Upload model.py
Browse files
model.py
CHANGED
@@ -1,112 +1,35 @@
|
|
1 |
-
"""
|
|
|
|
|
2 |
|
3 |
-
Based on the nanoGPT implementation: https://github.com/karpathy/nanoGPT and
|
4 |
-
https://github.com/EleutherAI/gpt-neox/tree/main/megatron/model.
|
5 |
"""
|
|
|
|
|
6 |
import math, random
|
7 |
import numpy as np
|
8 |
from typing import Any, List, Optional, Tuple
|
|
|
|
|
9 |
|
10 |
import torch
|
11 |
import torch.nn as nn
|
12 |
-
from lightning_utilities.core.imports import RequirementCache
|
13 |
-
from typing_extensions import Self
|
14 |
-
from flash_attn import flash_attn_func
|
15 |
-
# from lit_gpt.config import Config
|
16 |
-
from xformers.ops import SwiGLU
|
17 |
-
|
18 |
import torch.nn.functional as F
|
19 |
-
# from .fused_rotary_embedding import apply_rotary_emb_func
|
20 |
-
RoPECache = Tuple[torch.Tensor, torch.Tensor]
|
21 |
-
KVCache = Tuple[torch.Tensor, torch.Tensor]
|
22 |
-
PretokenCache = torch.Tensor
|
23 |
-
# Tuple[torch.Tensor, torch.Tensor]
|
24 |
-
FlashAttention2Available = RequirementCache("flash-attn>=2.0.0.post1")
|
25 |
-
from einops import rearrange
|
26 |
-
from transformers import PreTrainedModel, Cache, DynamicCache
|
27 |
-
|
28 |
-
from huggingface_hub import PyTorchModelHubMixin
|
29 |
-
from .model_config import YingLongConfig
|
30 |
-
|
31 |
-
# from torch.distributions import Normal, LowRankMultivariateNormal, kl_divergence,MultivariateNormal
|
32 |
|
33 |
-
class quantitleLoss(torch.nn.Module):
|
34 |
-
def __init__(self,
|
35 |
-
qSize = 99,
|
36 |
-
patch_size = 16,
|
37 |
-
*args,**kwargs) -> None:
|
38 |
-
|
39 |
-
super().__init__()
|
40 |
-
self.qSize = qSize
|
41 |
-
self.patch_size = patch_size
|
42 |
-
|
43 |
-
|
44 |
-
q = np.array([i+1 for i in range(self.qSize)])
|
45 |
-
q = q / (self.qSize + 1)
|
46 |
-
q = q.reshape((1,1,-1))
|
47 |
-
|
48 |
-
q_variance = q*(1-q)
|
49 |
-
|
50 |
-
self.register_buffer('q', torch.tensor(q))
|
51 |
-
self.register_buffer('q_variance', torch.tensor(q_variance))
|
52 |
|
53 |
-
|
54 |
-
|
55 |
-
|
56 |
-
|
57 |
-
|
58 |
-
target = target.unsqueeze(-1)
|
59 |
-
input = input[:,:target.shape[1],:,:]
|
60 |
-
|
61 |
-
|
62 |
-
posPart = input - target
|
63 |
-
negPart = -posPart
|
64 |
-
|
65 |
-
raw_loss = torch.maximum(self.q * negPart, (1-self.q) * posPart)
|
66 |
|
67 |
-
|
68 |
-
|
69 |
-
|
70 |
-
return torch.mean(raw_loss)
|
71 |
-
|
72 |
|
73 |
-
def haarMatrix_unnormalized(n):
|
74 |
-
# Allow only size n of power 2
|
75 |
-
n = 2**np.ceil(np.log2(n))
|
76 |
-
if n > 2:
|
77 |
-
h = haarMatrix(n / 2)
|
78 |
-
else:
|
79 |
-
return np.array([[1, 1], [1, -1]])
|
80 |
-
|
81 |
-
# calculate upper haar part
|
82 |
-
h_n = np.kron(h, [1, 1])
|
83 |
-
# calculate lower haar part
|
84 |
-
# if normalized:
|
85 |
-
# h_i = np.sqrt(n/2)*np.kron(np.eye(len(h)), [1, -1])
|
86 |
-
# else:
|
87 |
-
h_i = np.kron(np.eye(len(h)), [1, -1])
|
88 |
-
# combine parts
|
89 |
-
h = np.vstack((h_n, h_i))
|
90 |
-
return h
|
91 |
|
|
|
|
|
92 |
|
93 |
-
def haarMatrix(n,normalized = 'ortho'):
|
94 |
-
h = haarMatrix_unnormalized(n)
|
95 |
-
scaler = np.diag(1/np.sqrt(np.diag(h@h.transpose())))
|
96 |
-
if normalized == 'ortho':
|
97 |
-
return scaler @ h
|
98 |
-
elif normalized == 'forward':
|
99 |
-
return scaler @ h/ np.sqrt(n)
|
100 |
-
|
101 |
-
else:
|
102 |
-
return scaler @ h * np.sqrt(n)
|
103 |
-
# else:
|
104 |
-
# scaler = 1
|
105 |
|
106 |
|
107 |
|
108 |
-
|
109 |
-
|
110 |
class Tokenizer(torch.nn.Module):
|
111 |
def __init__(self, config: YingLongConfig, *args,**kwargs) -> None:
|
112 |
super().__init__()
|
@@ -119,6 +42,7 @@ class Tokenizer(torch.nn.Module):
|
|
119 |
|
120 |
self.register_buffer('mask_token', torch.zeros(1000))
|
121 |
if self.config.haar_trans:
|
|
|
122 |
self.register_buffer('haar_transform',torch.Tensor(haarMatrix(self.config.patch_size,normalized = self.config.haar_trans_norm)))
|
123 |
|
124 |
|
@@ -167,7 +91,6 @@ class Tokenizer(torch.nn.Module):
|
|
167 |
|
168 |
else:
|
169 |
|
170 |
-
|
171 |
factor = 1
|
172 |
more_rows = future_token // self.patch_size + 1
|
173 |
prev_more_rows = prev_token // self.patch_size + 1
|
@@ -187,7 +110,6 @@ class Tokenizer(torch.nn.Module):
|
|
187 |
masks = [jj for jj in range(x_featured.shape[1])]
|
188 |
masks = masks[-more_rows:]
|
189 |
|
190 |
-
# if not mean_replace:
|
191 |
x_featured[:,-more_rows:] = self.mask0(self.mask_token[:len(masks)].unsqueeze(-1)).repeat(x_featured.shape[0],1,1)
|
192 |
x_featured[:,:prev_more_rows] = self.mask0(self.mask_token[:prev_more_rows].unsqueeze(-1)).repeat(x_featured.shape[0],1,1)
|
193 |
|
@@ -199,26 +121,30 @@ class Tokenizer(torch.nn.Module):
|
|
199 |
