Fix batch generation
Browse files- modeling_internlm.py +162 -57
modeling_internlm.py
CHANGED
@@ -1,5 +1,5 @@
|
|
1 |
# coding=utf-8
|
2 |
-
# Copyright
|
3 |
#
|
4 |
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
5 |
# and OPT implementations in this library. It has been modified from its
|
@@ -19,14 +19,14 @@
|
|
19 |
# limitations under the License.
|
20 |
""" PyTorch InternLM model."""
|
21 |
import math
|
|
|
|
|
22 |
from typing import List, Optional, Tuple, Union
|
23 |
-
import threading, queue
|
24 |
|
25 |
import torch
|
26 |
import torch.utils.checkpoint
|
27 |
from torch import nn
|
28 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
29 |
-
|
30 |
from transformers.activations import ACT2FN
|
31 |
from transformers.modeling_outputs import (
|
32 |
BaseModelOutputWithPast,
|
@@ -34,15 +34,19 @@ from transformers.modeling_outputs import (
|
|
34 |
SequenceClassifierOutputWithPast,
|
35 |
)
|
36 |
from transformers.modeling_utils import PreTrainedModel
|
37 |
-
from transformers.generation.streamers import BaseStreamer
|
38 |
from transformers.utils import (
|
39 |
add_start_docstrings,
|
40 |
add_start_docstrings_to_model_forward,
|
41 |
logging,
|
42 |
replace_return_docstrings,
|
43 |
)
|
44 |
-
from .configuration_internlm import InternLMConfig
|
45 |
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
logger = logging.get_logger(__name__)
|
48 |
|
@@ -82,7 +86,20 @@ def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int]
|
|
82 |
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
|
83 |
|
84 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
85 |
class InternLMRMSNorm(nn.Module):
|
|
|
|
|
86 |
def __init__(self, hidden_size, eps=1e-6):
|
87 |
"""
|
88 |
InternLMRMSNorm is equivalent to T5LayerNorm
|
@@ -103,10 +120,19 @@ class InternLMRMSNorm(nn.Module):
|
|
103 |
|
104 |
|
105 |
class InternLMRotaryEmbedding(torch.nn.Module):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
106 |
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
107 |
super().__init__()
|
108 |
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
109 |
-
self.register_buffer("inv_freq", inv_freq)
|
110 |
|
111 |
# Build here to make `torch.jit.trace` work.
|
112 |
self.max_seq_len_cached = max_position_embeddings
|
@@ -114,8 +140,8 @@ class InternLMRotaryEmbedding(torch.nn.Module):
|
|
114 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
115 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
116 |
emb = torch.cat((freqs, freqs), dim=-1)
|
117 |
-
self.register_buffer("cos_cached", emb.cos()
|
118 |
-
self.register_buffer("sin_cached", emb.sin()
|
119 |
|
120 |
def forward(self, x, seq_len=None):
|
121 |
# x: [bs, num_attention_heads, seq_len, head_size]
|
@@ -126,11 +152,71 @@ class InternLMRotaryEmbedding(torch.nn.Module):
|
|
126 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
127 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
128 |
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
129 |
-
self.register_buffer("cos_cached", emb.cos()
|
130 |
-
self.register_buffer("sin_cached", emb.sin()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
131 |
return (
|
132 |
-
self.cos_cached[
|
133 |
-
self.sin_cached[
|
134 |
)
|
135 |
|
136 |
|
@@ -140,15 +226,23 @@ def rotate_half(x):
|
|
140 |
x2 = x[..., x.shape[-1] // 2 :]
|
141 |
return torch.cat((-x2, x1), dim=-1)
|
142 |
|
143 |
-
|
144 |
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
145 |
-
|
146 |
-
|
147 |
-
|
148 |
-
|
149 |
-
|
150 |
-
|
151 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
152 |
return q_embed, k_embed
|
153 |
|
154 |
|
@@ -178,6 +272,8 @@ class InternLMAttention(nn.Module):
|
|
178 |
self.hidden_size = config.hidden_size
|
179 |
self.num_heads = config.num_attention_heads
|
180 |
self.head_dim = self.hidden_size // self.num_heads
|
|
|
|
|
181 |
self.max_position_embeddings = config.max_position_embeddings
|
182 |
|
183 |
if (self.head_dim * self.num_heads) != self.hidden_size:
|
@@ -186,10 +282,31 @@ class InternLMAttention(nn.Module):
|
|
186 |
f" and `num_heads`: {self.num_heads})."
