KaleiNeely
commited on
Commit
•
14f6f8d
1
Parent(s):
af0c7e3
Upload 3 files
Browse files- configuration_rwkv5.py +22 -27
- modeling_rwkv5.py +160 -218
- tokenization_rwkv_world.py +91 -91
configuration_rwkv5.py
CHANGED
@@ -21,46 +21,44 @@ from transformers.utils import logging
|
|
21 |
|
22 |
logger = logging.get_logger(__name__)
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
}
|
27 |
|
28 |
|
29 |
class Rwkv5Config(PretrainedConfig):
|
30 |
"""
|
31 |
-
This is the configuration class to store the configuration of a [`
|
32 |
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
33 |
defaults will yield a similar configuration to that of the RWVK-4
|
34 |
-
[RWKV/rwkv-
|
35 |
|
36 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
37 |
documentation from [`PretrainedConfig`] for more information.
|
38 |
|
39 |
|
40 |
Args:
|
41 |
-
vocab_size (`int`, *optional*, defaults to
|
42 |
-
Vocabulary size of the
|
43 |
-
`inputs_ids` passed when calling [`
|
44 |
-
|
45 |
-
The maximum sequence length that this model can be be used with in a single forward (using it in RNN mode
|
46 |
-
lets use any sequence length).
|
47 |
-
hidden_size (`int`, *optional*, defaults to 4096):
|
48 |
Dimensionality of the embeddings and hidden states.
|
49 |
-
num_hidden_layers (`int`, *optional*, defaults to
|
50 |
Number of hidden layers in the model.
|
51 |
attention_hidden_size (`int`, *optional*):
|
52 |
Dimensionality of the attention hidden states. Will default to `hidden_size` if unset.
|
|
|
|
|
|
|
53 |
intermediate_size (`int`, *optional*):
|
54 |
Dimensionality of the inner feed-forward layers. Will default to 4 times `hidden_size` if unset.
|
55 |
-
|
56 |
The epsilon to use in the layer normalization layers.
|
57 |
bos_token_id (`int`, *optional*, defaults to 0):
|
58 |
-
The id of the beginning of sentence token in the vocabulary. Defaults to 0 as
|
59 |
as GPTNeoX.
|
60 |
eos_token_id (`int`, *optional*, defaults to 0):
|
61 |
-
The id of the end of sentence token in the vocabulary. Defaults to 0 as
|
62 |
GPTNeoX.
|
63 |
-
rescale_every (`int`, *optional*,
|
64 |
At inference, the hidden states (and weights of the correponding output layers) are divided by 2 every
|
65 |
`rescale_every` layer. If set to 0 or a negative number, no rescale is done.
|
66 |
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
@@ -72,28 +70,27 @@ class Rwkv5Config(PretrainedConfig):
|
|
72 |
Example:
|
73 |
|
74 |
```python
|
75 |
-
>>> from transformers import
|
76 |
|
77 |
-
>>> # Initializing a
|
78 |
-
>>> configuration =
|
79 |
|
80 |
>>> # Initializing a model (with random weights) from the configuration
|
81 |
-
>>> model =
|
82 |
|
83 |
>>> # Accessing the model configuration
|
84 |
>>> configuration = model.config
|
85 |
```"""
|
86 |
|
87 |
model_type = "rwkv5"
|
88 |
-
attribute_map = {"max_position_embeddings": "context_length"}
|
89 |
|
90 |
-
def __init__(
|
91 |
self,
|
92 |
vocab_size=65536,
|
93 |
-
context_length=4096,
|
94 |
hidden_size=768,
|
95 |
num_hidden_layers=24,
|
96 |
attention_hidden_size=None,
|
|
|
97 |
head_size=64,
|
98 |
intermediate_size=None,
|
99 |
layer_norm_epsilon=1e-5,
|
@@ -102,14 +99,13 @@ class Rwkv5Config(PretrainedConfig):
|
|
102 |
rescale_every=6,
|
103 |
tie_word_embeddings=False,
|
104 |
use_cache=True,
|
105 |
-
model_version="5_2",
|
106 |
**kwargs,
|
107 |
):
|
108 |
self.vocab_size = vocab_size
|
109 |
-
self.context_length = context_length
|
110 |
self.hidden_size = hidden_size
|
111 |
self.num_hidden_layers = num_hidden_layers
|
112 |
self.attention_hidden_size = attention_hidden_size if attention_hidden_size is not None else hidden_size
|
|
|
113 |
self.head_size = head_size
|
114 |
self.intermediate_size = None
|
115 |
self.layer_norm_epsilon = layer_norm_epsilon
|
@@ -118,7 +114,6 @@ class Rwkv5Config(PretrainedConfig):
|
|
118 |
|
119 |
self.bos_token_id = bos_token_id
|
120 |
self.eos_token_id = eos_token_id
|
121 |
-
self.model_version = model_version
|
122 |
|
123 |
super().__init__(
|
124 |
tie_word_embeddings=tie_word_embeddings, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs
|
|
|
21 |
|
22 |
logger = logging.get_logger(__name__)
|
23 |
|
24 |
+
RWKV5_PRETRAINED_CONFIG_ARCHIVE_MAP = {}
|
|
|
|
|
25 |
|
26 |
|
27 |
class Rwkv5Config(PretrainedConfig):
|
28 |
"""
|
29 |
+
This is the configuration class to store the configuration of a [`Rwkv5Model`]. It is used to instantiate a RWKV5
|
30 |
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
31 |
defaults will yield a similar configuration to that of the RWVK-4
|
32 |
+
[RWKV/rwkv-5-world-1b5](https://huggingface.co/RWKV/rwkv-5-world-1b5) architecture.
|
33 |
|
34 |
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
35 |
documentation from [`PretrainedConfig`] for more information.
|
36 |
|
37 |
|
38 |
Args:
|
39 |
+
vocab_size (`int`, *optional*, defaults to 65536):
|
40 |
+
Vocabulary size of the RWKV5 model. Defines the number of different tokens that can be represented by the
|
41 |
+
`inputs_ids` passed when calling [`Rwkv5Model`].
|
42 |
+
hidden_size (`int`, *optional*, defaults to 768):
|
|
|
|
|
|
|
43 |
Dimensionality of the embeddings and hidden states.
|
44 |
+
num_hidden_layers (`int`, *optional*, defaults to 24):
|
45 |
Number of hidden layers in the model.
|
46 |
attention_hidden_size (`int`, *optional*):
|
47 |
Dimensionality of the attention hidden states. Will default to `hidden_size` if unset.
|
48 |
+
num_attention_heads (`int`, *optional*, defaults to 64):
|
49 |
+
The attention heads to use in rwkv5 self_attention module.
|
50 |
+
head_size (`int`, *optional*, defaults to 64): head_size of rwkv5 self_attention module.
|
51 |
intermediate_size (`int`, *optional*):
|
52 |
Dimensionality of the inner feed-forward layers. Will default to 4 times `hidden_size` if unset.
|
53 |
+
layer_norm_epsilon (`float`, *optional*, defaults to 1e-05):
|
54 |
The epsilon to use in the layer normalization layers.
|
55 |
bos_token_id (`int`, *optional*, defaults to 0):
|
56 |
+
The id of the beginning of sentence token in the vocabulary. Defaults to 0 as RWKV5 uses the same tokenizer
|
57 |
as GPTNeoX.
|
58 |
eos_token_id (`int`, *optional*, defaults to 0):
|
59 |
+
The id of the end of sentence token in the vocabulary. Defaults to 0 as RWKV5 uses the same tokenizer as
|
60 |
GPTNeoX.
|
61 |
+
rescale_every (`int`, *optional*, defaults to 6):
|
62 |
At inference, the hidden states (and weights of the correponding output layers) are divided by 2 every
|
63 |
`rescale_every` layer. If set to 0 or a negative number, no rescale is done.