class model_tmp(PreTrainedModel):
|
200 |
config_class = YingLongConfig
|
201 |
base_model_prefix = "model"
|
202 |
-
# supports_gradient_checkpointing = True
|
203 |
-
# _no_split_modules = ["TimeMoeDecoderLayer"]
|
204 |
-
# _skip_keys_device_placement = "past_key_values"
|
205 |
-
_supports_flash_attn_2 = True
|
206 |
-
_supports_sdpa = False
|
207 |
-
_supports_cache_class = True
|
208 |
|
209 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
210 |
class GPT(model_tmp):
|
211 |
def __init__(self, config: YingLongConfig, *args,**kwargs) -> None:
|
212 |
|
213 |
|
214 |
-
# config_class = YingLongConfig
|
215 |
-
# base_model_prefix = "model"
|
216 |
-
# # supports_gradient_checkpointing = True
|
217 |
-
# # _no_split_modules = ["TimeMoeDecoderLayer"]
|
218 |
-
# # _skip_keys_device_placement = "past_key_values"
|
219 |
-
# _supports_flash_attn_2 = True
|
220 |
-
# _supports_sdpa = False
|
221 |
-
# _supports_cache_class = True
|
222 |
super().__init__(config)
|
223 |
|
224 |
self.config = config
|
@@ -227,49 +153,28 @@ class GPT(model_tmp):
|
|
227 |
|
228 |
|
229 |
if self.config._norm_class == "RMSNorm":
|
230 |
-
|
231 |
self.config.norm_class = RMSNorm
|
232 |
elif self.config._norm_class == "FusedRMSNorm":
|
233 |
-
# from .model import FusedRMSNorm
|
234 |
self.config.norm_class = FusedRMSNorm
|
235 |
elif self.config._norm_class == 'BatchNorm':
|
236 |
-
# from .model import iBatchNorm
|
237 |
self.config.norm_class = iBatchNorm
|
238 |
|
239 |
-
|
240 |
|
241 |
if self.config._mlp_class == "GptNeoxMLP":
|
242 |
-
# from .model import GptNeoxMLP
|
243 |
self.config.mlp_class = GptNeoxMLP
|
244 |
elif self.config._mlp_class == "LLaMAMLP":
|
245 |
-
# from .model import LLaMAMLP
|
246 |
self.config.mlp_class = LLaMAMLP
|
247 |
|
248 |
-
|
249 |
-
|
250 |
-
self.stat_tokens = 1
|
251 |
-
else:
|
252 |
-
self.stat_tokens = 0
|
253 |
-
|
254 |
-
|
255 |
-
|
256 |
-
|
257 |
|
258 |
self.tokenizer = Tokenizer(config)
|
259 |
|
260 |
-
|
261 |
-
# nn.Linear(config.n_embd, config.n_embd*4),
|
262 |
-
# nn.ReLU(),
|
263 |
-
# nn.Linear(config.n_embd*4, 99*self.patch_size),
|
264 |
-
# )
|
265 |
-
|
266 |
-
|
267 |
self.lm_head = nn.Linear(config.n_embd, 99*self.patch_size)
|
268 |
-
|
269 |
-
|
270 |
-
# self.gate = nn.Linear(config.n_embd, 1)
|
271 |
|
272 |
-
|
273 |
self.quantitleLoss = quantitleLoss(99,patch_size = self.patch_size)
|
274 |
|
275 |
|
@@ -296,48 +201,16 @@ class GPT(model_tmp):
|
|
296 |
|
297 |
|
298 |
|
299 |
-
self.rope_cache
|
300 |
-
self.mask_cache: Optional[torch.Tensor] = None
|
301 |
-
self.kv_caches: List[KVCache] = []
|
302 |
|
303 |
|
304 |
-
def _init_weights(self, module: nn.Module) -> None:
|
305 |
-
"""Meant to be used with `gpt.apply(gpt._init_weights)`."""
|
306 |
-
# GPT-NeoX https://arxiv.org/pdf/2204.06745.pdf
|
307 |
-
if isinstance(module, nn.Embedding):
|
308 |
-
torch.nn.init.normal_(module.weight, mean=0.0, std=math.sqrt(2.0 / 5 / self.config.n_embd))
|
309 |
-
# RWKV: set it to 1e-4
|
310 |
-
# torch.nn.init.uniform_(module.weight, -1e-4, 1e-4)
|
311 |
-
elif isinstance(module, nn.Linear):
|
312 |
-
torch.nn.init.normal_(module.weight, mean=0.0, std=math.sqrt(2.0 / 5 / self.config.n_embd))
|
313 |
-
if module.bias is not None:
|
314 |
-
torch.nn.init.zeros_(module.bias)
|
315 |
-
# GPT-NeoX
|
316 |
-
for name, p in module.named_parameters():
|
317 |
-
if (name == "proj.weight" and isinstance(module, LLaMAMLP)) or (name == "w3.weight" and isinstance(module, SwiGLU) or (name=="proj.weight" and isinstance(module, BidirectedlSelfAttention))): #if use xformer swiglu, fc2 layer will be renamed to w3
|
318 |
-
nn.init.normal_(p, mean=0.0, std=1 / math.sqrt(self.config.n_embd) / self.config.n_layer)
|
319 |
-
|
320 |
-
|
321 |
-
def reset_cache(self) -> None:
|
322 |
-
self.kv_caches.clear()
|
323 |
-
if self.mask_cache is not None and self.mask_cache.device.type == "xla":
|
324 |
-
# https://github.com/Lightning-AI/lit-gpt/pull/83#issuecomment-1558150179
|
325 |
-
self.rope_cache = None
|
326 |
-
self.mask_cache = None
|
327 |
|
328 |
def forward(
|
329 |
self, idx: torch.Tensor,
|
330 |
-
max_seq_length: Optional[int] = None,
|
331 |
-
input_pos: Optional[torch.Tensor] = None,
|
332 |
-
next_token: torch.Tensor = None,
|
333 |
future_token: int = 0,
|
334 |
prev_token: int = 0,
|
335 |
-
val: bool = False,
|
336 |
-
print_intermediate: bool = False,
|
337 |
-
cot_rounds: int = -1,
|
338 |
-
sequential: bool = False,
|
339 |
*args,**kwargs,
|
340 |
-
|
341 |
|
342 |
if future_token > 0:
|
343 |
more_rows = future_token // self.patch_size + 1
|
@@ -348,138 +221,53 @@ class GPT(model_tmp):
|
|
348 |
|
349 |
