+
output_hidden_states = (
|
272 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
273 |
+
)
|
274 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
275 |
+
|
276 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
277 |
+
|
278 |
+
# retrieve input_ids and inputs_embeds
|
279 |
+
if input_ids is not None and inputs_embeds is not None:
|
280 |
+
raise ValueError(
|
281 |
+
"You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
|
282 |
+
elif input_ids is not None:
|
283 |
+
batch_size, seq_length = input_ids.shape
|
284 |
+
elif inputs_embeds is not None:
|
285 |
+
batch_size, seq_length, _ = inputs_embeds.shape
|
286 |
+
else:
|
287 |
+
raise ValueError(
|
288 |
+
"You have to specify either decoder_input_ids or decoder_inputs_embeds")
|
289 |
+
|
290 |
+
if inputs_embeds is None:
|
291 |
+
inputs_embeds = self.embed_layer(input_ids)
|
292 |
+
seq_length = inputs_embeds.shape[1]
|
293 |
+
|
294 |
+
if self.gradient_checkpointing and self.training:
|
295 |
+
if use_cache:
|
296 |
+
use_cache = False
|
297 |
+
|
298 |
+
past_key_values_length = 0
|
299 |
+
|
300 |
+
if use_cache:
|
301 |
+
use_legacy_cache = not isinstance(past_key_values, Cache)
|
302 |
+
if use_legacy_cache:
|
303 |
+
past_key_values = DynamicCache.from_legacy_cache(
|
304 |
+
past_key_values)
|
305 |
+
past_key_values_length = past_key_values.get_usable_length(
|
306 |
+
seq_length)
|
307 |
+
|
308 |
+
if position_ids is None:
|
309 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
310 |
+
position_ids = torch.arange(
|
311 |
+
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
|
312 |
+
)
|
313 |
+
# position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
|
314 |
+
position_ids = position_ids.view(-1, seq_length)
|
315 |
+
else:
|
316 |
+
position_ids = position_ids.view(-1, seq_length).long()
|
317 |
+
|
318 |
+
# 4d mask is passed through the layers
|
319 |
+
attention_mask = _prepare_4d_causal_attention_mask(
|
320 |
+
attention_mask,
|
321 |
+
(batch_size, seq_length),
|
322 |
+
inputs_embeds,
|
323 |
+
past_key_values_length,
|
324 |
+
sliding_window=None,
|
325 |
+
)
|
326 |
+
|
327 |
+
hidden_states = inputs_embeds
|
328 |
+
|
329 |
+
# decoder layers
|
330 |
+
all_hidden_states = () if output_hidden_states else None
|
331 |
+
all_self_attns = () if output_attentions else None
|
332 |
+
next_decoder_cache = None
|
333 |
+
|
334 |
+
for decoder_layer in self.layers:
|
335 |
+
if output_hidden_states:
|
336 |
+
all_hidden_states += (hidden_states,)
|
337 |
+
|
338 |
+
if self.gradient_checkpointing and self.training:
|
339 |
+
layer_outputs = self._gradient_checkpointing_func(
|
340 |
+
decoder_layer.__call__,
|
341 |
+
hidden_states,
|
342 |
+
attention_mask,
|
343 |
+
position_ids,
|
344 |
+
past_key_values,
|
345 |
+
output_attentions,
|
346 |
+
use_cache,
|
347 |
+
)
|
348 |
+
else:
|
349 |
+
layer_outputs = decoder_layer(
|
350 |
+
hidden_states,
|
351 |
+
attention_mask=attention_mask,
|
352 |
+
position_ids=position_ids,
|
353 |
+
past_key_value=past_key_values,
|
354 |
+
output_attentions=output_attentions,
|
355 |
+
use_cache=use_cache,
|
356 |
+
)
|
357 |
+
|
358 |
+
hidden_states = layer_outputs[0]
|
359 |
+
|
360 |
+
if output_attentions:
|
361 |
+
all_self_attns += (layer_outputs[1],)
|
362 |
+
|
363 |
+
if use_cache:
|
364 |
+
next_decoder_cache = layer_outputs[2]
|
365 |
+
|
366 |
+
hidden_states = self.norm(hidden_states)
|
367 |
+
# add hidden states from the last decoder layer
|
368 |
+
if output_hidden_states:
|
369 |
+
all_hidden_states += (hidden_states,)
|
370 |
+
|
371 |
+
next_cache = None
|
372 |
+
if use_cache:
|
373 |
+
next_cache = next_decoder_cache.to_legacy_cache(
|
374 |
+
) if use_legacy_cache else next_decoder_cache
|
375 |
+
|
376 |
+
if not return_dict:
|
377 |
+
return tuple(
|
378 |
+
v
|
379 |
+
for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
|
380 |
+
if v is not None
|
381 |
+
)
|
382 |
+
return MoeModelOutputWithPast(
|
383 |
+
last_hidden_state=hidden_states,