|
64 |
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
|
|
70 |
Example:
|
71 |
|
72 |
```python
|
73 |
+
>>> from transformers import Rwkv5Config, Rwkv5Model
|
74 |
|
75 |
+
>>> # Initializing a Rwkv5 configuration
|
76 |
+
>>> configuration = Rwkv5Config()
|
77 |
|
78 |
>>> # Initializing a model (with random weights) from the configuration
|
79 |
+
>>> model = Rwkv5Model(configuration)
|
80 |
|
81 |
>>> # Accessing the model configuration
|
82 |
>>> configuration = model.config
|
83 |
```"""
|
84 |
|
85 |
model_type = "rwkv5"
|
|
|
86 |
|
87 |
+
def __init__(
|
88 |
self,
|
89 |
vocab_size=65536,
|
|
|
90 |
hidden_size=768,
|
91 |
num_hidden_layers=24,
|
92 |
attention_hidden_size=None,
|
93 |
+
num_attention_heads=64,
|
94 |
head_size=64,
|
95 |
intermediate_size=None,
|
96 |
layer_norm_epsilon=1e-5,
|
|
|
99 |
rescale_every=6,
|
100 |
tie_word_embeddings=False,
|
101 |
use_cache=True,
|
|
|
102 |
**kwargs,
|
103 |
):
|
104 |
self.vocab_size = vocab_size
|
|
|
105 |
self.hidden_size = hidden_size
|
106 |
self.num_hidden_layers = num_hidden_layers
|
107 |
self.attention_hidden_size = attention_hidden_size if attention_hidden_size is not None else hidden_size
|
108 |
+
self.num_attention_heads = num_attention_heads
|
109 |
self.head_size = head_size
|
110 |
self.intermediate_size = None
|
111 |
self.layer_norm_epsilon = layer_norm_epsilon
|
|
|
114 |
|
115 |
self.bos_token_id = bos_token_id
|
116 |
self.eos_token_id = eos_token_id
|
|
|
117 |
|
118 |
super().__init__(
|
119 |
tie_word_embeddings=tie_word_embeddings, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs
|
modeling_rwkv5.py
CHANGED
@@ -15,16 +15,13 @@
|
|
15 |
# limitations under the License.
|
16 |
"""PyTorch RWKV5 World model."""
|
17 |
|
18 |
-
import math
|
19 |
from dataclasses import dataclass
|
20 |
-
from pathlib import Path
|
21 |
from typing import List, Optional, Tuple, Union
|
22 |
|
23 |
import torch
|
|
|
24 |
import torch.utils.checkpoint
|
25 |
from torch import nn
|
26 |
-
import torch.nn.functional as F
|
27 |
-
from torch.nn import CrossEntropyLoss
|
28 |
|
29 |
from transformers.modeling_utils import PreTrainedModel
|
30 |
from transformers.utils import (
|
@@ -32,77 +29,59 @@ from transformers.utils import (
|
|
32 |
add_code_sample_docstrings,
|
33 |
add_start_docstrings,
|
34 |
add_start_docstrings_to_model_forward,
|
35 |
-
is_ninja_available,
|
36 |
-
is_torch_cuda_available,
|
37 |
logging,
|
38 |
)
|
|
|
39 |
from .configuration_rwkv5 import Rwkv5Config
|
40 |
|
41 |
|
42 |
logger = logging.get_logger(__name__)
|
43 |
|
44 |
-
_CHECKPOINT_FOR_DOC = "RWKV/rwkv-5-world"
|
45 |
_CONFIG_FOR_DOC = "Rwkv5Config"
|
46 |
|
47 |
-
|
48 |
-
|
|
|
49 |
]
|
50 |
|
51 |
-
def rwkv_linear_attention_v5_0(H, S, T, hidden, time_decay, time_first, receptance, key, value, lxw, lxb, ow, state, return_state=False, seq_mode=True):
|
52 |
-
time_decay = torch.exp(-torch.exp(time_decay.float())).reshape(-1,1,1)
|
53 |
-
time_first = torch.exp(time_first.float()).reshape(-1,1,1)
|
54 |
-
lxw = lxw.float()
|
55 |
-
lxb = lxb.float()
|
56 |
-
|
57 |
-
if seq_mode:
|
58 |
-
w = time_decay.reshape(-1, 1)
|
59 |
-
u = time_first.reshape(-1, 1)
|
60 |
-
ws = w.pow(T).reshape(H, 1, 1)
|
61 |
-
ind = torch.arange(T-1, -1, -1, device=w.device).unsqueeze(0).repeat(H, 1)
|
62 |
-
w = w.repeat(1, T).pow(ind)
|
63 |
-
wk = w.reshape(H, 1, T)
|
64 |
-
wb = wk.transpose(-2, -1).flip(1)
|
65 |
-
w = torch.cat([w[:, 1:], u], dim=1)
|
66 |
-
w = F.pad(w, (0, T))
|
67 |
-
w = torch.tile(w, [T])
|
68 |
-
w = w[:, :-T].reshape(-1, T, 2 * T - 1)
|
69 |
-
w = w[:, :, T-1:].reshape(H, T, T)
|
70 |
-
out = ((receptance @ key) * w) @ value + (receptance @ state) * wb
|
71 |
-
state = ws * state + (key * wk) @ value
|
72 |
-
|
73 |
-
out = out.transpose(1, 2).contiguous().reshape(T, H*S)
|
74 |
-
out = F.group_norm(out, num_groups=H, weight=lxw, bias=lxb)
|
75 |
-
out = out.to(dtype=hidden.dtype)
|
76 |
-
out = out @ ow
|
77 |
-
else:
|
78 |
-
a = key @ value
|
79 |
-
out = receptance @ (time_first * a + state)
|
80 |
-
state = a + time_decay * state
|
81 |
-
out = out.flatten()
|
82 |
-
out = F.group_norm(out.unsqueeze(0), num_groups=H, weight=lxw, bias=lxb)
|
83 |
-
out = out.to(dtype=hidden.dtype)
|
84 |
-
out = out @ ow
|
85 |
|
86 |
-
|
87 |
-
|
88 |
-
|
89 |
-
|
90 |
-
|
91 |
-
|
92 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
93 |
lxw = lxw.float()
|
94 |
lxb = lxb.float()
|
|
|
95 |
out = torch.empty((B, T, H, S), dtype=receptance.dtype, device=receptance.device)
|
96 |
for t in range(T):
|
97 |
-
rt = receptance[
|
98 |
-
kt = key[
|
99 |
-
vt = value[
|
100 |
at = kt @ vt
|
101 |
out[:, t] = (rt @ (time_first * at + state)).squeeze(2)
|
102 |
state = at + time_decay * state
|
103 |
|
104 |
-
out = out.reshape(B*T, H*S)
|
105 |
-
out = F.group_norm(out, num_groups=H, weight=lxw, bias=lxb).reshape(B, T, H*S)
|
106 |
out = out.to(dtype=hidden.dtype) * gate
|
107 |
out = out @ ow
|
108 |
|
@@ -123,13 +102,9 @@ class RwkvSelfAttention(nn.Module):
|
|
123 |
)
|
124 |
self.attention_hidden_size = attention_hidden_size
|
125 |
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
self.time_mix_gate = nn.Parameter(torch.empty(1, 1, hidden_size))
|
130 |
-
else:
|
131 |
-
self.time_decay = nn.Parameter(torch.empty(num_attention_heads))
|
132 |
-
self.time_first = nn.Parameter(torch.empty(num_attention_heads))
|
133 |
|
134 |
self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
|
135 |
self.time_mix_value = nn.Parameter(torch.empty(1, 1, hidden_size))
|
@@ -139,8 +114,7 @@ class RwkvSelfAttention(nn.Module):
|
|
139 |
self.key = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
140 |
self.value = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
141 |
self.receptance = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
142 |
-
|
143 |
-
self.gate = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
144 |
self.output = nn.Linear(attention_hidden_size, hidden_size, bias=False)
|
145 |
# https://github.com/BlinkDL/RWKV-LM/blob/3db37a72356b736966ddd377268f02b80963af3f/RWKV-v4neo/src/model.py#L190C1-L190C1
|
146 |
self.ln_x = nn.GroupNorm(hidden_size // config.head_size, hidden_size)
|
@@ -155,32 +129,22 @@ class RwkvSelfAttention(nn.Module):
|
|
155 |
if state is not None:
|
156 |
shifted[:, 0] = state[0][:, :, self.layer_id]
|
157 |
if len(shifted.size()) == 2:
|
158 |
-
shifted = shifted.unsqueeze(1)
|
159 |
key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key)
|
160 |
value = hidden * self.time_mix_value + shifted * (1 - self.time_mix_value)
|
161 |
receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance)
|
162 |
-
|
163 |
-
|
164 |
-
|
165 |
-
# if hidden.size(1) == 1 and state is not None:
|
166 |
-
# receptance = self.receptance(receptance).to(torch.float32).view(B, H, 1, S)
|
167 |
-
# key = self.key(key).to(torch.float32).view(B, H, S, 1)
|
168 |
-
# value = self.value(value).to(torch.float32).view(B, H, 1, S)
|
169 |
-
# else:
|
170 |
# https://github.com/BlinkDL/ChatRWKV/blob/main/rwkv_pip_package/src/rwkv/model.py#L693
|
171 |
key = self.key(key).to(torch.float32).view(B, T, H, S).transpose(1, 2).transpose(-2, -1)
|
172 |
value = self.value(value).to(torch.float32).view(B, T, H, S).transpose(1, 2)
|
173 |
receptance = self.receptance(receptance).to(torch.float32).view(B, T, H, S).transpose(1, 2)