B, T = idx.size()
|
350 |
|
351 |
-
|
352 |
|
353 |
block_size = self.config.block_size
|
354 |
-
|
355 |
-
max_seq_length = block_size
|
356 |
-
|
357 |
|
358 |
-
if use_kv_cache: # not relevant otherwise
|
359 |
-
assert (
|
360 |
-
max_seq_length >= T
|
361 |
-
), f"Cannot forward sequence of length {T}, max seq length is only {max_seq_length}"
|
362 |
assert max_seq_length <= block_size, f"Cannot attend to {max_seq_length}, block size is only {block_size}"
|
363 |
-
|
364 |
-
|
365 |
-
|
366 |
-
self.mask_cache = self.build_mask_cache(idx)
|
367 |
cos, sin = self.rope_cache
|
368 |
-
if use_kv_cache:
|
369 |
-
if self.stat_tokens:
|
370 |
-
if len(input_pos) == 1:
|
371 |
-
idx = idx[:,input_pos]
|
372 |
-
input_pos = input_pos.add_(1)
|
373 |
-
else:
|
374 |
-
input_pos = torch.arange(0, input_pos[-1]+2, device=idx.device)
|
375 |
-
|
376 |
-
cos = cos.index_select(0, input_pos)
|
377 |
-
sin = sin.index_select(0, input_pos)
|
378 |
-
mask = self.mask_cache.index_select(2, input_pos)
|
379 |
-
mask = mask[:, :, :, :max_seq_length]
|
380 |
-
|
381 |
-
else:
|
382 |
-
cos = cos.index_select(0, input_pos)
|
383 |
-
sin = sin.index_select(0, input_pos)
|
384 |
-
idx = idx[:,input_pos]
|
385 |
-
else:
|
386 |
-
cos = cos[:max(T,1024) + self.stat_tokens]
|
387 |
-
sin = sin[:max(T,1024) + self.stat_tokens]
|
388 |
-
mask = None
|
389 |
|
390 |
-
|
|
|
391 |
|
|
|
392 |
|
393 |
|
394 |
-
|
395 |
-
pass
|
396 |
-
else:
|
397 |
-
x,x_raw,masks,mean,std,x_0 = self.tokenizer(idx,
|
398 |
-
future_token =future_token,
|
399 |
-
prev_token = prev_token,
|
400 |
-
sequential = sequential,
|
401 |
-
)
|
402 |
|
403 |
|
404 |
|
405 |
-
|
406 |
if self.unet:
|
407 |
skips = []
|
408 |
|
409 |
|
410 |
|
411 |
-
res_intermediate = []
|
412 |
-
target_intermediate = []
|
413 |
-
if not use_kv_cache:
|
414 |
-
|
415 |
-
if cot_rounds <0:
|
416 |
-
cot_rounds = self.config.n_cot
|
417 |
-
|
418 |
-
res_list = []
|
419 |
-
gate_list = []
|
420 |
-
for rep in range(cot_rounds):
|
421 |
-
for block_idx in range(len( self.transformer.h)):
|
422 |
-
|
423 |
|
|
|
424 |
|
425 |
-
block = self.transformer.h[block_idx]
|
426 |
|
427 |
-
|
428 |
-
x = self.unet_projection[block_idx - len(self.transformer.h) //2](torch.cat((skips.pop(),x),dim = -1))
|
429 |
|
430 |
-
|
431 |
-
|
432 |
-
if self.unet and block_idx <len(self.transformer.h) //2:
|
433 |
-
skips.append(x)
|
434 |
-
x_delay = torch.cat((x[:,0,:].unsqueeze(1),x[:,:-1,:]),dim = 1)
|
435 |
-
x = self.unet_merge[block_idx](torch.cat((x_delay,x),dim = -1))
|
436 |
-
# if block_idx <len(self.transformer.h) //2:
|
437 |
-
# x_delay = torch.cat((x[:,0,:].unsqueeze(1),x[:,:-1,:]),dim = 1)
|
438 |
-
# x = self.unet_merge[block_idx](torch.cat((x_delay,x),dim = -1))
|
439 |
-
|
440 |
-
|
441 |
-
|
442 |
|
443 |
-
|
444 |
-
# gate_list.append(self.gate(x).unsqueeze(-1))
|
445 |
-
# gate_list.append(self.gate(x))
|
446 |
-
# if print_intermediate:
|
447 |
-
# res_intermediate.append(res_list[-1])
|
448 |
-
# if print_intermediate:
|
449 |
-
# res_tmp = self.lm_head(x[:,self.stat_tokens:])
|
450 |
-
# res_tmp = rearrange(res_tmp,'b c (l1 l2) -> b c l1 l2', l2 = 99)
|
451 |
-
# if self.config.haar_trans_inv:
|
452 |
-
|
453 |
-
# res_tmp = torch.einsum('bcal,ad->bcdl',res_tmp,self.tokenizer.haar_transform)
|
454 |
-
# if self.config.haar_trans_norm == "backward":
|
455 |
-
# res_tmp = res_tmp / np.sqrt(res_tmp.shape[-2])
|
456 |
-
# elif self.config.haar_trans_norm == "forward":
|
457 |
-
# res_tmp = res_tmp * np.sqrt(res_tmp.shape[-2])
|
458 |
-
# res_tmp = res_tmp * (std.unsqueeze(-1) + 1e-4) + mean.unsqueeze(-1)
|
459 |
-
# res_intermediate.append(res_tmp[:,masks,:,:])
|
460 |
|
461 |
-
|
|
|
|
|
|
|
462 |
|
463 |
-
|
464 |
-
|
465 |
-
else:
|
466 |
-
self.kv_caches = self.kv_caches or self.build_kv_caches(x, max_seq_length, cos.size(-1) * 2)
|
467 |
-
for block_idx in range(len( self.transformer.h)):
|
468 |
-
block = self.transformer.h[block_idx]
|
469 |
-
if self.unet and block_idx >=len(self.transformer.h) //2:
|
470 |
-
x = F.silu(skips.pop()) * x
|
471 |
-
x, self.kv_caches[block_idx] = block(x, (cos, sin), max_seq_length, mask, input_pos, self.kv_caches[block_idx])
|
472 |
-
if self.unet and block_idx <len(self.transformer.h) //2:
|
473 |
-
skips.append(x)
|
474 |
-
|
475 |
-
|
476 |
|
477 |
|
478 |
res = self.lm_head(x)
|
479 |
-
|
480 |
-
# gate = F.softmax(gate,dim = -1)
|
481 |
-
# res = torch.cat(res_list,dim = -1) * gate
|
482 |
-
# res = res.sum(dim = -1)
|
483 |
|
484 |
|
485 |
res = rearrange(res,'b c (l1 l2) -> b c l1 l2', l2 = 99)
|
@@ -487,7 +275,6 @@ class GPT(model_tmp):
|
|
487 |
|
488 |
|
489 |
if self.config.haar_trans_inv:
