-
Notifications
You must be signed in to change notification settings - Fork 18
/
flash_decoding_llama.py
374 lines (305 loc) · 14.7 KB
/
flash_decoding_llama.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
# -*- coding:utf-8 -*-
from typing import List, Optional, Tuple, Union, Dict
from torch import nn
import math
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
from transformers.modeling_outputs import BaseModelOutputWithPast
from transformers.models.llama.modeling_llama import rotate_half, repeat_kv, LLAMA_INPUTS_DOCSTRING, apply_rotary_pos_emb
import torch
import transformers
from transformers.utils import add_start_docstrings_to_model_forward
from flash_attn import flash_attn_func, flash_attn_with_kvcache
def do_flash_attn(query_states, key_states, value_states, causal=True):
output, softmax_lse, _ = flash_attn_func(query_states.transpose(1, 2), key_states.transpose(1, 2),
value_states.transpose(1, 2), causal=causal, return_attn_probs=True)
return output.transpose(1, 2)
def do_flash_decoding(query_states, key_states, value_states, k_cache, v_cache, cache_seqlens):
output = flash_attn_with_kvcache(query_states.transpose(1, 2), k_cache, v_cache,
key_states.transpose(1, 2),
value_states.transpose(1, 2), cache_seqlens=cache_seqlens)
return output.transpose(1, 2)
def forward(
self,
hidden_states: torch.Tensor,
attention_mask: Optional[torch.LongTensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_value=None,
output_attentions: bool = False,
use_cache: bool = False,
**kwargs,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
bsz, q_len, _ = hidden_states.size()
query_states = self.q_proj(hidden_states)
key_states = self.k_proj(hidden_states)
value_states = self.v_proj(hidden_states)
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
kv_seq_len = key_states.shape[-2]
kv_seq_len += past_key_value["cache_seqlens"].item()
past_key_value["cache_seqlens"] += key_states.shape[-2]
q_seq_len = query_states.shape[-2]
has_kv_cache = q_seq_len != kv_seq_len
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
# update kv cache
key_cache = past_key_value[0][:, :, 0, :, :]
value_cache = past_key_value[0][:, :, 1, :, :]
if not has_kv_cache:
key_cache[:, kv_seq_len - key_states.shape[-2]:kv_seq_len, :, :] = key_states.transpose(1, 2)
value_cache[:, kv_seq_len - key_states.shape[-2]:kv_seq_len, :, :] = value_states.transpose(1, 2)
if not has_kv_cache:
attn_output = do_flash_attn(query_states, key_states, value_states)
else:
cache_seqlens = kv_seq_len-1
attn_output = do_flash_decoding(query_states, key_states, value_states, key_cache, value_cache,
cache_seqlens=cache_seqlens)
attn_output = attn_output.transpose(1, 2).contiguous()
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
attn_output = self.o_proj(attn_output)
return attn_output, None, past_key_value
def allocate_inference_cache(
max_batch_size,
max_seqlen,
nheads,
headdim,
layers,
dtype=torch.float16,
):
assert dtype in [torch.float16, torch.bfloat16, torch.float32]
kv_cache_shape = (max_batch_size, max_seqlen, 2, nheads, headdim)
allc_kv_cache = {i: {0: torch.empty(kv_cache_shape, device=layer.self_attn.k_proj.weight.device, dtype=dtype),
"cache_seqlens": torch.tensor([0], device=layer.self_attn.k_proj.weight.device).long()} for
i, layer in enumerate(layers)}
return allc_kv_cache
@add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING)
def LlamaModel_forward(
self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, BaseModelOutputWithPast]:
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
use_cache = use_cache if use_cache is not None else self.config.use_cache
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# retrieve input_ids and inputs_embeds
if input_ids is not None and inputs_embeds is not None:
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
elif input_ids is not None:
batch_size, seq_length = input_ids.shape[:2]
elif inputs_embeds is not None:
batch_size, seq_length = inputs_embeds.shape[:2]
else:
raise ValueError("You have to specify either input_ids or inputs_embeds")
if self.gradient_checkpointing and self.training:
if use_cache:
logger.warning_once(
"`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
)
use_cache = False
past_key_values_length = 0
if past_key_values:
input_ids = input_ids[:, -1].unsqueeze(-1)
position_ids = position_ids[:, -1].unsqueeze(-1) if position_ids is not None else None
if use_cache and past_key_values is None:
num_kv_heads = self.config.num_key_value_heads
num_attention_heads = self.config.num_attention_heads
head_dim = self.config.hidden_size // num_attention_heads
past_key_values = allocate_inference_cache(
batch_size,
MAX_CACHE_LEN,
num_kv_heads,
head_dim,
self.layers,
dtype=self.dtype,
)
if position_ids is None:
device = input_ids.device if input_ids is not None else inputs_embeds.device
position_ids = torch.arange(
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
)
position_ids = position_ids.unsqueeze(0)
if inputs_embeds is None:
inputs_embeds = self.embed_tokens(input_ids)
if self._use_flash_attention_2:
# 2d mask is passed through the layers
attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
elif self._use_sdpa and not output_attentions:
# output_attentions=True can not be supported when using SDPA, and we fall back on
# the manual implementation that requires a 4D causal mask in all cases.
attention_mask = _prepare_4d_causal_attention_mask_for_sdpa(
attention_mask,
(batch_size, seq_length),
inputs_embeds,
past_key_values_length,
)
else:
# 4d mask is passed through the layers
attention_mask = _prepare_4d_causal_attention_mask(
attention_mask, (batch_size, seq_length), inputs_embeds, past_key_values_length
)
# embed positions
hidden_states = inputs_embeds
# decoder layers
all_hidden_states = () if output_hidden_states else None
all_self_attns = () if output_attentions else None
next_decoder_cache = None
for i, decoder_layer in enumerate(self.layers):
if output_hidden_states:
all_hidden_states += (hidden_states,)
if self.gradient_checkpointing and self.training:
layer_outputs = self._gradient_checkpointing_func(
decoder_layer.__call__,
hidden_states,
attention_mask,
position_ids,
past_key_values,
output_attentions,
use_cache,
)
else:
layer_outputs = decoder_layer(
hidden_states,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_value=past_key_values[i],
output_attentions=output_attentions,
use_cache=use_cache,
)
hidden_states = layer_outputs[0]
if output_attentions:
all_self_attns += (layer_outputs[1],)
hidden_states = self.norm(hidden_states)
# add hidden states from the last decoder layer
if output_hidden_states:
all_hidden_states += (hidden_states,)
if not return_dict:
return tuple(v for v in [hidden_states, past_key_values, all_hidden_states, all_self_attns] if v is not None)
return BaseModelOutputWithPast(
last_hidden_state=hidden_states,
past_key_values=past_key_values,
hidden_states=all_hidden_states,
attentions=all_self_attns,
)
def causal_forward(self,
input_ids: torch.LongTensor = None,
attention_mask: Optional[torch.Tensor] = None,
position_ids: Optional[torch.LongTensor] = None,
past_key_values: Optional[List[torch.FloatTensor]] = None,
inputs_embeds: Optional[torch.FloatTensor] = None,
labels: Optional[torch.LongTensor] = None,
use_cache: Optional[bool] = None,
output_attentions: Optional[bool] = None,
output_hidden_states: Optional[bool] = None,
return_dict: Optional[bool] = None,
) -> Union[Tuple, CausalLMOutputWithPast]:
r"""
Args:
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
Returns:
Example:
```python
>>> from transformers import AutoTokenizer, LlamaForCausalLM
>>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
>>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")
>>> # Generate
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
```"""
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
output_hidden_states = (
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
)
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
# decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
outputs = self.model(
input_ids=input_ids,
attention_mask=attention_mask,
position_ids=position_ids,
past_key_values=past_key_values,
inputs_embeds=inputs_embeds,
use_cache=use_cache,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = outputs[0]
full_logits_length = 32000
if hidden_states.shape[-2] < full_logits_length:
logits = self.lm_head(hidden_states)
logits = logits.float()
loss = None
if labels is not None:
# Shift so that tokens < n predict n
shift_logits = logits[..., :-1, :].contiguous()
shift_labels = labels[..., 1:].contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
shift_logits = shift_logits.view(-1, self.config.vocab_size)
shift_labels = shift_labels.view(-1)
# Enable model parallelism
shift_labels = shift_labels.to(shift_logits.device)
loss = loss_fct(shift_logits, shift_labels)
else:
res = 0
div_len = full_logits_length // 2
if labels is None:
# only produce the last logits
logits = self.lm_head(hidden_states[..., -1:, :])
logits = logits.float()
# logits = logits.expand(-1, hidden_states.shape[-2], -1)
loss = None
else:
# calculate loss by chunk
shift_hidden_states = hidden_states[..., :-1, :]
shift_labels = labels[..., 1:].contiguous()
for i in range(0, shift_hidden_states.shape[-2], div_len):
st = i
ed = min(i + div_len, shift_hidden_states.shape[-2])
logits = self.lm_head(shift_hidden_states[..., st:ed, :])
logits = logits.float()
shift_logits = logits.contiguous()
# Flatten the tokens
loss_fct = CrossEntropyLoss()
shift_logits = shift_logits.view(-1, self.config.vocab_size)
shift_labels = shift_labels.view(-1)
# Enable model parallelism
shift_labels = shift_labels.to(shift_logits.device)
res = res + loss_fct(shift_logits, shift_labels[st:ed]) * (ed - st)
loss = res / (hidden_states.shape[-2] - 1)
logits = None
if not return_dict:
output = (logits,) + outputs[1:]
return (loss,) + output if loss is not None else output
return CausalLMOutputWithPast(
loss=loss,
logits=logits,
past_key_values=outputs.past_key_values,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
MAX_CACHE_LEN = 32 * 1024
MAX_NEW_TOKENS = 512
def replace_with_flashdecoding(max_prompt_length=None):
global MAX_CACHE_LEN
if max_prompt_length is not None:
MAX_CACHE_LEN = max_prompt_length + MAX_NEW_TOKENS
transformers.models.llama.modeling_llama.LlamaForCausalLM.forward = causal_forward
transformers.models.llama.modeling_llama.LlamaModel.forward = LlamaModel_forward
transformers.models.llama.modeling_llama.LlamaAttention.forward = forward
transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward = forward