|
|
|
174 |
|
175 |
-
if self.config.model_version == "5_2":
|
176 |
-
gate = F.silu(self.gate(gate))
|
177 |
-
|
178 |
if state is not None:
|
179 |
state[0][:, :, self.layer_id] = hidden[:, -1]
|
180 |
-
|
181 |
-
|
182 |
-
return receptance, key, value, gate, state
|
183 |
-
return receptance, key, value, state
|
184 |
|
185 |
def forward(self, hidden, state=None, use_cache=False, seq_mode=True):
|
186 |
B = hidden.shape[0]
|
@@ -188,13 +152,9 @@ class RwkvSelfAttention(nn.Module):
|
|
188 |
S = hidden.shape[-1] // H
|
189 |
T = hidden.shape[1]
|
190 |
|
191 |
-
|
192 |
-
receptance, key, value, gate, state = self.extract_key_value(B, H, S, T, hidden, state=state)
|
193 |
-
else:
|
194 |
-
receptance, key, value, state = self.extract_key_value(H, S, T, hidden, state=state)
|
195 |
layer_state = state[1][:, :, :, :, self.layer_id] if state is not None else None
|
196 |
-
|
197 |
-
rwkv, layer_state = rwkv_linear_attention_v5_2(
|
198 |
B,
|
199 |
H,
|
200 |
S,
|
@@ -214,24 +174,6 @@ class RwkvSelfAttention(nn.Module):
|
|
214 |
return_state=use_cache,
|
215 |
seq_mode=seq_mode,
|
216 |
)
|
217 |
-
else:
|
218 |
-
rwkv, layer_state = rwkv_linear_attention_v5_0(
|
219 |
-
H,
|
220 |
-
S,
|
221 |
-
T,
|
222 |
-
hidden,
|
223 |
-
self.time_decay,
|
224 |
-
self.time_first,
|
225 |
-
receptance,
|
226 |
-
key,
|
227 |
-
value,
|
228 |
-
self.ln_x.weight,
|
229 |
-
self.ln_x.bias,
|
230 |
-
self.output.weight.t(),
|
231 |
-
state=layer_state,
|
232 |
-
return_state=use_cache,
|
233 |
-
seq_mode=seq_mode,
|
234 |
-
)
|
235 |
|
236 |
if layer_state is not None:
|
237 |
state[1][:, :, :, :, self.layer_id] = layer_state
|
@@ -246,19 +188,16 @@ class RwkvFeedForward(nn.Module):
|
|
246 |
self.layer_id = layer_id
|
247 |
hidden_size = config.hidden_size
|
248 |
# https://github.com/BlinkDL/RWKV-LM/blob/3db37a72356b736966ddd377268f02b80963af3f/RWKV-v4neo/train.py#L168
|
249 |
-
|
250 |
-
intermediate_size
|
251 |
-
|
252 |
-
)
|
253 |
-
|
254 |
-
intermediate_size = (
|
255 |
-
config.intermediate_size if config.intermediate_size is not None else 4 * config.hidden_size
|
256 |
-
)
|
257 |
|
258 |
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
259 |
self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
|
260 |
self.time_mix_receptance = nn.Parameter(torch.empty(1, 1, hidden_size))
|
261 |
-
|
262 |
self.key = nn.Linear(hidden_size, intermediate_size, bias=False)
|
263 |
self.receptance = nn.Linear(hidden_size, hidden_size, bias=False)
|
264 |
self.value = nn.Linear(intermediate_size, hidden_size, bias=False)
|
@@ -301,7 +240,6 @@ class RwkvBlock(nn.Module):
|
|
301 |
self.feed_forward = RwkvFeedForward(config, layer_id)
|
302 |
|
303 |
def forward(self, hidden, state=None, use_cache=False, output_attentions=False, seq_mode=True):
|
304 |
-
|
305 |
attention, state = self.attention(self.ln1(hidden), state=state, use_cache=use_cache, seq_mode=seq_mode)
|
306 |
hidden = hidden + attention
|
307 |
|
@@ -317,16 +255,18 @@ class RwkvBlock(nn.Module):
|
|
317 |
return outputs
|
318 |
|
319 |
|
320 |
-
class
|
321 |
"""
|
322 |
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
323 |
models.
|
324 |
"""
|
325 |
|
326 |
config_class = Rwkv5Config
|
327 |
-
base_model_prefix = "
|
328 |
_no_split_modules = ["RwkvBlock"]
|
329 |
_keep_in_fp32_modules = ["time_decay", "time_first"]
|
|
|
|
|
330 |
|
331 |
def _init_weights(self, module):
|
332 |
"""Initialize the weights."""
|
@@ -335,7 +275,7 @@ class RwkvPreTrainedModel(PreTrainedModel):
|
|
335 |
num_hidden_layers = module.config.num_hidden_layers
|
336 |
hidden_size = module.config.hidden_size
|
337 |
attention_hidden_size = module.attention_hidden_size
|
338 |
-
num_attention_heads = hidden_size // module.config.
|
339 |
|
340 |
ratio_0_to_1 = layer_id / (num_hidden_layers - 1) # 0 to 1
|
341 |
ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers) # 1 to ~0
|
@@ -347,43 +287,30 @@ class RwkvPreTrainedModel(PreTrainedModel):
|
|
347 |
)
|
348 |
time_weight = time_weight[None, None, :]
|
349 |
|
350 |
-
|
351 |
-
|
352 |
-
|
353 |
-
|
354 |
-
|
355 |
-
]
|
356 |
-
else:
|
357 |
-
# https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-v4neo/src/model.py#L172
|
358 |
-
decay_speed = [
|
359 |
-
-6.0 + 5.0 * (h / (num_attention_heads - 1)) ** (0.7 + 1.3 * ratio_0_to_1)
|
360 |
-
for h in range(num_attention_heads)
|
361 |
-
]
|
362 |
decay_speed = torch.tensor(decay_speed, dtype=module.time_decay.dtype, device=module.time_decay.device)
|
363 |
-
|
364 |
-
|
365 |
-
|
366 |
-
|
367 |
-
|
368 |
-
|
369 |
-
|
370 |
-
|
371 |
-
else:
|
372 |
-
tmp = torch.ones(num_attention_heads) * (-3.0)
|
373 |
|
374 |
with torch.no_grad():
|
375 |
-
|
376 |
-
|
377 |
-
module.time_faaaa.data = tmp.reshape(num_attention_heads, module.config.head_size)
|
378 |
-
else:
|
379 |
-
module.time_decay.data = decay_speed
|
380 |
-
module.time_first.data = tmp
|
381 |
-
|
382 |
module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
|
|
|
383 |
module.time_mix_value.data = torch.pow(time_weight, ratio_1_to_almost0) + 0.3 * ratio_0_to_1
|
384 |
module.time_mix_receptance.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0)
|
385 |
-
|
386 |
-
|
387 |
elif isinstance(module, RwkvFeedForward):
|
388 |
layer_id = module.layer_id
|
389 |
num_hidden_layers = module.config.num_hidden_layers
|
@@ -402,13 +329,9 @@ class RwkvPreTrainedModel(PreTrainedModel):
|
|
402 |
module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
|
403 |
module.time_mix_receptance.data = torch.pow(time_weight, ratio_1_to_almost0)
|
404 |
|
405 |
-
def _set_gradient_checkpointing(self, module, value=False):
|
406 |
-
if isinstance(module, RwkvModel):
|
407 |
-
module.gradient_checkpointing = value
|
408 |
-
|
409 |
|
410 |
@dataclass
|
411 |
-
class
|
412 |
"""
|
413 |
Class for the RWKV model outputs.
|
414 |
|
@@ -420,15 +343,12 @@ class RwkvOutput(ModelOutput):
|
|
420 |
avoid providing the old `input_ids`.
|
421 |
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
422 |
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
423 |
-
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
424 |
-
|
425 |
-
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
426 |
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
427 |
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
428 |
-
sequence_length)`.
|
429 |
-
|
430 |
-
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
431 |
-
heads.
|
432 |
"""
|
433 |
|
434 |
last_hidden_state: torch.FloatTensor = None
|
@@ -438,7 +358,7 @@ class RwkvOutput(ModelOutput):
|
|
438 |
|
439 |
|
440 |
@dataclass
|
441 |
-
class
|
442 |
"""
|
443 |
Base class for causal language model (or autoregressive) outputs.
|
444 |
|
@@ -452,33 +372,27 @@ class RwkvCausalLMOutput(ModelOutput):
|
|
452 |
avoid providing the old `input_ids`.
|
453 |
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
454 |
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
455 |
-
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
456 |
-
|
457 |
-
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
458 |
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
459 |
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
460 |
-
sequence_length)`.
|
461 |
-
|
462 |
-
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
463 |
-
heads.