|
490 |
-
# print('res',res.shape,self.tokenizer.haar_transform.shape)
|
491 |
res = torch.einsum('bcal,ad->bcdl',res,self.tokenizer.haar_transform)
|
492 |
if self.config.haar_trans_norm == "backward":
|
493 |
res = res / np.sqrt(res.shape[-2])
|
@@ -495,32 +282,33 @@ class GPT(model_tmp):
|
|
495 |
res = res * np.sqrt(res.shape[-2])
|
496 |
|
497 |
|
|
|
|
|
|
|
498 |
res = res * (std.unsqueeze(-1) + 1e-4) + mean.unsqueeze(-1)
|
499 |
|
500 |
|
501 |
|
502 |
-
|
503 |
if future_token == 0:
|
504 |
-
return res[:,masks,:,:], x_raw[:,masks,:]
|
505 |
else:
|
506 |
-
return res[:,masks,:,:]
|
507 |
|
508 |
def generate(self,*args,**kwargs):
|
509 |
-
|
510 |
-
res, _ = self.forward(*args,**kwargs)
|
511 |
-
# logits_all,res_intermediate = model(idx = x_train, future_token = (pred_len//32 + 1)* 32, prev_token = 0,print_intermediate = False,cot_rounds = 1)
|
512 |
-
|
513 |
res = rearrange(res, 'b l c d -> b (l c) d')
|
514 |
return res[:,:kwargs['future_token'],:]
|
515 |
|
516 |
|
|
|
517 |
@classmethod
|
518 |
def from_name(cls, name: str, **kwargs: Any) -> Self:
|
519 |
return cls(Config.from_name(name, **kwargs))
|
520 |
|
521 |
-
def build_rope_cache(self, idx: torch.Tensor)
|
522 |
return build_rope_cache(
|
523 |
-
seq_len=self.config.block_size
|
524 |
n_elem=int(self.config.rotary_percentage * self.config.head_size),
|
525 |
dtype=torch.bfloat16,
|
526 |
device=idx.device,
|
@@ -528,27 +316,6 @@ class GPT(model_tmp):
|
|
528 |
condense_ratio=self.config.condense_ratio,
|
529 |
)
|
530 |
|
531 |
-
def build_mask_cache(self, idx: torch.Tensor) -> torch.Tensor:
|
532 |
-
ones = torch.ones((self.config.block_size+self.stat_tokens, self.config.block_size+self.stat_tokens), device=idx.device, dtype=torch.bool)
|
533 |
-
return torch.tril(ones).unsqueeze(0).unsqueeze(0)
|
534 |
-
|
535 |
-
def build_kv_caches(self, idx: torch.Tensor, max_seq_length: int, rope_cache_length: int) -> List[KVCache]:
|
536 |
-
B = idx.size(0)
|
537 |
-
heads = 1 if self.config.n_query_groups == 1 else self.config.n_query_groups
|
538 |
-
|
539 |
-
k_cache_shape = (
|
540 |
-
B,
|
541 |
-
max_seq_length,
|
542 |
-
heads,
|
543 |
-
rope_cache_length + self.config.head_size - int(self.config.rotary_percentage * self.config.head_size),
|
544 |
-
)
|
545 |
-
v_cache_shape = (B, max_seq_length, heads, self.config.head_size)
|
546 |
-
device = idx.device
|
547 |
-
return [
|
548 |
-
(torch.zeros(k_cache_shape, device=device), torch.zeros(v_cache_shape, device=device))
|
549 |
-
for _ in range(self.config.n_layer)
|
550 |
-
]
|
551 |
-
|
552 |
|
553 |
class Block(nn.Module):
|
554 |
def __init__(self, config:YingLongConfig) -> None:
|
@@ -562,15 +329,14 @@ class Block(nn.Module):
|
|
562 |
def forward(
|
563 |
self,
|
564 |
x: torch.Tensor,
|
565 |
-
rope:
|
566 |
max_seq_length: int,
|
567 |
mask: Optional[torch.Tensor] = None,
|
568 |
input_pos: Optional[torch.Tensor] = None,
|
569 |
-
|
570 |
-
) -> Tuple[torch.Tensor, Optional[KVCache]]:
|
571 |
|
572 |
n_1 = self.norm_1(x)
|
573 |
-
h
|
574 |
if self.config.parallel_residual:
|
575 |
n_2 = n_1 if self.config.shared_attention_norm else self.norm_2(x)
|
576 |
x = x + h + self.mlp(n_2)
|
@@ -583,29 +349,25 @@ class Block(nn.Module):
|
|
583 |
|
584 |
x = x + h
|
585 |
x = x + self.mlp(self.norm_2(x))
|
586 |
-
return x
|
587 |
|
588 |
|
589 |
class BidirectedlSelfAttention(nn.Module):
|
590 |
def __init__(self, config:YingLongConfig) -> None:
|
591 |
super().__init__()
|
592 |
shape = (config.n_head + 2 * config.n_query_groups) * config.head_size
|
593 |
-
# key, query, value projections for all heads, but in a batch
|
594 |
self.attn = nn.Linear(config.n_embd, shape, bias=config.bias)
|
595 |
-
# output projection
|
596 |
self.proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
|
597 |
-
|
598 |
self.config = config
|
599 |
|
600 |
def forward(
|
601 |
self,
|
602 |
x: torch.Tensor,
|
603 |
-
rope:
|
604 |
max_seq_length: int,
|
605 |
mask: Optional[torch.Tensor] = None,
|
606 |
input_pos: Optional[torch.Tensor] = None,
|
607 |
-
|
608 |
-
) -> Tuple[torch.Tensor, Optional[KVCache]]:
|
609 |
|
610 |
|
611 |
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
|
@@ -616,52 +378,20 @@ class BidirectedlSelfAttention(nn.Module):
|
|
616 |
q_per_kv = self.config.n_head // self.config.n_query_groups
|
617 |
total_qkv = q_per_kv + 2 # each group has 1+ queries, 1 key, and 1 value
|
618 |
qkv = qkv.view(B, T, self.config.n_query_groups, total_qkv, self.config.head_size) # (B, T, n_query_groups, total_qkv, hs)
|
619 |
-
|
620 |
|
621 |
# split batched computation into three
|
622 |
q, k, v = qkv.split((q_per_kv, 1, 1), dim=-2)