|
384 |
+
past_key_values=next_cache,
|
385 |
+
hidden_states=all_hidden_states,
|
386 |
+
attentions=all_self_attns,
|
387 |
+
)
|
388 |
+
|
389 |
+
|
390 |
+
class TimerForPrediction(TimerPreTrainedModel, TSGenerationMixin):
|
391 |
+
def __init__(self, config: TimerConfig):
|
392 |
+
super().__init__(config)
|
393 |
+
self.config = config
|
394 |
+
self.model = TimerModel(self.config)
|
395 |
+
lm_head_list = []
|
396 |
+
self.output_token_len_map = {}
|
397 |
+
for i, output_token_len in enumerate(self.config.output_token_lens):
|
398 |
+
lm_head_list.append(
|
399 |
+
nn.Linear(self.config.hidden_size, output_token_len, bias=False))
|
400 |
+
self.output_token_len_map[output_token_len] = i
|
401 |
+
self.lm_heads = nn.ModuleList(lm_head_list)
|
402 |
+
self.loss_function = torch.nn.MSELoss(reduction='none')
|
403 |
+
self.post_init()
|
404 |
+
|
405 |
+
def set_decoder(self, decoder):
|
406 |
+
self.model = decoder
|
407 |
+
|
408 |
+
def get_decoder(self):
|
409 |
+
return self.model
|
410 |
+
|
411 |
+
def forward(
|
412 |
+
self,
|
413 |
+
input_ids: torch.FloatTensor = None,
|
414 |
+
attention_mask: Optional[torch.Tensor] = None,
|
415 |
+
position_ids: Optional[torch.LongTensor] = None,
|
416 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
417 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
418 |
+
labels: Optional[torch.FloatTensor] = None,
|
419 |
+
loss_masks: Optional[torch.FloatTensor] = None,
|
420 |
+
use_cache: Optional[bool] = None,
|
421 |
+
output_attentions: Optional[bool] = None,
|
422 |
+
output_hidden_states: Optional[bool] = None,
|
423 |
+
return_dict: Optional[bool] = None,
|
424 |
+
max_output_length: Optional[int] = None,
|
425 |
+
revin: Optional[bool] = False,
|
426 |
+
) -> Union[Tuple, MoeCausalLMOutputWithPast]:
|
427 |
+
|
428 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
429 |
+
output_hidden_states = (
|
430 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
431 |
+
)
|
432 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
433 |
+
|
434 |
+
if revin:
|
435 |
+
mean, std = input_ids.mean(dim=-1, keepdim=True), input_ids.std(dim=-1, keepdim=True)
|
436 |
+
input_ids = (input_ids - mean) / std
|
437 |
+
outputs = self.model(
|
438 |
+
input_ids=input_ids,
|
439 |
+
attention_mask=attention_mask,
|
440 |
+
position_ids=position_ids,
|
441 |
+
past_key_values=past_key_values,
|
442 |
+
inputs_embeds=inputs_embeds,
|
443 |
+
use_cache=use_cache,
|
444 |
+
output_attentions=output_attentions,
|
445 |
+
output_hidden_states=output_hidden_states,
|
446 |
+
return_dict=return_dict,
|
447 |
+
)
|
448 |
+
|
449 |
+
hidden_states = outputs[0] if not return_dict else outputs.last_hidden_state
|
450 |
+
predictions = None
|
451 |
+
|
452 |
+
loss = None
|
453 |
+
if labels is not None:
|
454 |
+
ar_loss = 0.0
|
455 |
+
for lm_head, output_token_len in zip(self.lm_heads, self.config.output_token_lens):
|
456 |
+
one_predictions = lm_head(hidden_states)
|
457 |
+
one_loss = self.calc_ar_loss(
|
458 |
+
one_predictions, labels, loss_masks, output_token_len)
|
459 |
+
ar_loss += one_loss
|
460 |
+
if predictions is None:
|
461 |
+
predictions = one_predictions
|
462 |
+
loss = ar_loss / len(self.config.output_token_lens)
|
463 |
+
else:
|
464 |
+
if max_output_length is None:
|
465 |
+
output_token_len = self.config.output_token_lens[0]
|
466 |
+
max_output_length = output_token_len
|
467 |
+
else:
|
468 |
+
output_token_len = self.config.output_token_lens[0]
|
469 |
+
for h in self.config.output_token_lens[1:]:
|
470 |
+
if h > max_output_length:
|
471 |
+
break
|
472 |
+
else:
|
473 |
+
output_token_len = h
|
474 |
+
lm_head = self.lm_heads[self.output_token_len_map[output_token_len]]
|
475 |
+
predictions = lm_head(hidden_states)
|
476 |
+
if output_token_len > max_output_length:
|
477 |
+
predictions = predictions[:, :, :max_output_length]
|
478 |
+
if revin:
|
479 |
+
predictions = predictions * std + mean
|
480 |
+
if not return_dict:
|
481 |
+