|
464 |
"""
|
465 |
|
466 |
loss: Optional[torch.FloatTensor] = None
|
467 |
logits: torch.FloatTensor = None
|
468 |
state: Optional[List[torch.FloatTensor]] = None
|
469 |
-
|
470 |
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
471 |
|
472 |
|
473 |
RWKV_START_DOCSTRING = r"""
|
474 |
-
|
475 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
476 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
477 |
-
etc.)
|
478 |
-
|
479 |
-
|
480 |
-
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
|
481 |
-
and behavior.
|
482 |
|
483 |
Parameters:
|
484 |
config ([`Rwkv5Config`]): Model configuration class with all the parameters of the model.
|
@@ -491,15 +405,10 @@ RWKV_INPUTS_DOCSTRING = r"""
|
|
491 |
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
|
492 |
`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
|
493 |
`past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
|
494 |
-
sequence tokens in the vocabulary.
|
495 |
-
|
496 |
-
|
497 |
-
|
498 |
-
|
499 |
-
Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
|
500 |
-
[`PreTrainedTokenizer.__call__`] for details.
|
501 |
-
|
502 |
-
[What are input IDs?](../glossary#input-ids)
|
503 |
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
504 |
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
505 |
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
@@ -524,7 +433,7 @@ RWKV_INPUTS_DOCSTRING = r"""
|
|
524 |
"The bare RWKV Model transformer outputting raw hidden-states without any specific head on top.",
|
525 |
RWKV_START_DOCSTRING,
|
526 |
)
|
527 |
-
class
|
528 |
def __init__(self, config):
|
529 |
super().__init__(config)
|
530 |
|
@@ -535,6 +444,8 @@ class RwkvModel(RwkvPreTrainedModel):
|
|
535 |
self.layers_are_rescaled = False
|
536 |
self.pre_ln_flag = False
|
537 |
|
|
|
|
|
538 |
# Initialize weights and apply final processing
|
539 |
self.post_init()
|
540 |
|
@@ -547,28 +458,31 @@ class RwkvModel(RwkvPreTrainedModel):
|
|
547 |
@add_start_docstrings_to_model_forward(RWKV_INPUTS_DOCSTRING)
|
548 |
@add_code_sample_docstrings(
|
549 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
550 |
-
output_type=
|
551 |
config_class=_CONFIG_FOR_DOC,
|
552 |
)
|
553 |
def forward(
|
554 |
self,
|
555 |
input_ids: Optional[torch.LongTensor] = None,
|
|
|
556 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
557 |
state: Optional[List[torch.FloatTensor]] = None,
|
558 |
use_cache: Optional[bool] = None,
|
559 |
output_attentions: Optional[bool] = None,
|
560 |
output_hidden_states: Optional[bool] = None,
|
561 |
return_dict: Optional[bool] = None,
|
562 |
-
) -> Union[Tuple,
|
563 |
-
seq_mode = input_ids.shape[1] > 1
|
564 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
565 |
output_hidden_states = (
|
566 |
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
567 |
)
|
568 |
-
|
|
|
569 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
570 |
|
571 |
-
if self.training == self.layers_are_rescaled and (
|
|
|
|
|
572 |
self._rescale_layers()
|
573 |
|
574 |
if input_ids is not None and inputs_embeds is not None:
|
@@ -578,23 +492,55 @@ class RwkvModel(RwkvPreTrainedModel):
|
|
578 |
|
579 |
if inputs_embeds is None:
|
580 |
if not self.pre_ln_flag:
|
581 |
-
normalized_weight = F.layer_norm(
|
|
|
|
|
|
|
|
|
|
|
582 |
self.embeddings.weight = nn.Parameter(normalized_weight)
|
583 |
self.pre_ln_flag = True
|
|
|
584 |
inputs_embeds = self.embeddings(input_ids)
|
585 |
|
586 |
if use_cache and state is None:
|
587 |
# https://github.com/BlinkDL/ChatRWKV/blob/main/rwkv_pip_package/src/rwkv/model.py#L904-L906
|
588 |
state = []
|
589 |
-
num_attention_heads = self.config.hidden_size // self.config.
|
590 |
-
state.append(
|
591 |
-
|
592 |
-
|
593 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
594 |
|
|
|
595 |
hidden_states = inputs_embeds
|
596 |
-
|
597 |
-
cnt += 1
|
598 |
all_self_attentions = () if output_attentions else None
|
599 |
all_hidden_states = () if output_hidden_states else None
|
600 |
for idx, block in enumerate(self.blocks):
|
@@ -622,11 +568,11 @@ class RwkvModel(RwkvPreTrainedModel):
|
|
622 |
if not return_dict:
|
623 |
return (hidden_states, state, all_hidden_states, all_self_attentions)
|
624 |
|
625 |
-
return
|
626 |
last_hidden_state=hidden_states,
|
627 |
state=state,
|
628 |
-
hidden_states=all_hidden_states,
|
629 |
-
attentions=all_self_attentions,
|
630 |
)
|
631 |
|
632 |
def _rescale_layers(self):
|
@@ -645,6 +591,7 @@ class RwkvModel(RwkvPreTrainedModel):
|
|
645 |
|
646 |
self.layers_are_rescaled = not self.training
|
647 |
|
|
|
648 |
@add_start_docstrings(
|
649 |
"""
|
650 |
The RWKV Model transformer with a language modeling head on top (linear layer with weights tied to the input
|
@@ -652,10 +599,12 @@ class RwkvModel(RwkvPreTrainedModel):
|
|
652 |
""",
|
653 |
RWKV_START_DOCSTRING,
|
654 |
)
|
655 |
-
class
|
|
|
|
|
656 |
def __init__(self, config):
|
657 |
super().__init__(config)
|
658 |
-
self.rwkv =
|
659 |
self.head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
660 |
|
661 |
# Initialize weights and apply final processing
|
@@ -684,7 +633,7 @@ class RwkvForCausalLM(RwkvPreTrainedModel):
|
|
684 |
@add_start_docstrings_to_model_forward(RWKV_INPUTS_DOCSTRING)
|
685 |
@add_code_sample_docstrings(
|
686 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
687 |
-
output_type=
|
688 |
config_class=_CONFIG_FOR_DOC,
|
689 |
)
|
690 |
def forward(
|
@@ -698,7 +647,7 @@ class RwkvForCausalLM(RwkvPreTrainedModel):
|
|
698 |
output_attentions: Optional[bool] = None,
|
699 |
output_hidden_states: Optional[bool] = None,
|
700 |
return_dict: Optional[bool] = None,
|
701 |
-
) -> Union[Tuple,
|
702 |
r"""
|
703 |
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
704 |
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
@@ -716,30 +665,23 @@ class RwkvForCausalLM(RwkvPreTrainedModel):
|
|
716 |
output_hidden_states=output_hidden_states,
|
717 |
return_dict=return_dict,
|
718 |
)
|
719 |
-
|
720 |
-
state = rwkv_outputs.state
|
721 |
|
722 |
-
logits = self.head(
|
723 |
|
724 |
loss = None
|
725 |
if labels is not None:
|
726 |
-
#
|
727 |
-
|
728 |
-
# Shift so that tokens < n predict n
|
729 |
-
shift_logits = logits[..., :-1, :].contiguous()
|
730 |
-
shift_labels = labels[..., 1:].contiguous()
|
731 |
-
# Flatten the tokens
|
732 |
-
loss_fct = CrossEntropyLoss()
|
733 |
-
loss = loss_fct(shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1))
|
734 |
|
735 |
if not return_dict:
|
736 |
output = (logits,) + rwkv_outputs[1:]
|
737 |
return ((loss,) + output) if loss is not None else output
|
738 |
|
739 |
-
return
|
740 |
loss=loss,
|
741 |
logits=logits,
|
742 |
state=rwkv_outputs.state,
|
743 |
-
|
744 |
attentions=rwkv_outputs.attentions,
|
745 |
)
|
|
|
15 |
# limitations under the License.
|
16 |
"""PyTorch RWKV5 World model."""
