GradientGuru
commited on
Commit
•
85278de
1
Parent(s):
f5f47be
fix alibi problem and support attention_mask
Browse files- modeling_baichuan.py +72 -26
modeling_baichuan.py
CHANGED
@@ -16,6 +16,7 @@ from .configuration_baichuan import BaichuanConfig
|
|
16 |
|
17 |
logger = logging.get_logger(__name__)
|
18 |
|
|
|
19 |
def _get_interleave(n):
|
20 |
def _get_interleave_power_of_2(n):
|
21 |
start = (2 ** (-2 ** -(math.log2(n) - 3)))
|
@@ -44,6 +45,16 @@ def _gen_alibi_mask(n_head, max_pos):
|
|
44 |
alibi_mask = alibi_mask.unsqueeze(0) + alibi
|
45 |
return alibi_mask
|
46 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
47 |
|
48 |
class RMSNorm(torch.nn.Module):
|
49 |
def __init__(self, hidden_size, epsilon=1e-6):
|
@@ -80,7 +91,6 @@ class MLP(torch.nn.Module):
|
|
80 |
|
81 |
|
82 |
class BaichuanAttention(torch.nn.Module):
|
83 |
-
|
84 |
def __init__(self, config: BaichuanConfig):
|
85 |
super().__init__()
|
86 |
self.config = config
|
@@ -130,12 +140,16 @@ class BaichuanAttention(torch.nn.Module):
|
|
130 |
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
131 |
|
132 |
if attention_mask is not None:
|
133 |
-
if
|
134 |
-
|
135 |
-
|
|
|
|
|
|
|
136 |
attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
|
137 |
|
138 |
attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
|
|
|
139 |
attn_output = torch.matmul(attn_weights, value_states)
|
140 |
|
141 |
attn_output = attn_output.transpose(1, 2)
|
@@ -239,23 +253,32 @@ class BaichuanModel(BaichuanPreTrainedModel):
|
|
239 |
|
240 |
def get_input_embeddings(self):
|
241 |
return self.embed_tokens
|
242 |
-
|
243 |
def set_input_embeddings(self, value):
|
244 |
self.embed_tokens = value
|
245 |
-
|
246 |
def get_alibi_mask(self, tensor, seq_length_with_past):
|
247 |
-
if self.
|
248 |
-
|
249 |
-
|
250 |
-
|
251 |
-
|
252 |
-
|
253 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
254 |
return mask
|
255 |
|
256 |
def forward(
|
257 |
self,
|
258 |
input_ids: torch.LongTensor = None,
|
|
|
259 |
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
260 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
261 |
use_cache: Optional[bool] = False,
|
@@ -283,8 +306,23 @@ class BaichuanModel(BaichuanPreTrainedModel):
|
|
283 |
if inputs_embeds is None:
|
284 |
inputs_embeds = self.embed_tokens(input_ids)
|
285 |
|
286 |
-
|
287 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
288 |
|
289 |
hidden_states = inputs_embeds
|
290 |
|
@@ -353,7 +391,7 @@ class BaichuanModel(BaichuanPreTrainedModel):
|
|
353 |
hidden_states=all_hidden_states,
|
354 |
attentions=all_self_attns,
|
355 |
)
|
356 |
-
|
357 |
|
358 |
class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
359 |
def __init__(self, config):
|
@@ -363,7 +401,7 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
|
363 |
|
364 |
# Initialize weights and apply final processing
|
365 |
self.post_init()
|
366 |
-
|
367 |
def get_input_embeddings(self):
|
368 |
return self.model.embed_tokens
|
369 |
|
@@ -381,10 +419,11 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
|
381 |
|
382 |
def get_decoder(self):
|
383 |
return self.model
|
384 |
-
|
385 |
def forward(
|
386 |
self,
|
387 |
input_ids: torch.LongTensor = None,
|
|
|
388 |
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
389 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
390 |
labels: Optional[torch.LongTensor] = None,
|
@@ -399,13 +438,14 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
|
399 |
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
400 |
outputs = self.model(
|
401 |
input_ids=input_ids,
|
|
|
402 |
past_key_values=past_key_values,
|
403 |
inputs_embeds=inputs_embeds,
|
404 |
use_cache=use_cache,
|
405 |
output_attentions=output_attentions,
|
406 |
output_hidden_states=output_hidden_states,
|
407 |
return_dict=return_dict,
|
408 |
-
)
|
409 |
|
410 |
hidden_states = outputs[0]
|
411 |
logits = self.lm_head(hidden_states)
|
@@ -436,8 +476,13 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
|
436 |
)
|
437 |
|
438 |
def prepare_inputs_for_generation(
|
439 |
-
self,
|
440 |
-
|
|
|
|
|
|
|
|
|
|
|
441 |
if past_key_values:
|
442 |
input_ids = input_ids[:, -1:]
|
443 |
|
@@ -448,11 +493,12 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
|
448 |
model_inputs = {"input_ids": input_ids}
|
449 |
|
450 |
model_inputs.update(
|
451 |
-
{
|
452 |
"past_key_values": past_key_values,
|
453 |
"use_cache": kwargs.get("use_cache"),
|
454 |
-
|
455 |
-
|
|
|
456 |
return model_inputs
|
457 |
|
458 |
@staticmethod
|
@@ -470,7 +516,7 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
|
470 |
raise ImportError(
|
471 |
f"Needs QLinear to run quantize."