|
187 |
)
|
188 |
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
|
189 |
-
self.k_proj = nn.Linear(self.hidden_size, self.
|
190 |
-
self.v_proj = nn.Linear(self.hidden_size, self.
|
191 |
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.bias)
|
192 |
-
self.rotary_emb =
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
193 |
|
194 |
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
|
195 |
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
@@ -206,15 +323,12 @@ class InternLMAttention(nn.Module):
|
|
206 |
bsz, q_len, _ = hidden_states.size()
|
207 |
|
208 |
query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
209 |
-
key_states =
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
-
|
214 |
-
|
215 |
-
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
216 |
-
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
217 |
-
# [bsz, nh, t, hd]
|
218 |
|
219 |
if past_key_value is not None:
|
220 |
# reuse k, v, self_attention
|
@@ -223,6 +337,13 @@ class InternLMAttention(nn.Module):
|
|
223 |
|
224 |
past_key_value = (key_states, value_states) if use_cache else None
|
225 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
226 |
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
227 |
|
228 |
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
@@ -332,11 +453,9 @@ INTERNLM_START_DOCSTRING = r"""
|
|
332 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
333 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
334 |
etc.)
|
335 |
-
|
336 |
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
337 |
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
338 |
and behavior.
|
339 |
-
|
340 |
Parameters:
|
341 |
config ([`InternLMConfig`]):
|
342 |
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
@@ -377,44 +496,34 @@ INTERNLM_INPUTS_DOCSTRING = r"""
|
|
377 |
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
378 |
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
379 |
it.
|
380 |
-
|
381 |
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
382 |
[`PreTrainedTokenizer.__call__`] for details.
|
383 |
-
|
384 |
[What are input IDs?](../glossary#input-ids)
|
385 |
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
386 |
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
387 |
-
|
388 |
- 1 for tokens that are **not masked**,
|
389 |
- 0 for tokens that are **masked**.
|
390 |
-
|
391 |
[What are attention masks?](../glossary#attention-mask)
|
392 |
-
|
393 |
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
394 |
[`PreTrainedTokenizer.__call__`] for details.
|
395 |
-
|
396 |
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
397 |
`past_key_values`).
|
398 |
-
|
399 |
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
400 |
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
401 |
information on the default strategy.
|
402 |
-
|
403 |
- 1 indicates the head is **not masked**,
|
404 |
- 0 indicates the head is **masked**.
|
405 |
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
406 |
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
407 |
config.n_positions - 1]`.
|
408 |
-
|
409 |
[What are position IDs?](../glossary#position-ids)
|
410 |
-
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or
|
|
|
411 |
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
412 |
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
|
413 |
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
|
414 |
-
|
415 |
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
416 |
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
417 |
-
|
418 |
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
419 |
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
420 |
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
@@ -443,7 +552,6 @@ INTERNLM_INPUTS_DOCSTRING = r"""
|
|
443 |
class InternLMModel(InternLMPreTrainedModel):
|
444 |
"""
|
445 |
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`InternLMDecoderLayer`]
|
446 |
-
|
447 |
Args:
|
448 |
config: InternLMConfig
|
449 |
"""
|
@@ -673,20 +781,14 @@ class InternLMForCausalLM(InternLMPreTrainedModel):
|
|
673 |
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
674 |
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
675 |
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
676 |
-
|
677 |
Returns:
|
678 |
-
|
679 |
Example:
|
680 |
-
|
681 |
```python
|
682 |
>>> from transformers import AutoTokenizer, InternLMForCausalLM
|
683 |
-
|
684 |
>>> model = InternLMForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
|
685 |
>>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
|
686 |
-
|
687 |
>>> prompt = "Hey, are you consciours? Can you talk to me?"
|
688 |
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
689 |
-
|
690 |
>>> # Generate
|
691 |
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
692 |
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
@@ -776,7 +878,7 @@ class InternLMForCausalLM(InternLMPreTrainedModel):
|
|
776 |
for layer_past in past_key_values:
|
777 |
reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
|
778 |
return reordered_past
|
779 |
-
|
780 |
def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = [], meta_instruction=""):
|
781 |
prompt = ""
|
782 |
if meta_instruction:
|
@@ -839,6 +941,11 @@ class InternLMForCausalLM(InternLMPreTrainedModel):
|
|
839 |
('你好,有什么可以帮助您的吗', [('你好', '你好,有什么可以帮助您的吗')])
|
840 |
('你好,有什么可以帮助您的吗?', [('你好', '你好,有什么可以帮助您的吗?')])