|
623 |
|
624 |
-
# repeat k and v if necessary
|
625 |
-
# Peiyuan: we do not need to do this as flash attention 2 already support GQA
|
626 |
-
# if self.config.n_query_groups != 1: # doing this would require a full kv cache with MQA (inefficient!)
|
627 |
-
# # for MHA this is a no-op
|
628 |
-
# k = k.expand(B, self.config.n_query_groups, q_per_kv, T, self.config.head_size)
|
629 |
-
# v = v.expand(B, self.config.n_query_groups, q_per_kv, T, self.config.head_size)
|
630 |
-
|
631 |
q = q.reshape(B, T, -1, self.config.head_size) # (B, T, nh_q, hs)
|
632 |
k = k.reshape(B, T, -1, self.config.head_size)
|
633 |
v = v.reshape(B, T, -1, self.config.head_size)
|
634 |
|
635 |
cos, sin = rope
|
636 |
|
637 |
-
# apply rope in fp32 significanly stabalize training
|
638 |
-
# fused rope expect (batch_size, seqlen, nheads, headdim)
|
639 |
q = apply_rotary_emb_func(q, cos, sin, False, True)
|
640 |
k = apply_rotary_emb_func(k, cos, sin, False, True)
|
641 |
-
|
642 |
-
# n_elem = int(self.config.rotary_percentage * self.config.head_size)
|
643 |
-
|
644 |
-
# q_roped = apply_rope(q[..., :n_elem], cos.repeat(1,2), sin.repeat(1,2))
|
645 |
-
# k_roped = apply_rope(k[..., :n_elem], cos.repeat(1,2), sin.repeat(1,2))
|
646 |
-
# print( (q_roped - q).sum())
|
647 |
-
# q = torch.cat((q_roped, q[..., n_elem:]), dim=-1)
|
648 |
-
# k = torch.cat((k_roped, k[..., n_elem:]), dim=-1)
|
649 |
-
|
650 |
-
if kv_cache is not None:
|
651 |
-
cache_k, cache_v = kv_cache
|
652 |
-
cache_k, cache_v = cache_k.to(dtype=k.dtype), cache_v.to(dtype=v.dtype)
|
653 |
-
# check if reached token limit
|
654 |
-
if input_pos[-1] >= max_seq_length:
|
655 |
-
input_pos = torch.tensor(max_seq_length - 1, device=input_pos.device)
|
656 |
-
# shift 1 position to the left
|
657 |
-
cache_k = torch.roll(cache_k, -1, dims=1)
|
658 |
-
cache_v = torch.roll(cache_v, -1, dims=1)
|
659 |
-
|
660 |
-
k = cache_k.index_copy_(1, input_pos, k)
|
661 |
-
v = cache_v.index_copy_(1, input_pos, v)
|
662 |
-
kv_cache = k, v
|
663 |
|
664 |
-
|
665 |
|
666 |
y = self.scaled_dot_product_attention(q, k, v, mask=mask)
|
667 |
|
@@ -670,7 +400,9 @@ class BidirectedlSelfAttention(nn.Module):
|
|
670 |
# output projection
|
671 |
y = self.proj(y)
|
672 |
|
673 |
-
return y
|
|
|
|
|
674 |
|
675 |
def scaled_dot_product_attention(
|
676 |
self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, mask: Optional[torch.Tensor] = None
|
@@ -690,14 +422,83 @@ class BidirectedlSelfAttention(nn.Module):
|
|
690 |
k = k.transpose(1, 2)
|
691 |
v = v.transpose(1, 2)
|
692 |
if q.size() != k.size():
|
693 |
-
|
694 |
-
|
695 |
y = torch.nn.functional.scaled_dot_product_attention(
|
696 |
q, k, v, attn_mask=mask, dropout_p=0.0, scale=scale, is_causal=False
|
697 |
)
|
698 |
return y.transpose(1, 2)
|
699 |
|
700 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
701 |
class GptNeoxMLP(nn.Module):
|
702 |
def __init__(self, config:YingLongConfig) -> None:
|
703 |
super().__init__()
|
@@ -713,21 +514,15 @@ class GptNeoxMLP(nn.Module):
|
|
713 |
class LLaMAMLP(nn.Module):
|
714 |
def __init__(self, config:YingLongConfig) -> None:
|
715 |
super().__init__()
|
716 |
-
|
717 |
-
# self.fc_2 = nn.Linear(config.n_embd, config.intermediate_size, bias=config.bias)
|
718 |
-
# self.proj = nn.Linear(config.intermediate_size, config.n_embd, bias=config.bias)
|
719 |
self.swiglu = SwiGLU(config.n_embd,config.intermediate_size, bias=False, _pack_weights=False)
|
720 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
721 |
-
# x_fc_1 = self.fc_1(x)
|
722 |
-
# x_fc_2 = self.fc_2(x)
|
723 |
-
# x = torch.nn.functional.silu(x_fc_1) * x_fc_2
|
724 |
-
# return self.proj(x)
|
725 |
return self.swiglu(x)
|
726 |
|
727 |
|
728 |
def build_rope_cache(
|
729 |
seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000, condense_ratio: int = 1
|
730 |
-
) ->
|
731 |
"""Enhanced Transformer with Rotary Position Embedding.
|
732 |
|
733 |
Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
|
@@ -745,7 +540,6 @@ def build_rope_cache(
|
|
745 |
|
746 |
cos, sin = torch.cos(idx_theta), torch.sin(idx_theta)
|
747 |
|
748 |
-
# print(' print(seq_idx.shape,theta.shape,sin.shape,cos.shape,idx_theta.shape)',seq_idx.shape,theta.shape,sin.shape,cos.shape,idx_theta.shape)
|
749 |
# added by peiyuan to ensure same data type with q, k, to use fused rotary embedding
|
750 |
if dtype == torch.bfloat16:
|
751 |
return cos.bfloat16(), sin.bfloat16()
|
@@ -766,6 +560,14 @@ def apply_rope(x: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor) -> torch.T
|
|
766 |
|
767 |
|
768 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
769 |
import torch
|
770 |
# Copyright (c) 2022, Tri Dao.
|
771 |
# Adapted from https://github.com/NVIDIA/apex/blob/master/apex/contrib/layer_norm/layer_norm.py AND https://github.com/Dao-AILab/flash-attention/blob/7a983df74215e035e566e37125b0a71e3618f39d/flash_attn/ops/layer_norm.py#L16
|
@@ -1611,7 +1413,18 @@ class RMSNorm(torch.nn.Module):