|
17 |
|
|
|
18 |
from dataclasses import dataclass
|
|
|
19 |
from typing import List, Optional, Tuple, Union
|
20 |
|
21 |
import torch
|
22 |
+
import torch.nn.functional as F
|
23 |
import torch.utils.checkpoint
|
24 |
from torch import nn
|
|
|
|
|
25 |
|
26 |
from transformers.modeling_utils import PreTrainedModel
|
27 |
from transformers.utils import (
|
|
|
29 |
add_code_sample_docstrings,
|
30 |
add_start_docstrings,
|
31 |
add_start_docstrings_to_model_forward,
|
|
|
|
|
32 |
logging,
|
33 |
)
|
34 |
+
|
35 |
from .configuration_rwkv5 import Rwkv5Config
|
36 |
|
37 |
|
38 |
logger = logging.get_logger(__name__)
|
39 |
|
40 |
+
_CHECKPOINT_FOR_DOC = "RWKV/rwkv-5-world-1b5"
|
41 |
_CONFIG_FOR_DOC = "Rwkv5Config"
|
42 |
|
43 |
+
RWKV5_PRETRAINED_MODEL_ARCHIVE_LIST = [
|
44 |
+
"RWKV/rwkv-5-world-1b5",
|
45 |
+
# See all RWKV models at https://huggingface.co/models?filter=rwkv
|
46 |
]
|
47 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
48 |
|
49 |
+
def rwkv_linear_attention_v5(
|
50 |
+
B,
|
51 |
+
H,
|
52 |
+
S,
|
53 |
+
T,
|
54 |
+
n_head,
|
55 |
+
hidden,
|
56 |
+
time_decay,
|
57 |
+
time_first,
|
58 |
+
receptance,
|
59 |
+
key,
|
60 |
+
value,
|
61 |
+
gate,
|
62 |
+
lxw,
|
63 |
+
lxb,
|
64 |
+
ow,
|
65 |
+
state,
|
66 |
+
return_state=False,
|
67 |
+
seq_mode=True,
|
68 |
+
):
|
69 |
+
time_decay = torch.exp(-torch.exp(time_decay.float())).reshape(-1, 1, 1).reshape(n_head, -1, 1)
|
70 |
+
time_first = time_first.float().reshape(-1, 1, 1).reshape(n_head, -1, 1)
|
71 |
lxw = lxw.float()
|
72 |
lxb = lxb.float()
|
73 |
+
# if seq_mode:
|
74 |
out = torch.empty((B, T, H, S), dtype=receptance.dtype, device=receptance.device)
|
75 |
for t in range(T):
|
76 |
+
rt = receptance[:, :, t : t + 1, :]
|
77 |
+
kt = key[:, :, :, t : t + 1]
|
78 |
+
vt = value[:, :, t : t + 1, :]
|
79 |
at = kt @ vt
|
80 |
out[:, t] = (rt @ (time_first * at + state)).squeeze(2)
|
81 |
state = at + time_decay * state
|
82 |
|
83 |
+
out = out.reshape(B * T, H * S)
|
84 |
+
out = F.group_norm(out, num_groups=H, weight=lxw, bias=lxb).reshape(B, T, H * S)
|
85 |
out = out.to(dtype=hidden.dtype) * gate
|
86 |
out = out @ ow
|
87 |
|
|
|
102 |
)
|
103 |
self.attention_hidden_size = attention_hidden_size
|
104 |
|
105 |
+
self.time_decay = nn.Parameter(torch.empty(num_attention_heads, config.head_size))
|
106 |
+
self.time_faaaa = nn.Parameter(torch.empty(num_attention_heads, config.head_size))
|
107 |
+
self.time_mix_gate = nn.Parameter(torch.empty(1, 1, hidden_size))
|
|
|
|
|
|
|
|
|
108 |
|
109 |
self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
|
110 |
self.time_mix_value = nn.Parameter(torch.empty(1, 1, hidden_size))
|
|
|
114 |
self.key = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
115 |
self.value = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
116 |
self.receptance = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
117 |
+
self.gate = nn.Linear(hidden_size, attention_hidden_size, bias=False)
|
|
|
118 |
self.output = nn.Linear(attention_hidden_size, hidden_size, bias=False)
|
119 |
# https://github.com/BlinkDL/RWKV-LM/blob/3db37a72356b736966ddd377268f02b80963af3f/RWKV-v4neo/src/model.py#L190C1-L190C1
|
120 |
self.ln_x = nn.GroupNorm(hidden_size // config.head_size, hidden_size)
|
|
|
129 |
if state is not None:
|
130 |
shifted[:, 0] = state[0][:, :, self.layer_id]
|
131 |
if len(shifted.size()) == 2:
|
132 |
+
shifted = shifted.unsqueeze(1)
|
133 |
key = hidden * self.time_mix_key + shifted * (1 - self.time_mix_key)
|
134 |
value = hidden * self.time_mix_value + shifted * (1 - self.time_mix_value)
|
135 |
receptance = hidden * self.time_mix_receptance + shifted * (1 - self.time_mix_receptance)
|
136 |
+
gate = hidden * self.time_mix_gate + shifted * (1 - self.time_mix_gate)
|
137 |
+
|
|
|
|
|
|
|
|
|
|
|
|
|
138 |
# https://github.com/BlinkDL/ChatRWKV/blob/main/rwkv_pip_package/src/rwkv/model.py#L693
|
139 |
key = self.key(key).to(torch.float32).view(B, T, H, S).transpose(1, 2).transpose(-2, -1)
|
140 |
value = self.value(value).to(torch.float32).view(B, T, H, S).transpose(1, 2)
|
141 |
receptance = self.receptance(receptance).to(torch.float32).view(B, T, H, S).transpose(1, 2)
|
142 |
+
gate = F.silu(self.gate(gate))
|
143 |
|
|
|
|
|
|
|
144 |
if state is not None:
|
145 |
state[0][:, :, self.layer_id] = hidden[:, -1]
|
146 |
+
|
147 |
+
return receptance, key, value, gate, state
|
|
|
|
|
148 |
|
149 |
def forward(self, hidden, state=None, use_cache=False, seq_mode=True):
|
150 |
B = hidden.shape[0]
|
|
|
152 |
S = hidden.shape[-1] // H
|
153 |
T = hidden.shape[1]
|
154 |
|
155 |
+
receptance, key, value, gate, state = self.extract_key_value(B, H, S, T, hidden, state=state)
|
|
|
|
|
|
|
156 |
layer_state = state[1][:, :, :, :, self.layer_id] if state is not None else None
|
157 |
+
rwkv, layer_state = rwkv_linear_attention_v5(
|
|
|
158 |
B,
|
159 |
H,
|
160 |
S,
|
|
|
174 |
return_state=use_cache,
|
175 |
seq_mode=seq_mode,
|
176 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
177 |
|
178 |
if layer_state is not None:
|
179 |
state[1][:, :, :, :, self.layer_id] = layer_state
|
|
|
188 |
self.layer_id = layer_id
|
189 |
hidden_size = config.hidden_size
|
190 |
# https://github.com/BlinkDL/RWKV-LM/blob/3db37a72356b736966ddd377268f02b80963af3f/RWKV-v4neo/train.py#L168
|
191 |
+
intermediate_size = (
|
192 |
+
config.intermediate_size
|
193 |
+
if config.intermediate_size is not None
|
194 |
+
else int((config.hidden_size * 3.5) // 32 * 32)
|
195 |
+
)
|
|
|
|
|
|
|
196 |
|
197 |
self.time_shift = nn.ZeroPad2d((0, 0, 1, -1))
|
198 |
self.time_mix_key = nn.Parameter(torch.empty(1, 1, hidden_size))
|
199 |
self.time_mix_receptance = nn.Parameter(torch.empty(1, 1, hidden_size))
|
200 |
+
|
201 |
self.key = nn.Linear(hidden_size, intermediate_size, bias=False)
|
202 |
self.receptance = nn.Linear(hidden_size, hidden_size, bias=False)
|
203 |
self.value = nn.Linear(intermediate_size, hidden_size, bias=False)
|
|
|
240 |
self.feed_forward = RwkvFeedForward(config, layer_id)
|
241 |
|
242 |
def forward(self, hidden, state=None, use_cache=False, output_attentions=False, seq_mode=True):
|
|
|
243 |
attention, state = self.attention(self.ln1(hidden), state=state, use_cache=use_cache, seq_mode=seq_mode)
|
244 |
hidden = hidden + attention
|
245 |
|
|
|
255 |
return outputs
|
256 |
|
257 |
|
258 |
+
class Rwkv5PreTrainedModel(PreTrainedModel):
|
259 |
"""
|
260 |
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
261 |
models.
|
262 |
"""
|
263 |
|
264 |
config_class = Rwkv5Config
|
265 |
+
base_model_prefix = "rwkv"
|
266 |
_no_split_modules = ["RwkvBlock"]
|
267 |
_keep_in_fp32_modules = ["time_decay", "time_first"]
|
268 |
+
supports_gradient_checkpointing = True
|
269 |
+
training = False
|
270 |
|
271 |
def _init_weights(self, module):
|
272 |
"""Initialize the weights."""