|
841 |
"""
|
|
|
|
|
|
|
|
|
|
|
842 |
|
843 |
response_queue = queue.Queue(maxsize=20)
|
844 |
|
@@ -891,7 +998,7 @@ class InternLMForCausalLM(InternLMPreTrainedModel):
|
|
891 |
producer.start()
|
892 |
while True:
|
893 |
res = response_queue.get()
|
894 |
-
if res is
|
895 |
return
|
896 |
yield res
|
897 |
|
@@ -901,10 +1008,8 @@ class InternLMForCausalLM(InternLMPreTrainedModel):
|
|
901 |
@add_start_docstrings(
|
902 |
"""
|
903 |
The InternLM Model transformer with a sequence classification head on top (linear layer).
|
904 |
-
|
905 |
[`InternLMForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
906 |
(e.g. GPT-2) do.
|
907 |
-
|
908 |
Since it does classification on the last token, it requires to know the position of the last token. If a
|
909 |
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
910 |
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|
|
|
1 |
# coding=utf-8
|
2 |
+
# Copyright (c) InternLM. All rights reserved.
|
3 |
#
|
4 |
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
5 |
# and OPT implementations in this library. It has been modified from its
|
|
|
19 |
# limitations under the License.
|
20 |
""" PyTorch InternLM model."""
|
21 |
import math
|
22 |
+
import queue
|
23 |
+
import threading
|
24 |
from typing import List, Optional, Tuple, Union
|
|
|
25 |
|
26 |
import torch
|
27 |
import torch.utils.checkpoint
|
28 |
from torch import nn
|
29 |
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
|
|
|
30 |
from transformers.activations import ACT2FN
|
31 |
from transformers.modeling_outputs import (
|
32 |
BaseModelOutputWithPast,
|
|
|
34 |
SequenceClassifierOutputWithPast,
|
35 |
)
|
36 |
from transformers.modeling_utils import PreTrainedModel
|
|
|
37 |
from transformers.utils import (
|
38 |
add_start_docstrings,
|
39 |
add_start_docstrings_to_model_forward,
|
40 |
logging,
|
41 |
replace_return_docstrings,
|
42 |
)
|
|
|
43 |
|
44 |
+
try:
|
45 |
+
from transformers.generation.streamers import BaseStreamer
|
46 |
+
except: # noqa # pylint: disable=bare-except
|
47 |
+
BaseStreamer = None
|
48 |
+
|
49 |
+
from .configuration_internlm import InternLMConfig
|
50 |
|
51 |
logger = logging.get_logger(__name__)
|
52 |
|
|
|
86 |
return inverted_mask.masked_fill(inverted_mask.to(torch.bool), torch.finfo(dtype).min)
|
87 |
|
88 |
|
89 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
90 |
+
"""
|
91 |
+
(batch, num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
92 |
+
"""
|
93 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
94 |
+
if n_rep == 1:
|
95 |
+
return hidden_states
|
96 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
97 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
98 |
+
|
99 |
+
|
100 |
class InternLMRMSNorm(nn.Module):
|
101 |
+
"""RMSNorm implemention."""
|
102 |
+
|
103 |
def __init__(self, hidden_size, eps=1e-6):
|
104 |
"""
|
105 |
InternLMRMSNorm is equivalent to T5LayerNorm
|
|
|
120 |
|
121 |
|
122 |
class InternLMRotaryEmbedding(torch.nn.Module):
|
123 |
+
"""Implement InternLM's rotary embedding.
|
124 |
+
|
125 |
+
Args:
|
126 |
+
dim (int): Characteristic dimension of each self-attentional head.
|
127 |
+
max_position_embeddings (int, optional): Model's training length. Defaults to 2048.
|
128 |
+
base (int, optional): The rotation position encodes the rotation Angle base number. Defaults to 10000.
|
129 |
+
device (Any, optional): Running device. Defaults to None.