|
|
1611 |
|
1612 |
|
1613 |
|
1614 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1615 |
|
1616 |
|
1617 |
# Copyright (c) 2023, Tri Dao.
|
|
|
1 |
+
"""
|
2 |
+
|
3 |
+
Based on the tinyllama implementation: https://github.com/jzhang38/TinyLlama
|
4 |
|
|
|
|
|
5 |
"""
|
6 |
+
|
7 |
+
|
8 |
import math, random
|
9 |
import numpy as np
|
10 |
from typing import Any, List, Optional, Tuple
|
11 |
+
from typing_extensions import Self
|
12 |
+
|
13 |
|
14 |
import torch
|
15 |
import torch.nn as nn
|
|
|
|
|
|
|
|
|
|
|
|
|
16 |
import torch.nn.functional as F
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
17 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
|
19 |
+
from lightning_utilities.core.imports import RequirementCache
|
20 |
+
FlashAttention2Available = RequirementCache("flash-attn>=2.0.0.post1")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
21 |
|
22 |
+
from flash_attn import flash_attn_func
|
23 |
+
from xformers.ops import SwiGLU
|
24 |
+
from einops import rearrange
|
|
|
|
|
25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
26 |
|
27 |
+
from transformers import PreTrainedModel
|
28 |
+
from .model_config import YingLongConfig
|
29 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
30 |
|
31 |
|
32 |
|
|
|
|
|
33 |
class Tokenizer(torch.nn.Module):
|
34 |
def __init__(self, config: YingLongConfig, *args,**kwargs) -> None:
|
35 |
super().__init__()
|
|
|
42 |
|
43 |
self.register_buffer('mask_token', torch.zeros(1000))
|
44 |
if self.config.haar_trans:
|
45 |
+
|
46 |
self.register_buffer('haar_transform',torch.Tensor(haarMatrix(self.config.patch_size,normalized = self.config.haar_trans_norm)))
|
47 |
|
48 |
|
|
|
91 |
|
92 |
else:
|
93 |
|
|
|
94 |
factor = 1
|
95 |
more_rows = future_token // self.patch_size + 1
|
96 |
prev_more_rows = prev_token // self.patch_size + 1
|
|
|
110 |
masks = [jj for jj in range(x_featured.shape[1])]
|
111 |
masks = masks[-more_rows:]
|
112 |
|
|
|
113 |
x_featured[:,-more_rows:] = self.mask0(self.mask_token[:len(masks)].unsqueeze(-1)).repeat(x_featured.shape[0],1,1)
|
114 |
x_featured[:,:prev_more_rows] = self.mask0(self.mask_token[:prev_more_rows].unsqueeze(-1)).repeat(x_featured.shape[0],1,1)
|
115 |
|
|
|
121 |
class model_tmp(PreTrainedModel):
|
122 |
config_class = YingLongConfig
|
123 |
base_model_prefix = "model"
|
|
|
|
|
|
|
|
|
|
|
|
|
124 |
|
125 |
+
|
126 |
+
|
127 |
+
def _init_weights(self, module: nn.Module) -> None:
|
128 |
+
if isinstance(module, nn.Embedding):
|
129 |
+
torch.nn.init.normal_(module.weight, mean=0.0, std=math.sqrt(2.0 / 5 / self.config.n_embd))
|
130 |
+
elif isinstance(module, nn.Linear):
|
131 |
+
torch.nn.init.normal_(module.weight, mean=0.0, std=math.sqrt(2.0 / 5 / self.config.n_embd))
|
132 |
+
if module.bias is not None:
|
133 |
+
torch.nn.init.zeros_(module.bias)
|
134 |
+
for name, p in module.named_parameters():
|
135 |
+
if (name == "proj.weight" and isinstance(module, LLaMAMLP)) or (name == "w3.weight" and isinstance(module, SwiGLU) or (name=="proj.weight" and isinstance(module, BidirectedlSelfAttention))):
|
136 |
+
nn.init.normal_(p, mean=0.0, std=1 / math.sqrt(self.config.n_embd) / self.config.n_layer)
|
137 |
+
|
138 |
+
|
139 |
+
|
140 |
+
|
141 |
+
|
142 |
+
|
143 |
+
|
144 |
class GPT(model_tmp):
|
145 |
def __init__(self, config: YingLongConfig, *args,**kwargs) -> None:
|
146 |
|
147 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
148 |
super().__init__(config)
|
149 |
|
150 |
self.config = config
|
|
|
153 |
|
154 |
|
155 |
if self.config._norm_class == "RMSNorm":
|
156 |
+
|
157 |
self.config.norm_class = RMSNorm
|
158 |
elif self.config._norm_class == "FusedRMSNorm":
|
|
|
159 |
self.config.norm_class = FusedRMSNorm
|
160 |
elif self.config._norm_class == 'BatchNorm':
|
|
|
161 |
self.config.norm_class = iBatchNorm
|
162 |
|
|
|
163 |
|
164 |
if self.config._mlp_class == "GptNeoxMLP":
|
|
|
165 |
self.config.mlp_class = GptNeoxMLP
|
166 |
elif self.config._mlp_class == "LLaMAMLP":
|
|
|
167 |
self.config.mlp_class = LLaMAMLP
|
168 |
|
169 |
+
|
170 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
171 |
|
172 |
self.tokenizer = Tokenizer(config)
|
173 |
|
174 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
175 |
self.lm_head = nn.Linear(config.n_embd, 99*self.patch_size)
|
176 |
+
|
|
|
|
|
177 |
|
|
|
178 |
self.quantitleLoss = quantitleLoss(99,patch_size = self.patch_size)
|
179 |
|
180 |
|
|
|
201 |
|
202 |
|
203 |
|
204 |
+
self.rope_cache = None
|
|
|
|
|
205 |
|
206 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
207 |
|
208 |
def forward(
|
209 |
self, idx: torch.Tensor,
|
|
|
|
|
|
|
210 |
future_token: int = 0,
|
211 |
prev_token: int = 0,
|
|
|
|
|
|
|
|
|
212 |
*args,**kwargs,
|
213 |
+
) -> torch.Tensor:
|
214 |
|
215 |
if future_token > 0:
|
216 |
more_rows = future_token // self.patch_size + 1
|
|
|
221 |
|
222 |
B, T = idx.size()
|
223 |
|
224 |
+
|
225 |
|
226 |
block_size = self.config.block_size
|
227 |
+
max_seq_length = T
|
|
|
|
|
228 |
|
|
|
|
|
|
|
|