|
|
|
275 |
num_hidden_layers = module.config.num_hidden_layers
|
276 |
hidden_size = module.config.hidden_size
|
277 |
attention_hidden_size = module.attention_hidden_size
|
278 |
+
num_attention_heads = hidden_size // module.config.num_attention_heads
|
279 |
|
280 |
ratio_0_to_1 = layer_id / (num_hidden_layers - 1) # 0 to 1
|
281 |
ratio_1_to_almost0 = 1.0 - (layer_id / num_hidden_layers) # 1 to ~0
|
|
|
287 |
)
|
288 |
time_weight = time_weight[None, None, :]
|
289 |
|
290 |
+
# https://github.com/BlinkDL/RWKV-LM/blob/main/RWKV-v4neo/src/model.py#L398
|
291 |
+
decay_speed = [
|
292 |
+
-6.0 + 5.0 * (h / (attention_hidden_size - 1)) ** (0.7 + 1.3 * ratio_0_to_1)
|
293 |
+
for h in range(attention_hidden_size)
|
294 |
+
]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
295 |
decay_speed = torch.tensor(decay_speed, dtype=module.time_decay.dtype, device=module.time_decay.device)
|
296 |
+
tmp = torch.tensor(
|
297 |
+
[
|
298 |
+
(1.0 - (i / (attention_hidden_size - 1.0))) * ratio_0_to_1 + 0.1 * ((i + 1) % 3 - 1)
|
299 |
+
for i in range(attention_hidden_size)
|
300 |
+
],
|
301 |
+
dtype=module.time_faaaa.dtype,
|
302 |
+
device=module.time_faaaa.device,
|
303 |
+
)
|
|
|
|
|
304 |
|
305 |
with torch.no_grad():
|
306 |
+
module.time_decay.data = decay_speed.reshape(num_attention_heads, module.config.num_attention_heads)
|
307 |
+
module.time_faaaa.data = tmp.reshape(num_attention_heads, module.config.num_attention_heads)
|
|
|
|
|
|
|
|
|
|
|
308 |
module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
|
309 |
+
|
310 |
module.time_mix_value.data = torch.pow(time_weight, ratio_1_to_almost0) + 0.3 * ratio_0_to_1
|
311 |
module.time_mix_receptance.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0)
|
312 |
+
module.time_mix_gate.data = torch.pow(time_weight, 0.5 * ratio_1_to_almost0)
|
313 |
+
|
314 |
elif isinstance(module, RwkvFeedForward):
|
315 |
layer_id = module.layer_id
|
316 |
num_hidden_layers = module.config.num_hidden_layers
|
|
|
329 |
module.time_mix_key.data = torch.pow(time_weight, ratio_1_to_almost0)
|
330 |
module.time_mix_receptance.data = torch.pow(time_weight, ratio_1_to_almost0)
|
331 |
|
|
|
|
|
|
|
|
|
332 |
|
333 |
@dataclass
|
334 |
+
class Rwkv5Output(ModelOutput):
|
335 |
"""
|
336 |
Class for the RWKV model outputs.
|
337 |
|
|
|
343 |
avoid providing the old `input_ids`.
|
344 |
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
345 |
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
346 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of
|
347 |
+
the model at the output of each layer plus the optional initial embedding outputs.
|
|
|
348 |
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
349 |
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
350 |
+
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
|
351 |
+
the self-attention heads.
|
|
|
|
|
352 |
"""
|
353 |
|
354 |
last_hidden_state: torch.FloatTensor = None
|
|
|
358 |
|
359 |
|
360 |
@dataclass
|
361 |
+
class Rwkv5CausalLMOutput(ModelOutput):
|
362 |
"""
|
363 |
Base class for causal language model (or autoregressive) outputs.
|
364 |
|
|
|
372 |
avoid providing the old `input_ids`.
|
373 |
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
374 |
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
375 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of
|
376 |
+
the model at the output of each layer plus the optional initial embedding outputs.
|
|
|
377 |
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
378 |
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
379 |
+
sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in
|
380 |
+
the self-attention heads.
|
|
|
|
|
381 |
"""
|
382 |
|
383 |
loss: Optional[torch.FloatTensor] = None
|
384 |
logits: torch.FloatTensor = None
|
385 |
state: Optional[List[torch.FloatTensor]] = None
|
386 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
387 |
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
388 |
|
389 |
|
390 |
RWKV_START_DOCSTRING = r"""
|
|
|
391 |
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
|
392 |
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
|
393 |
+
etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module)
|
394 |
+
subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to
|
395 |
+
general usage and behavior.
|
|
|
|
|
396 |
|
397 |
Parameters:
|
398 |
config ([`Rwkv5Config`]): Model configuration class with all the parameters of the model.
|
|
|
405 |
input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`):
|
406 |
`input_ids_length` = `sequence_length` if `past_key_values` is `None` else
|
407 |
`past_key_values[0][0].shape[-2]` (`sequence_length` of input past key value states). Indices of input
|
408 |
+
sequence tokens in the vocabulary. If `past_key_values` is used, only `input_ids` that do not have their
|
409 |
+
past calculated should be passed as `input_ids`. Indices can be obtained using [`AutoTokenizer`]. See
|
410 |
+
[`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input
|
411 |
+
IDs?](../glossary#input-ids)
|
|
|
|
|
|
|
|
|
|
|
412 |
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
|
413 |
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
|
414 |
is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
|
|
|
433 |
"The bare RWKV Model transformer outputting raw hidden-states without any specific head on top.",
|
434 |
RWKV_START_DOCSTRING,
|
435 |
)
|
436 |
+
class Rwkv5Model(Rwkv5PreTrainedModel):
|
437 |
def __init__(self, config):
|
438 |
super().__init__(config)
|
439 |
|
|
|
444 |
self.layers_are_rescaled = False
|
445 |
self.pre_ln_flag = False
|
446 |
|
447 |
+
self.gradient_checkpointing = False
|
448 |
+
|
449 |
# Initialize weights and apply final processing
|
450 |
self.post_init()
|
451 |
|
|
|
458 |
@add_start_docstrings_to_model_forward(RWKV_INPUTS_DOCSTRING)
|
459 |
@add_code_sample_docstrings(
|
460 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
461 |
+
output_type=Rwkv5Output,
|
462 |
config_class=_CONFIG_FOR_DOC,
|
463 |
)
|
464 |
def forward(
|
465 |
self,
|
466 |
input_ids: Optional[torch.LongTensor] = None,
|
467 |
+
attention_mask: Optional[torch.LongTensor] = None, # noqa
|
468 |
inputs_embeds: Optional[torch.FloatTensor] = None,
|
469 |
state: Optional[List[torch.FloatTensor]] = None,
|
470 |
use_cache: Optional[bool] = None,
|
471 |
output_attentions: Optional[bool] = None,
|
472 |
output_hidden_states: Optional[bool] = None,
|
473 |
return_dict: Optional[bool] = None,
|
474 |
+
) -> Union[Tuple, Rwkv5Output]:
|
|
|
475 |
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
476 |
output_hidden_states = (
|
477 |
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
478 |
)