|
130 |
+
"""
|
131 |
+
|
132 |
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
133 |
super().__init__()
|
134 |
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
135 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
136 |
|
137 |
# Build here to make `torch.jit.trace` work.
|
138 |
self.max_seq_len_cached = max_position_embeddings
|
|
|
140 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
141 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
142 |
emb = torch.cat((freqs, freqs), dim=-1)
|
143 |
+
self.register_buffer("cos_cached", emb.cos().to(torch.float32), persistent=False)
|
144 |
+
self.register_buffer("sin_cached", emb.sin().to(torch.float32), persistent=False)
|
145 |
|
146 |
def forward(self, x, seq_len=None):
|
147 |
# x: [bs, num_attention_heads, seq_len, head_size]
|
|
|
152 |
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
153 |
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
154 |
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
|
155 |
+
self.register_buffer("cos_cached", emb.cos(), persistent=False)
|
156 |
+
self.register_buffer("sin_cached", emb.sin(), persistent=False)
|
157 |
+
return (
|
158 |
+
self.cos_cached[:seq_len, ...].to(dtype=x.dtype),
|
159 |
+
self.sin_cached[:seq_len, ...].to(dtype=x.dtype),
|
160 |
+
)
|
161 |
+
|
162 |
+
|
163 |
+
class InternLMDynamicNTKScalingRotaryEmbedding(torch.nn.Module):
|
164 |
+
"""Implement InternLM's DyanmicNTK extrapolation method, thereby broadening the model support context to 16K.
|
165 |
+
|
166 |
+
Args:
|
167 |
+
dim (int): Characteristic dimension of each self-attentional head.
|
168 |
+
max_position_embeddings (int, optional): Model's training length. Defaults to 2048.
|
169 |
+
base (int, optional): The rotation position encodes the rotation Angle base number. Defaults to 10000.
|
170 |
+
device (Any, optional): Running device. Defaults to None.
|
171 |
+
scaling_factor (float, optional): NTK method extrapolation coefficient. Defaults to 1.0.
|
172 |
+
"""
|
173 |
+
|
174 |
+
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None, scaling_factor=1.0):
|
175 |
+
super().__init__()
|
176 |
+
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
|
177 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
178 |
+
self.dim = dim
|
179 |
+
self.base = base
|
180 |
+
self.scaling_factor = scaling_factor
|
181 |
+
|
182 |
+
# Build here to make `torch.jit.trace` work.
|
183 |
+
self.max_position_embeddings = max_position_embeddings
|
184 |
+
self.max_seq_len_cached = max_position_embeddings
|
185 |
+
t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
|
186 |
+
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
|
187 |
+
# Different from paper, but it uses a different permutation in order to obtain the same calculation
|
188 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
189 |
+
self.register_buffer("cos_cached", emb.cos(), persistent=False)
|
190 |
+
self.register_buffer("sin_cached", emb.sin(), persistent=False)
|
191 |
+
|
192 |
+
def _update_cached(self, x, seq_len=None):
|
193 |
+
self.max_seq_len_cached = max(seq_len, self.max_position_embeddings)
|
194 |
+
if seq_len > self.max_position_embeddings:
|
195 |
+
base = self.base * (
|
196 |
+
(self.scaling_factor * seq_len / self.max_position_embeddings) - (self.scaling_factor - 1)
|
197 |
+
) ** (self.dim / (self.dim - 2))
|
198 |
+
inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2).float().to(x.device) / self.dim))
|
199 |
+
else:
|
200 |
+
inv_freq = self.inv_freq
|
201 |
+
t = torch.arange(self.max_seq_len_cached, device=inv_freq.device, dtype=inv_freq.dtype)
|
202 |
+
freqs = torch.einsum("i,j->ij", t, inv_freq)
|
203 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
204 |
+
self.register_buffer("cos_cached", emb.cos(), persistent=False)
|
205 |
+
self.register_buffer("sin_cached", emb.sin(), persistent=False)
|
206 |
+
|
207 |
+
def forward(self, x, seq_len=None):
|
208 |
+
# x: [bs, num_attention_heads, seq_len, head_size]
|
209 |
+
# This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
|
210 |
+
if seq_len <= self.max_position_embeddings:
|
211 |
+
# Reset the tables if the sequence length has changed,
|
212 |
+
if self.max_seq_len_cached > self.max_position_embeddings:
|
213 |
+
self._update_cached(x, seq_len)