|
229 |
assert max_seq_length <= block_size, f"Cannot attend to {max_seq_length}, block size is only {block_size}"
|
230 |
+
|
231 |
+
|
232 |
+
self.rope_cache = self.build_rope_cache(idx)
|
|
|
233 |
cos, sin = self.rope_cache
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
234 |
|
235 |
+
cos = cos[:max(T,1024)]
|
236 |
+
sin = sin[:max(T,1024)]
|
237 |
|
238 |
+
|
239 |
|
240 |
|
241 |
+
x,x_raw,masks,mean,std,_ = self.tokenizer(idx, future_token =future_token,prev_token = prev_token)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
242 |
|
243 |
|
244 |
|
|
|
245 |
if self.unet:
|
246 |
skips = []
|
247 |
|
248 |
|
249 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
250 |
|
251 |
+
for block_idx in range(len( self.transformer.h)):
|
252 |
|
|
|
253 |
|
254 |
+
block = self.transformer.h[block_idx]
|
|
|
255 |
|
256 |
+
if self.unet and block_idx >=len(self.transformer.h) //2:
|
257 |
+
x = self.unet_projection[block_idx - len(self.transformer.h) //2](torch.cat((skips.pop(),x),dim = -1))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
258 |
|
259 |
+
x = block(x, (cos, sin), max_seq_length)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
260 |
|
261 |
+
if self.unet and block_idx <len(self.transformer.h) //2:
|
262 |
+
skips.append(x)
|
263 |
+
x_delay = torch.cat((x[:,0,:].unsqueeze(1),x[:,:-1,:]),dim = 1)
|
264 |
+
x = self.unet_merge[block_idx](torch.cat((x_delay,x),dim = -1))
|
265 |
|
266 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
267 |
|
268 |
|
269 |
res = self.lm_head(x)
|
270 |
+
|
|
|
|
|
|
|
271 |
|
272 |
|
273 |
res = rearrange(res,'b c (l1 l2) -> b c l1 l2', l2 = 99)
|
|
|
275 |
|
276 |
|
277 |
if self.config.haar_trans_inv:
|
|
|
278 |
res = torch.einsum('bcal,ad->bcdl',res,self.tokenizer.haar_transform)
|
279 |
if self.config.haar_trans_norm == "backward":
|
280 |
res = res / np.sqrt(res.shape[-2])
|
|
|
282 |
res = res * np.sqrt(res.shape[-2])
|
283 |
|
284 |
|
285 |
+
|
286 |
+
|
287 |
+
|
288 |
res = res * (std.unsqueeze(-1) + 1e-4) + mean.unsqueeze(-1)
|
289 |
|
290 |
|
291 |
|
292 |
+
|
293 |
if future_token == 0:
|
294 |
+
return res[:,masks,:,:], x_raw[:,masks,:]
|
295 |
else:
|
296 |
+
return res[:,masks,:,:]
|
297 |
|
298 |
def generate(self,*args,**kwargs):
|
299 |
+
res = self.forward(*args,**kwargs)
|
|
|
|
|
|
|
300 |
res = rearrange(res, 'b l c d -> b (l c) d')
|
301 |
return res[:,:kwargs['future_token'],:]
|
302 |
|
303 |
|
304 |
+
|
305 |
@classmethod
|
306 |
def from_name(cls, name: str, **kwargs: Any) -> Self:
|
307 |
return cls(Config.from_name(name, **kwargs))
|
308 |
|
309 |
+
def build_rope_cache(self, idx: torch.Tensor) :
|
310 |
return build_rope_cache(
|
311 |
+
seq_len=self.config.block_size,
|
312 |
n_elem=int(self.config.rotary_percentage * self.config.head_size),
|
313 |
dtype=torch.bfloat16,
|
314 |
device=idx.device,
|
|
|
316 |
condense_ratio=self.config.condense_ratio,
|
317 |
)
|
318 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
319 |
|
320 |
class Block(nn.Module):
|
321 |
def __init__(self, config:YingLongConfig) -> None:
|
|
|
329 |
def forward(
|
330 |
self,
|
331 |
x: torch.Tensor,
|
332 |
+
rope: Optional[Tuple[torch.Tensor, torch.Tensor]],
|
333 |
max_seq_length: int,
|
334 |
mask: Optional[torch.Tensor] = None,
|
335 |
input_pos: Optional[torch.Tensor] = None,
|
336 |
+
) -> torch.Tensor:
|
|
|
337 |
|
338 |
n_1 = self.norm_1(x)
|
339 |
+
h = self.attn(n_1, rope, max_seq_length, mask, input_pos)
|
340 |
if self.config.parallel_residual:
|
341 |
n_2 = n_1 if self.config.shared_attention_norm else self.norm_2(x)
|
342 |
x = x + h + self.mlp(n_2)
|
|
|
349 |
|
350 |
x = x + h
|
351 |
x = x + self.mlp(self.norm_2(x))
|
352 |
+
return x
|
353 |
|
354 |
|
355 |
class BidirectedlSelfAttention(nn.Module):
|
356 |
def __init__(self, config:YingLongConfig) -> None:
|
357 |
super().__init__()
|
358 |
shape = (config.n_head + 2 * config.n_query_groups) * config.head_size
|
|
|
359 |
self.attn = nn.Linear(config.n_embd, shape, bias=config.bias)
|
|
|
360 |
self.proj = nn.Linear(config.n_embd, config.n_embd, bias=config.bias)
|
|
|
361 |
self.config = config
|
362 |
|
363 |
def forward(
|
364 |
self,
|
365 |
x: torch.Tensor,
|
366 |
+
rope: Tuple[torch.Tensor, torch.Tensor],
|
367 |
max_seq_length: int,
|
368 |
mask: Optional[torch.Tensor] = None,
|
369 |
input_pos: Optional[torch.Tensor] = None,
|
370 |
+
) -> torch.Tensor:
|
|
|
371 |
|
372 |
|
373 |
B, T, C = x.size() # batch size, sequence length, embedding dimensionality (n_embd)
|
|
|
378 |
q_per_kv = self.config.n_head // self.config.n_query_groups
|
379 |
total_qkv = q_per_kv + 2 # each group has 1+ queries, 1 key, and 1 value
|
380 |
qkv = qkv.view(B, T, self.config.n_query_groups, total_qkv, self.config.head_size) # (B, T, n_query_groups, total_qkv, hs)
|
381 |
+
|
382 |
|
383 |
# split batched computation into three
|
384 |
q, k, v = qkv.split((q_per_kv, 1, 1), dim=-2)
|
385 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
386 |
q = q.reshape(B, T, -1, self.config.head_size) # (B, T, nh_q, hs)