|
472 |
)
|
473 |
-
|
474 |
for layer in self.model.layers:
|
475 |
layer.self_attn.W_pack = QLinear(
|
476 |
bits=bits,
|
@@ -497,7 +543,7 @@ class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
|
497 |
weight=layer.mlp.up_proj.weight,
|
498 |
bias = None,
|
499 |
)
|
500 |
-
return self
|
501 |
|
502 |
def _build_chat_input(self, tokenizer, messages: List[dict], max_new_tokens: int=0):
|
503 |
max_new_tokens = max_new_tokens or self.generation_config.max_new_tokens
|
|
|
16 |
|
17 |
logger = logging.get_logger(__name__)
|
18 |
|
19 |
+
|
20 |
def _get_interleave(n):
|
21 |
def _get_interleave_power_of_2(n):
|
22 |
start = (2 ** (-2 ** -(math.log2(n) - 3)))
|
|
|
45 |
alibi_mask = alibi_mask.unsqueeze(0) + alibi
|
46 |
return alibi_mask
|
47 |
|
48 |
+
def _buffered_future_mask(tensor, maxpos, alibi, attn_heads):
|
49 |
+
"""for training only"""
|
50 |
+
dim = tensor.size(1)
|
51 |
+
_future_mask = torch.triu(
|
52 |
+
_fill_with_neg_inf(torch.zeros([maxpos, maxpos])), 1
|
53 |
+
)
|
54 |
+
_future_mask = _future_mask.unsqueeze(0) + alibi
|
55 |
+
_future_mask = _future_mask.to(tensor)
|
56 |
+
return _future_mask[:tensor.shape[0] * attn_heads, :maxpos, :maxpos]
|
57 |
+
|
58 |
|
59 |
class RMSNorm(torch.nn.Module):
|
60 |
def __init__(self, hidden_size, epsilon=1e-6):
|
|
|
91 |
|
92 |
|
93 |
class BaichuanAttention(torch.nn.Module):
|
|
|
94 |
def __init__(self, config: BaichuanConfig):
|
95 |
super().__init__()
|
96 |
self.config = config
|
|
|
140 |
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
141 |
|
142 |
if attention_mask is not None:
|
143 |
+
if q_len == 1: # inference with cache
|
144 |
+
if len(attention_mask.size()) == 4:
|
145 |
+
attention_mask = attention_mask[:, :, -1:, :]
|
146 |
+
else:
|
147 |
+
attention_mask = attention_mask[:, -1:, :]
|
148 |
+
attn_weights = attn_weights + attention_mask
|
149 |
attn_weights = torch.max(attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min))
|
150 |
|
151 |
attn_weights = torch.nn.functional.softmax(attn_weights, dim=-1)
|
152 |
+
|
153 |
attn_output = torch.matmul(attn_weights, value_states)
|
154 |
|
155 |
attn_output = attn_output.transpose(1, 2)
|
|
|
253 |
|
254 |
def get_input_embeddings(self):
|
255 |
return self.embed_tokens
|
256 |
+
|
257 |
def set_input_embeddings(self, value):
|
258 |
self.embed_tokens = value
|
259 |
+
|
260 |
def get_alibi_mask(self, tensor, seq_length_with_past):
|
261 |
+
if self.training:
|
262 |
+
slopes = torch.Tensor(_get_interleave(self.n_head))
|
263 |
+
alibi = slopes.unsqueeze(1).unsqueeze(1) * torch.arange(seq_length_with_past).unsqueeze(0).unsqueeze(0).expand(
|
264 |
+
self.n_head,
|
265 |
+
-1, -1)
|
266 |
+
alibi = alibi.view(self.n_head, 1, seq_length_with_past)
|
267 |
+
mask = _buffered_future_mask(tensor, seq_length_with_past, alibi, self.n_head)
|
268 |
+
else:
|
269 |
+
if self.first_run:
|
270 |
+
self.first_run = False
|
271 |
+
self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
|
272 |
+
if seq_length_with_past > self.max_cache_pos:
|
273 |
+
self.max_cache_pos = seq_length_with_past
|
274 |
+
self.register_buffer("future_mask", _gen_alibi_mask(self.n_head, self.max_cache_pos).to(tensor), persistent=False)
|
275 |
+
mask = self.future_mask[:self.n_head, :seq_length_with_past, :seq_length_with_past]
|
276 |
return mask
|
277 |
|
278 |
def forward(
|
279 |
self,
|
280 |
input_ids: torch.LongTensor = None,
|
281 |
+
attention_mask: Optional[torch.Tensor] = None,
|
282 |
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
283 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
284 |