|
214 |
+
else:
|
215 |
+
self._update_cached(x, seq_len)
|
216 |
+
|
217 |
return (
|
218 |
+
self.cos_cached[:seq_len, ...].to(dtype=x.dtype),
|
219 |
+
self.sin_cached[:seq_len, ...].to(dtype=x.dtype),
|
220 |
)
|
221 |
|
222 |
|
|
|
226 |
x2 = x[..., x.shape[-1] // 2 :]
|
227 |
return torch.cat((-x2, x1), dim=-1)
|
228 |
|
|
|
229 |
def apply_rotary_pos_emb(q, k, cos, sin, position_ids):
|
230 |
+
if position_ids.size(1) == 1:
|
231 |
+
q_cos = cos[position_ids].unsqueeze(1).expand(q.shape)
|
232 |
+
q_sin = sin[position_ids].unsqueeze(1).expand(q.shape)
|
233 |
+
q_embed = (q * q_cos) + (rotate_half(q) * q_sin)
|
234 |
+
|
235 |
+
position_ids = position_ids.flatten() + 1
|
236 |
+
max_length = max(position_ids)
|
237 |
+
position_ids = torch.stack([torch.cat([torch.ones(max_length - w, dtype=torch.long), torch.arange(w)]) for w in position_ids])
|
238 |
+
k_cos = cos[position_ids].unsqueeze(1).expand(k.shape)
|
239 |
+
k_sin = sin[position_ids].unsqueeze(1).expand(k.shape)
|
240 |
+
k_embed = (k * k_cos) + (rotate_half(k) * k_sin)
|
241 |
+
else:
|
242 |
+
cos = cos[position_ids].unsqueeze(1).expand(q.shape)
|
243 |
+
sin = sin[position_ids].unsqueeze(1).expand(q.shape)
|
244 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
245 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
246 |
return q_embed, k_embed
|
247 |
|
248 |
|
|
|
272 |
self.hidden_size = config.hidden_size
|
273 |
self.num_heads = config.num_attention_heads
|
274 |
self.head_dim = self.hidden_size // self.num_heads
|
275 |
+
self.num_key_value_heads = config.num_key_value_heads
|
276 |
+
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
277 |
self.max_position_embeddings = config.max_position_embeddings
|
278 |
|
279 |
if (self.head_dim * self.num_heads) != self.hidden_size:
|
|
|
282 |
f" and `num_heads`: {self.num_heads})."
|
283 |
)
|
284 |
self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.bias)
|
285 |
+
self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.bias)
|
286 |
+
self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.bias)
|
287 |
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.bias)
|
288 |
+
self.rotary_emb = self._init_rope()
|
289 |
+
|
290 |
+
def _init_rope(self):
|
291 |
+
if self.config.rope_scaling is None:
|
292 |
+
self.rotary_emb = InternLMRotaryEmbedding(
|
293 |
+
self.head_dim,
|
294 |
+
max_position_embeddings=self.max_position_embeddings,
|
295 |
+
base=self.config.rope_theta,
|
296 |
+
)
|
297 |
+
else:
|
298 |
+
scaling_type = self.config.rope_scaling["type"]
|
299 |
+
scaling_factor = self.config.rope_scaling["factor"]
|
300 |
+
if scaling_type == "dynamic":
|
301 |
+
self.rotary_emb = InternLMDynamicNTKScalingRotaryEmbedding(
|
302 |
+
self.head_dim,
|
303 |
+
max_position_embeddings=self.max_position_embeddings,
|
304 |
+
base=self.config.rope_theta,
|
305 |
+
scaling_factor=scaling_factor,
|
306 |
+
)
|
307 |
+
else:
|
308 |
+
raise ValueError("Currently we only support rotary embedding's type being 'dynamic'.")
|
309 |
+
return self.rotary_emb
|
310 |
|
311 |
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
|
312 |
return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
|
|
|
323 |
bsz, q_len, _ = hidden_states.size()
|
324 |
|
325 |
query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
326 |
+
key_states = (
|
327 |
+
self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
328 |
+
)
|
329 |
+
value_states = (
|
330 |
+
self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
331 |
+
)
|
|
|
|
|
|
|
332 |
|
333 |
if past_key_value is not None:
|
334 |
# reuse k, v, self_attention
|
|
|
337 |
|
338 |
past_key_value = (key_states, value_states) if use_cache else None
|
339 |
|
340 |
+
kv_seq_len = key_states.shape[-2]
|
341 |
+
cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
|
342 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
343 |
+
|
344 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
345 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
346 |
+
|
347 |
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
348 |
|
349 |
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
|
|
453 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
454 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
455 |
etc.)