|
387 |
k = k.reshape(B, T, -1, self.config.head_size)
|
388 |
v = v.reshape(B, T, -1, self.config.head_size)
|
389 |
|
390 |
cos, sin = rope
|
391 |
|
|
|
|
|
392 |
q = apply_rotary_emb_func(q, cos, sin, False, True)
|
393 |
k = apply_rotary_emb_func(k, cos, sin, False, True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
394 |
|
|
|
395 |
|
396 |
y = self.scaled_dot_product_attention(q, k, v, mask=mask)
|
397 |
|
|
|
400 |
# output projection
|
401 |
y = self.proj(y)
|
402 |
|
403 |
+
return y
|
404 |
+
|
405 |
+
|
406 |
|
407 |
def scaled_dot_product_attention(
|
408 |
self, q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, mask: Optional[torch.Tensor] = None
|
|
|
422 |
k = k.transpose(1, 2)
|
423 |
v = v.transpose(1, 2)
|
424 |
if q.size() != k.size():
|
425 |
+
k = k.repeat_interleave(q.shape[1]//k.shape[1], dim=1)
|
426 |
+
v = v.repeat_interleave(q.shape[1]//v.shape[1], dim=1)
|
427 |
y = torch.nn.functional.scaled_dot_product_attention(
|
428 |
q, k, v, attn_mask=mask, dropout_p=0.0, scale=scale, is_causal=False
|
429 |
)
|
430 |
return y.transpose(1, 2)
|
431 |
|
432 |
|
433 |
+
|
434 |
+
|
435 |
+
|
436 |
+
|
437 |
+
class quantitleLoss(torch.nn.Module):
|
438 |
+
def __init__(self,
|
439 |
+
qSize = 99,
|
440 |
+
patch_size = 16,
|
441 |
+
*args,**kwargs):
|
442 |
+
|
443 |
+
super().__init__()
|
444 |
+
self.qSize = qSize
|
445 |
+
self.patch_size = patch_size
|
446 |
+
|
447 |
+
|
448 |
+
q = np.array([i+1 for i in range(self.qSize)])
|
449 |
+
q = q / (self.qSize + 1)
|
450 |
+
q = q.reshape((1,1,-1))
|
451 |
+
|
452 |
+
q_variance = q*(1-q)
|
453 |
+
|
454 |
+
self.register_buffer('q', torch.tensor(q))
|
455 |
+
self.register_buffer('q_variance', torch.tensor(q_variance))
|
456 |
+
|
457 |
+
|
458 |
+
def forward(self, input: torch.Tensor, target: torch.Tensor,rel_loss = False):
|
459 |
+
|
460 |
+
|
461 |
+
|
462 |
+
target = target.unsqueeze(-1)
|
463 |
+
input = input[:,:target.shape[1],:,:]
|
464 |
+
|
465 |
+
|
466 |
+
posPart = input - target
|
467 |
+
negPart = -posPart
|
468 |
+
|
469 |
+
raw_loss = torch.maximum(self.q * negPart, (1-self.q) * posPart)
|
470 |
+
|
471 |
+
target_absmean = torch.mean(target.abs(),dim = (1,2),keepdims = True)
|
472 |
+
raw_loss = raw_loss / torch.sqrt(self.q_variance) / (target_absmean + 1e-4)
|
473 |
+
|
474 |
+
return torch.mean(raw_loss)
|
475 |
+
|
476 |
+
|
477 |
+
def haarMatrix_unnormalized(n):
|
478 |
+
|
479 |
+
n = 2**np.ceil(np.log2(n))
|
480 |
+
if n > 2:
|
481 |
+
h = haarMatrix(n / 2)
|
482 |
+
else:
|
483 |
+
return np.array([[1, 1], [1, -1]])
|
484 |
+
h_n = np.kron(h, [1, 1])
|
485 |
+
h_i = np.kron(np.eye(len(h)), [1, -1])
|
486 |
+
h = np.vstack((h_n, h_i))
|
487 |
+
return h
|
488 |
+
|
489 |
+
def haarMatrix(n,normalized = 'ortho'):
|
490 |
+
h = haarMatrix_unnormalized(n)
|
491 |
+
scaler = np.diag(1/np.sqrt(np.diag(h@h.transpose())))
|
492 |
+
if normalized == 'ortho':
|
493 |
+
return scaler @ h
|
494 |
+
elif normalized == 'forward':
|
495 |
+
return scaler @ h/ np.sqrt(n)
|
496 |
+
|
497 |
+
else:
|
498 |
+
return scaler @ h * np.sqrt(n)
|
499 |
+
|
500 |
+
|
501 |
+
|
502 |
class GptNeoxMLP(nn.Module):
|
503 |
def __init__(self, config:YingLongConfig) -> None:
|
504 |
super().__init__()
|
|
|
514 |
class LLaMAMLP(nn.Module):
|
515 |
def __init__(self, config:YingLongConfig) -> None:
|
516 |
super().__init__()
|
517 |
+
|
|
|
|
|
518 |
self.swiglu = SwiGLU(config.n_embd,config.intermediate_size, bias=False, _pack_weights=False)
|
519 |
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
|
|
|
|
|
|
|
520 |
return self.swiglu(x)
|
521 |
|
522 |
|
523 |
def build_rope_cache(
|
524 |
seq_len: int, n_elem: int, dtype: torch.dtype, device: torch.device, base: int = 10000, condense_ratio: int = 1
|
525 |
+
) -> Tuple[torch.Tensor,torch.Tensor]:
|
526 |
"""Enhanced Transformer with Rotary Position Embedding.
|
527 |
|
528 |
Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/
|
|
|
540 |
|
541 |
cos, sin = torch.cos(idx_theta), torch.sin(idx_theta)
|
542 |
|
|
|
543 |
# added by peiyuan to ensure same data type with q, k, to use fused rotary embedding
|
544 |
if dtype == torch.bfloat16:
|
545 |
return cos.bfloat16(), sin.bfloat16()
|
|
|
560 |
|
561 |
|
562 |
|
563 |
+
|
564 |
+
|
565 |
+
|
566 |
+
######################################
|
567 |
+
#layernorm
|
568 |
+
######################################
|
569 |
+
|
570 |
+
|
571 |
import torch
|
572 |
# Copyright (c) 2022, Tri Dao.
|
573 |
# Adapted from https://github.com/NVIDIA/apex/blob/master/apex/contrib/layer_norm/layer_norm.py AND https://github.com/Dao-AILab/flash-attention/blob/7a983df74215e035e566e37125b0a71e3618f39d/flash_attn/ops/layer_norm.py#L16
|
|
|
1413 |
|
1414 |
|
1415 |
|
1416 |
+
|
1417 |
+
|
1418 |
+
|
1419 |
+
|
1420 |
+
######################################
|
1421 |
+
#rope_emb
|
1422 |
+
######################################
|
1423 |
+
|
1424 |
+
|
1425 |
+
|
1426 |
+
|
1427 |
+
|
1428 |
|
1429 |
|
1430 |
# Copyright (c) 2023, Tri Dao.
|