|
479 |
+
# rwkv5 only support inference in huggingface.
|
480 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
481 |
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
482 |
|
483 |
+
if self.training == self.layers_are_rescaled and (
|
484 |
+
self.embeddings.weight.dtype == torch.float16 or self.embeddings.weight.dtype == torch.bfloat16
|
485 |
+
):
|
486 |
self._rescale_layers()
|
487 |
|
488 |
if input_ids is not None and inputs_embeds is not None:
|
|
|
492 |
|
493 |
if inputs_embeds is None:
|
494 |
if not self.pre_ln_flag:
|
495 |
+
normalized_weight = F.layer_norm(
|
496 |
+
self.embeddings.weight,
|
497 |
+
(self.config.hidden_size,),
|
498 |
+
weight=self.blocks[0].pre_ln.weight,
|
499 |
+
bias=self.blocks[0].pre_ln.bias,
|
500 |
+
)
|
501 |
self.embeddings.weight = nn.Parameter(normalized_weight)
|
502 |
self.pre_ln_flag = True
|
503 |
+
|
504 |
inputs_embeds = self.embeddings(input_ids)
|
505 |
|
506 |
if use_cache and state is None:
|
507 |
# https://github.com/BlinkDL/ChatRWKV/blob/main/rwkv_pip_package/src/rwkv/model.py#L904-L906
|
508 |
state = []
|
509 |
+
num_attention_heads = self.config.hidden_size // self.config.num_attention_heads
|
510 |
+
state.append(
|
511 |
+
torch.zeros(
|
512 |
+
(inputs_embeds.size(0), self.config.hidden_size, self.config.num_hidden_layers),
|
513 |
+
dtype=inputs_embeds.dtype,
|
514 |
+
requires_grad=False,
|
515 |
+
device=inputs_embeds.device,
|
516 |
+
).contiguous()
|
517 |
+
)
|
518 |
+
state.append(
|
519 |
+
torch.zeros(
|
520 |
+
(
|
521 |
+
inputs_embeds.size(0),
|
522 |
+
num_attention_heads,
|
523 |
+
self.config.hidden_size // num_attention_heads,
|
524 |
+
self.config.hidden_size // num_attention_heads,
|
525 |
+
self.config.num_hidden_layers,
|
526 |
+
),
|
527 |
+
dtype=torch.float32,
|
528 |
+
requires_grad=False,
|
529 |
+
device=inputs_embeds.device,
|
530 |
+
).contiguous()
|
531 |
+
)
|
532 |
+
state.append(
|
533 |
+
torch.zeros(
|
534 |
+
(inputs_embeds.size(0), self.config.hidden_size, self.config.num_hidden_layers),
|
535 |
+
dtype=inputs_embeds.dtype,
|
536 |
+
requires_grad=False,
|
537 |
+
device=inputs_embeds.device,
|
538 |
+
).contiguous()
|
539 |
+
)
|
540 |
|
541 |
+
seq_mode = inputs_embeds.shape[1] > 1
|
542 |
hidden_states = inputs_embeds
|
543 |
+
|
|
|
544 |
all_self_attentions = () if output_attentions else None
|
545 |
all_hidden_states = () if output_hidden_states else None
|
546 |
for idx, block in enumerate(self.blocks):
|
|
|
568 |
if not return_dict:
|
569 |
return (hidden_states, state, all_hidden_states, all_self_attentions)
|
570 |
|
571 |
+
return Rwkv5Output(
|
572 |
last_hidden_state=hidden_states,
|
573 |
state=state,
|
574 |
+
hidden_states=all_hidden_states, # None
|
575 |
+
attentions=all_self_attentions, # None
|
576 |
)
|
577 |
|
578 |
def _rescale_layers(self):
|
|
|
591 |
|
592 |
self.layers_are_rescaled = not self.training
|
593 |
|
594 |
+
|
595 |
@add_start_docstrings(
|
596 |
"""
|
597 |
The RWKV Model transformer with a language modeling head on top (linear layer with weights tied to the input
|
|
|
599 |
""",
|
600 |
RWKV_START_DOCSTRING,
|
601 |
)
|
602 |
+
class Rwkv5ForCausalLM(Rwkv5PreTrainedModel):
|
603 |
+
_tied_weights_keys = ["head.weight"]
|
604 |
+
|
605 |
def __init__(self, config):
|
606 |
super().__init__(config)
|
607 |
+
self.rwkv = Rwkv5Model(config)
|
608 |
self.head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
609 |
|
610 |
# Initialize weights and apply final processing
|
|
|
633 |
@add_start_docstrings_to_model_forward(RWKV_INPUTS_DOCSTRING)
|
634 |
@add_code_sample_docstrings(
|
635 |
checkpoint=_CHECKPOINT_FOR_DOC,
|
636 |
+
output_type=Rwkv5CausalLMOutput,
|
637 |
config_class=_CONFIG_FOR_DOC,
|
638 |
)
|
639 |
def forward(
|
|
|
647 |
output_attentions: Optional[bool] = None,
|
648 |
output_hidden_states: Optional[bool] = None,
|
649 |
return_dict: Optional[bool] = None,
|
650 |
+
) -> Union[Tuple, Rwkv5CausalLMOutput]:
|
651 |
r"""
|
652 |
labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
|
653 |
Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
|
|
|
665 |
output_hidden_states=output_hidden_states,
|
666 |
return_dict=return_dict,
|
667 |
)
|
668 |
+
hidden_states = rwkv_outputs[0]
|
|
|
669 |
|
670 |
+
logits = self.head(hidden_states)
|
671 |
|
672 |
loss = None
|
673 |
if labels is not None:
|
674 |
+
# https://github.com/BlinkDL/ChatRWKV/blob/main/rwkv_pip_package/src/rwkv/model.py#L984
|
675 |
+
loss = torch.tensor(0.0, device=logits.device, dtype=logits.dtype)
|
|
|
|
|
|
|
|
|
|
|
|
|
676 |
|
677 |
if not return_dict:
|
678 |
output = (logits,) + rwkv_outputs[1:]
|
679 |
return ((loss,) + output) if loss is not None else output
|
680 |
|
681 |
+
return Rwkv5CausalLMOutput(
|
682 |
loss=loss,
|
683 |
logits=logits,
|
684 |
state=rwkv_outputs.state,
|
685 |
+
hidden_states=rwkv_outputs.hidden_states,
|
686 |
attentions=rwkv_outputs.attentions,
|
687 |
)
|
tokenization_rwkv_world.py
CHANGED
@@ -12,38 +12,20 @@
|
|
12 |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
# See the License for the specific language governing permissions and
|
14 |
# limitations under the License.
|
15 |
-
"""Tokenization classes for
|
16 |
|
17 |
import json
|
18 |
import os
|
19 |
from typing import TYPE_CHECKING, List, Optional, Tuple, Union
|
20 |
-
from transformers.tokenization_utils import AddedToken, PreTrainedTokenizer
|
21 |
-
from transformers.utils import logging, to_py_obj
|
22 |
-
from transformers.tokenization_utils_base import BatchEncoding
|
23 |
-
|
24 |
-
import bisect
|
25 |
-
import itertools
|
26 |
-
import re
|
27 |
-
import unicodedata
|
28 |
-
from collections import OrderedDict
|
29 |
-
from typing import Any, Dict, List, Optional, Tuple, Union, overload
|
30 |
|
|
|
31 |
from transformers.tokenization_utils_base import (
|
32 |
-
ENCODE_KWARGS_DOCSTRING,
|
33 |
-
ENCODE_PLUS_ADDITIONAL_KWARGS_DOCSTRING,
|
34 |
-
INIT_TOKENIZER_DOCSTRING,
|
35 |
-
AddedToken,
|
36 |
BatchEncoding,
|
37 |
EncodedInput,
|
38 |
-
EncodedInputPair,
|
39 |
-
PreTokenizedInput,
|
40 |
-
PreTokenizedInputPair,
|
41 |
-
PreTrainedTokenizerBase,
|
42 |
TextInput,
|
43 |
-
TextInputPair,
|
44 |
TruncationStrategy,
|
45 |
)
|
46 |
-
from transformers.utils import PaddingStrategy, TensorType,
|
47 |
|
48 |
|
49 |
if TYPE_CHECKING:
|
@@ -54,11 +36,18 @@ logger = logging.get_logger(__name__)
|
|
54 |
VOCAB_FILES_NAMES = {
|
55 |
"vocab_file": "rwkv_vocab_v20230424.txt",
|
56 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
57 |
|
58 |
class TRIE:
|
59 |
__slots__ = tuple("ch,to,values,front".split(","))
|
60 |
-
to:list
|
61 |
-
values:set
|
|
|
62 |
def __init__(self, front=None, ch=None):
|
63 |
self.ch = ch
|
64 |
self.to = [None for ch in range(256)]
|
@@ -68,64 +57,59 @@ class TRIE:
|
|
68 |
def __repr__(self):
|
69 |
fr = self
|
70 |
ret = []
|
71 |
-
while
|
72 |
-
if
|
73 |
ret.append(fr.ch)
|
74 |
fr = fr.front
|
75 |
-
return "<TRIE %s %s>"%(ret[::-1], self.values)
|
76 |
-
|
77 |
-
def add(self, key:bytes, idx:int=0, val=None):
|
78 |
-
if
|
79 |
-
if
|
80 |
val = key
|
81 |
self.values.add(val)
|
82 |
return self
|
83 |
ch = key[idx]
|
84 |
-
if
|
85 |
self.to[ch] = TRIE(front=self, ch=ch)
|
86 |
-
return self.to[ch].add(key, idx=idx+1, val=val)
|
87 |
-
|
88 |
-
def find_longest(self, key:bytes, idx:int=0):
|
89 |
-
u:TRIE = self
|
90 |
-
ch:int = key[idx]
|
91 |
-
|
92 |
-
while
|
93 |
u = u.to[ch]
|
94 |
idx += 1
|
95 |
-
if
|
96 |
ret = idx, u, u.values
|
97 |
-
if
|
98 |
break
|
99 |
ch = key[idx]
|
100 |
return ret
|
101 |
|
|
|
102 |
class RWKVWorldTokenizer(PreTrainedTokenizer):
|
103 |
vocab_files_names = VOCAB_FILES_NAMES
|
104 |
model_input_names = ["input_ids", "attention_mask"]
|
105 |
|
106 |
-
def __init__(
|
107 |
-
self,
|
108 |
-
vocab_file,
|
109 |
-
errors="replace",
|
110 |
-
pad_token="0",
|
111 |
-
**kwargs
|
112 |
-
):
|
113 |
self.add_bos_token = False
|
114 |
self.encoder = {}
|
115 |
-
sorted = []
|
116 |
with open(vocab_file, "r", encoding="utf-8") as f:
|
117 |
lines = f.readlines()
|
118 |
for l in lines:
|
119 |
-
idx = int(l[:l.index(
|
120 |
-
x = eval(l[l.index(
|
121 |
x = x.encode("utf-8") if isinstance(x, str) else x
|
122 |
assert isinstance(x, bytes)
|
123 |
-
assert len(x) == int(l[l.rindex(
|
124 |
sorted += [x]
|
125 |
self.encoder[idx] = x
|
126 |
-
|
127 |
self.decoder = {}
|
128 |
-
for k,v in self.encoder.items():
|
129 |
self.decoder[v] = int(k)
|
130 |
|
131 |
self.trie = TRIE()
|
@@ -134,13 +118,18 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
134 |
self.errors = errors # how to handle errors in decoding
|
135 |
self.cache = {}
|
136 |
self.first_max_length = 0
|
137 |
-
|
138 |
-
# pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
|
139 |
super().__init__(
|
140 |
errors=errors,
|
141 |
-
# pad_token=pad_token,
|
142 |
**kwargs,
|
143 |
)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
|
145 |
@property
|
146 |
def vocab_size(self):
|
@@ -148,12 +137,12 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
148 |
|
149 |
def get_vocab(self):
|
150 |
return dict(self.encoder, **self.added_tokens_encoder)
|
151 |
-
|
152 |
def add_tokens(self, new_tokens, special_tokens: bool = False):
|
153 |
for token in new_tokens:
|
154 |
token_id = self.convert_tokens_to_ids(token)
|
155 |
self.added_tokens_decoder[token_id] = token
|
156 |
-
|
157 |
def convert_ids_to_tokens(self, ids, skip_special_tokens=False):
|
158 |
if isinstance(ids, int):
|
159 |
ids = [ids]
|
@@ -179,8 +168,7 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
179 |
return output + bos_token_ids + token_ids_1
|
180 |
|
181 |
def get_special_tokens_mask(
|
182 |
-
|
183 |
-
already_has_special_tokens: bool = False
|
184 |
) -> List[int]:
|
185 |
"""
|
186 |
Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
|
@@ -211,19 +199,19 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
211 |
return [1] + ([0] * len(token_ids_0))
|
212 |
return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1))
|
213 |
|
214 |
-
def encodeBytes(self, src:bytes):
|
215 |
-
idx:int = 0
|
216 |
tokens = []
|
217 |
-
while
|
218 |
-
_idx:int = idx
|
219 |
idx, _, values = self.trie.find_longest(src, idx)
|
220 |
-
assert
|
221 |
-
_, token = next(iter(values))
|
222 |
tokens.append(token)
|
223 |
return tokens
|
224 |
-
|
225 |
def decodeBytes(self, tokens):
|
226 |
-
return b''.join(map(lambda i: self.encoder[i], tokens))
|
227 |
|
228 |
def _tokenize(self, text, **kwargs):
|
229 |
"""Tokenize a string."""