|
|
|
456 |
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
|
457 |
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
458 |
and behavior.
|
|
|
459 |
Parameters:
|
460 |
config ([`InternLMConfig`]):
|
461 |
Model configuration class with all the parameters of the model. Initializing with a config file does not
|
|
|
496 |
input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
|
497 |
Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
|
498 |
it.
|
|
|
499 |
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
500 |
[`PreTrainedTokenizer.__call__`] for details.
|
|
|
501 |
[What are input IDs?](../glossary#input-ids)
|
502 |
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
503 |
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
|
|
504 |
- 1 for tokens that are **not masked**,
|
505 |
- 0 for tokens that are **masked**.
|
|
|
506 |
[What are attention masks?](../glossary#attention-mask)
|
|
|
507 |
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
508 |
[`PreTrainedTokenizer.__call__`] for details.
|
|
|
509 |
If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
|
510 |
`past_key_values`).
|
|
|
511 |
If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
|
512 |
and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
|
513 |
information on the default strategy.
|
|
|
514 |
- 1 indicates the head is **not masked**,
|
515 |
- 0 indicates the head is **masked**.
|
516 |
position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
517 |
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
|
518 |
config.n_positions - 1]`.
|
|
|
519 |
[What are position IDs?](../glossary#position-ids)
|
520 |
+
past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or
|
521 |
+
when `config.use_cache=True`):
|
522 |
Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
|
523 |
`(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
|
524 |
`(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
|
|
|
525 |
Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
|
526 |
blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
|
|
|
527 |
If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
|
528 |
don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
|
529 |
`decoder_input_ids` of shape `(batch_size, sequence_length)`.
|
|
|
552 |
class InternLMModel(InternLMPreTrainedModel):
|
553 |
"""
|
554 |
Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`InternLMDecoderLayer`]
|
|
|
555 |
Args:
|
556 |
config: InternLMConfig
|
557 |
"""
|
|
|
781 |
Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
|
782 |
config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
|
783 |
(masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
|
|
|
784 |
Returns:
|
|
|
785 |
Example:
|
|
|
786 |
```python
|
787 |
>>> from transformers import AutoTokenizer, InternLMForCausalLM
|
|
|
788 |
>>> model = InternLMForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS)
|
789 |
>>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER)
|
|
|
790 |
>>> prompt = "Hey, are you consciours? Can you talk to me?"
|
791 |
>>> inputs = tokenizer(prompt, return_tensors="pt")
|
|
|
792 |
>>> # Generate
|
793 |
>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
|
794 |
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
|
|
|
878 |
for layer_past in past_key_values:
|
879 |
reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
|
880 |
return reordered_past
|
881 |
+
|
882 |
def build_inputs(self, tokenizer, query: str, history: List[Tuple[str, str]] = [], meta_instruction=""):
|
883 |
prompt = ""
|
884 |
if meta_instruction:
|
|
|
941 |
('你好,有什么可以帮助您的吗', [('你好', '你好,有什么可以帮助您的吗')])
|
942 |
('你好,有什么可以帮助您的吗?', [('你好', '你好,有什么可以帮助您的吗?')])
|
943 |
"""
|
944 |
+
if BaseStreamer is None:
|
945 |
+
raise ModuleNotFoundError(
|
946 |
+
"The version of `transformers` is too low. Please make sure "
|
947 |
+
"that you have installed `transformers>=4.28.0`."
|
948 |
+
)
|
949 |
|
950 |
response_queue = queue.Queue(maxsize=20)
|
951 |
|
|
|
998 |
producer.start()
|
999 |
while True:
|
1000 |
res = response_queue.get()
|
1001 |
+
if res is None:
|
1002 |
return
|
1003 |
yield res
|
1004 |
|
|
|
1008 |
@add_start_docstrings(
|
1009 |
"""
|
1010 |
The InternLM Model transformer with a sequence classification head on top (linear layer).
|
|
|
1011 |
[`InternLMForSequenceClassification`] uses the last token in order to do the classification, as other causal models
|
1012 |
(e.g. GPT-2) do.
|
|
|
1013 |
Since it does classification on the last token, it requires to know the position of the last token. If a
|
1014 |
`pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
|
1015 |
no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
|