output = (predictions,) + outputs[1:]
|
482 |
+
return (loss) + output if loss is not None else output
|
483 |
+
|
484 |
+
return MoeCausalLMOutputWithPast(
|
485 |
+
loss=loss,
|
486 |
+
logits=predictions,
|
487 |
+
past_key_values=outputs.past_key_values,
|
488 |
+
hidden_states=outputs.hidden_states,
|
489 |
+
attentions=outputs.attentions,
|
490 |
+
)
|
491 |
+
|
492 |
+
def calc_ar_loss(self, predictions, labels, loss_masks, output_token_len):
|
493 |
+
seq_len = predictions.shape[1] * self.config.input_token_len
|
494 |
+
labels = labels[:, :seq_len -
|
495 |
+
self.config.input_token_len + output_token_len]
|
496 |
+
shift_labels = labels.unfold(
|
497 |
+
dimension=-1, size=output_token_len, step=self.config.input_token_len)
|
498 |
+
|
499 |
+
# Calculate loss with mask
|
500 |
+
losses = self.loss_function(predictions, shift_labels).mean(dim=-1)
|
501 |
+
if loss_masks is not None:
|
502 |
+
losses = losses * loss_masks
|
503 |
+
loss = losses.sum() / loss_masks.sum()
|
504 |
+
else:
|
505 |
+
loss = torch.mean(losses)
|
506 |
+
|
507 |
+
return loss
|
508 |
+
|
509 |
+
def prepare_inputs_for_generation(
|
510 |
+
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, revin=True, **kwargs
|
511 |
+
):
|
512 |
+
# Omit tokens covered by past_key_values
|
513 |
+
if past_key_values is not None:
|
514 |
+
if isinstance(past_key_values, Cache):
|
515 |
+
cache_length = past_key_values.get_seq_length()
|
516 |
+
if isinstance(past_key_values, DynamicCache):
|
517 |
+
past_length = past_key_values.seen_tokens
|
518 |
+
else:
|
519 |
+
past_length = cache_length
|
520 |
+
|
521 |
+
max_cache_length = past_key_values.get_max_length()
|
522 |
+
else:
|
523 |
+
cache_length = past_length = past_key_values[0][0].shape[2]
|
524 |
+
max_cache_length = None
|
525 |
+
|
526 |
+
# Keep only the unprocessed tokens:
|
527 |
+
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
528 |
+
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
529 |
+
# input)
|
530 |
+
if attention_mask is not None and attention_mask.shape[1] > (input_ids.shape[1] // self.config.input_token_len):
|
531 |
+
input_ids = input_ids[:, -
|
532 |
+
(attention_mask.shape[1] - past_length):]
|
533 |
+
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
|
534 |
+
# input_ids based on the past_length.
|
535 |
+
elif past_length < (input_ids.shape[1] // self.config.input_token_len):
|
536 |
+
input_ids = input_ids[:, past_length *
|
537 |
+
self.config.input_token_len:]
|
538 |
+
# 3 - Otherwise (past_length >= (input_ids.shape[1] // self.config.input_token_len)), let's assume input_ids only has unprocessed tokens.
|
539 |
+
|
540 |
+
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
|
541 |
+
if (
|
542 |
+
max_cache_length is not None
|
543 |
+
and attention_mask is not None
|
544 |
+
and cache_length + (input_ids.shape[1] // self.config.input_token_len) > max_cache_length
|
545 |
+
):
|
546 |
+
attention_mask = attention_mask[:, -max_cache_length:]
|
547 |
+
|
548 |
+
position_ids = kwargs.get("position_ids", None)
|
549 |
+
if attention_mask is not None and position_ids is None:
|
550 |
+
# create position_ids on the fly for batch generation
|
551 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
552 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
553 |
+
if past_key_values:
|
554 |
+
position_ids = position_ids[:, -
|
555 |
+
(input_ids.shape[1] // self.config.input_token_len):]
|
556 |
+
|
557 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
558 |
+
if inputs_embeds is not None and past_key_values is None:
|
559 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
560 |
+
else:
|
561 |
+
model_inputs = {"input_ids": input_ids}
|
562 |
+
|
563 |
+
model_inputs.update(
|
564 |
+
{
|
565 |
+
"position_ids": position_ids,
|
566 |
+
"past_key_values": past_key_values,
|
567 |
+
"use_cache": kwargs.get("use_cache"),
|
568 |
+
"attention_mask": attention_mask,
|
569 |
+
"revin": revin
|
570 |
+
}
|
571 |
+
)
|
572 |
+
return model_inputs
|