use_cache: Optional[bool] = False,
|
|
|
306 |
if inputs_embeds is None:
|
307 |
inputs_embeds = self.embed_tokens(input_ids)
|
308 |
|
309 |
+
|
310 |
+
alibi_mask = self.get_alibi_mask(inputs_embeds, seq_length_with_past)
|
311 |
+
if attention_mask is not None:
|
312 |
+
if len(attention_mask.shape) == 2:
|
313 |
+
expanded_mask = attention_mask.to(alibi_mask.dtype)
|
314 |
+
expanded_mask = torch.tril(torch.gt(expanded_mask[:, :, None] * expanded_mask[:, None, :], 0)
|
315 |
+
) * torch.eq(expanded_mask[:, :, None] - expanded_mask[:, None, :], 0)
|
316 |
+
else:
|
317 |
+
expanded_mask = attention_mask
|
318 |
+
bsz = inputs_embeds.size(0)
|
319 |
+
src_len, tgt_len = alibi_mask.size()[-2:]
|
320 |
+
expanded_mask = expanded_mask.unsqueeze(1).expand(bsz, 1, src_len, tgt_len).to(alibi_mask.dtype)
|
321 |
+
inverted_mask = 1.0 - expanded_mask
|
322 |
+
inverted_mask = inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(alibi_mask.dtype).min)
|
323 |
+
attention_mask = inverted_mask + alibi_mask.unsqueeze(0)
|
324 |
+
else:
|
325 |
+
attention_mask = alibi_mask
|
326 |
|
327 |
hidden_states = inputs_embeds
|
328 |
|
|
|
391 |
hidden_states=all_hidden_states,
|
392 |
attentions=all_self_attns,
|
393 |
)
|
394 |
+
|
395 |
|
396 |
class BaichuanForCausalLM(BaichuanPreTrainedModel):
|
397 |
def __init__(self, config):
|
|
|
401 |
|
402 |
# Initialize weights and apply final processing
|
403 |
self.post_init()
|
404 |
+
|
405 |
def get_input_embeddings(self):
|
406 |
return self.model.embed_tokens
|
407 |
|
|
|
419 |
|
420 |
def get_decoder(self):
|
421 |
return self.model
|
422 |
+
|
423 |
def forward(
|
424 |
self,
|
425 |
input_ids: torch.LongTensor = None,
|
426 |
+
attention_mask: Optional[torch.Tensor] = None,
|
427 |
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
428 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
429 |
labels: Optional[torch.LongTensor] = None,
|
|
|
438 |
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
|
439 |
outputs = self.model(
|
440 |
input_ids=input_ids,
|
441 |
+
attention_mask=attention_mask,
|
442 |
past_key_values=past_key_values,
|
443 |
inputs_embeds=inputs_embeds,
|
444 |
use_cache=use_cache,
|
445 |
output_attentions=output_attentions,
|
446 |
output_hidden_states=output_hidden_states,
|
447 |
return_dict=return_dict,
|
448 |
+
)
|
449 |
|
450 |
hidden_states = outputs[0]
|
451 |
logits = self.lm_head(hidden_states)
|
|
|
476 |
)
|
477 |
|
478 |
def prepare_inputs_for_generation(
|
479 |
+
self,
|
480 |
+
input_ids: torch.LongTensor,
|
481 |
+
past_key_values: Optional[torch.Tensor] = None,
|
482 |
+
attention_mask: Optional[torch.Tensor] = None,
|
483 |
+
inputs_embeds: Optional[torch.Tensor] = None,
|
484 |
+
**kwargs
|
485 |
+
):
|
486 |
if past_key_values:
|
487 |
input_ids = input_ids[:, -1:]
|
488 |
|
|
|
493 |
model_inputs = {"input_ids": input_ids}
|
494 |
|
495 |
model_inputs.update(
|
496 |
+
{
|
497 |
"past_key_values": past_key_values,
|
498 |
"use_cache": kwargs.get("use_cache"),
|
499 |
+
"attention_mask": attention_mask
|
500 |
+
}
|
501 |
+
)
|
502 |
return model_inputs
|
503 |
|
504 |
@staticmethod
|
|
|
516 |
raise ImportError(
|
517 |
f"Needs QLinear to run quantize."
|
518 |
)
|
519 |
+
|
520 |
for layer in self.model.layers:
|
521 |
layer.self_attn.W_pack = QLinear(
|
522 |
bits=bits,
|
|
|
543 |
weight=layer.mlp.up_proj.weight,
|
544 |
bias = None,
|
545 |
)
|
546 |
+
return self
|
547 |
|
548 |
def _build_chat_input(self, tokenizer, messages: List[dict], max_new_tokens: int=0):
|
549 |
max_new_tokens = max_new_tokens or self.generation_config.max_new_tokens
|