|
@@ -231,21 +219,21 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
231 |
|
232 |
def _decode_tokens(self, tokens):
|
233 |
try:
|
234 |
-
return self.decodeBytes(tokens).decode(
|
235 |
-
except:
|
236 |
-
return
|
237 |
-
|
238 |
-
def _decode(
|
239 |
-
|
240 |
-
|
241 |
-
|
242 |
-
|
243 |
-
|
244 |
def remove_zeros_from_first_segment(token_ids, first_max_length):
|
245 |
first_segment = token_ids[:first_max_length]
|
246 |
first_segment_cleaned = [token for token in first_segment if token != 0]
|
247 |
return first_segment_cleaned + token_ids[first_max_length:]
|
248 |
-
|
249 |
# Convert inputs to python lists
|
250 |
token_ids = to_py_obj(token_ids)
|
251 |
token_ids = remove_zeros_from_first_segment(token_ids, self.first_max_length)
|
@@ -263,7 +251,7 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
263 |
break
|
264 |
out_tokens += [token]
|
265 |
tmp = self._decode_tokens(out_tokens[out_last:])
|
266 |
-
if
|
267 |
out_str += tmp
|
268 |
out_last = i + 1
|
269 |
return out_str
|
@@ -318,16 +306,29 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
318 |
return_offsets_mapping: bool = False,
|
319 |
return_length: bool = False,
|
320 |
verbose: bool = True,
|
321 |
-
**kwargs
|
322 |
) -> BatchEncoding:
|
323 |
-
def get_input_ids(text):
|
|
|
|
|
|
|
324 |
if isinstance(text, str):
|
325 |
-
|
326 |
-
|
|
|
|
|
|
|
327 |
elif isinstance(text, list) and len(text) > 0 and isinstance(text[0], str):
|
328 |
-
|
|
|
|
|
|
|
|
|
329 |
elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], int):
|
|
|
|
|
330 |
return text
|
|
|
331 |
else:
|
332 |
raise ValueError(
|
333 |
"Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
|
@@ -383,7 +384,7 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
383 |
return_offsets_mapping: bool = False,
|
384 |
return_length: bool = False,
|
385 |
verbose: bool = True,
|
386 |
-
**kwargs
|
387 |
) -> BatchEncoding:
|
388 |
def get_input_ids(text, max_length=None, pad_token_id=0):
|
389 |
def pad_sequence(seq, max_len, pad_tok):
|
@@ -411,7 +412,6 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
411 |
"Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
|
412 |
)
|
413 |
|
414 |
-
|
415 |
if return_offsets_mapping:
|
416 |
raise NotImplementedError(
|
417 |
"return_offset_mapping is not available when using Python tokenizers. "
|
@@ -462,10 +462,10 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
462 |
)
|
463 |
|
464 |
return BatchEncoding(batch_outputs)
|
465 |
-
|
466 |
def decode(
|
467 |
self,
|
468 |
-
token_ids: Union[int, List[int]
|
469 |
skip_special_tokens: bool = False,
|
470 |
clean_up_tokenization_spaces: bool = None,
|
471 |
**kwargs,
|
@@ -500,7 +500,7 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
500 |
|
501 |
def batch_decode(
|
502 |
self,
|
503 |
-
sequences: Union[List[int], List[List[int]]
|
504 |
skip_special_tokens: bool = False,
|
505 |
clean_up_tokenization_spaces: bool = None,
|
506 |
**kwargs,
|
@@ -537,5 +537,5 @@ class RWKVWorldTokenizer(PreTrainedTokenizer):
|
|
537 |
for is_user, text in conversation.iter_texts():
|
538 |
input_ids.extend(self.encode(text, add_special_tokens=False) + [self.eos_token_id])
|
539 |
if len(input_ids) > self.model_max_length:
|
540 |
-
input_ids = input_ids[-self.model_max_length:]
|
541 |
return input_ids
|
|
|
12 |
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
13 |
# See the License for the specific language governing permissions and
|
14 |
# limitations under the License.
|
15 |
+
"""Tokenization classes for RWKV5."""
|
16 |
|
17 |
import json
|
18 |
import os
|
19 |
from typing import TYPE_CHECKING, List, Optional, Tuple, Union
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
20 |
|
21 |
+
from transformers.tokenization_utils import PreTrainedTokenizer
|
22 |
from transformers.tokenization_utils_base import (
|
|
|
|
|
|
|
|
|
23 |
BatchEncoding,
|
24 |
EncodedInput,
|
|
|
|
|
|
|
|
|
25 |
TextInput,
|
|
|
26 |
TruncationStrategy,
|
27 |
)
|
28 |
+
from transformers.utils import PaddingStrategy, TensorType, logging, to_py_obj
|
29 |
|
30 |
|
31 |
if TYPE_CHECKING:
|
|
|
36 |
VOCAB_FILES_NAMES = {
|
37 |
"vocab_file": "rwkv_vocab_v20230424.txt",
|
38 |
}
|
39 |
+
PRETRAINED_VOCAB_FILES_MAP = {
|
40 |
+
"vocab_file": {
|
41 |
+
"RWKV/rwkv-5-world-169m": "https://huggingface.co/RWKV/rwkv-5-world-169m/blob/main/rwkv_vocab_v20230424.txt",
|
42 |
+
},
|
43 |
+
}
|
44 |
+
|
45 |
|
46 |
class TRIE:
|
47 |
__slots__ = tuple("ch,to,values,front".split(","))
|
48 |
+
to: list
|
49 |
+
values: set
|
50 |
+
|
51 |
def __init__(self, front=None, ch=None):
|
52 |
self.ch = ch
|
53 |
self.to = [None for ch in range(256)]
|
|
|
57 |
def __repr__(self):
|
58 |
fr = self
|
59 |
ret = []
|
60 |
+
while fr is not None:
|
61 |
+
if fr.ch is not None:
|
62 |
ret.append(fr.ch)
|
63 |
fr = fr.front
|
64 |
+
return "<TRIE %s %s>" % (ret[::-1], self.values)
|
65 |
+
|
66 |
+
def add(self, key: bytes, idx: int = 0, val=None):
|
67 |
+
if idx == len(key):
|
68 |
+
if val is None:
|
69 |
val = key
|
70 |
self.values.add(val)
|
71 |
return self
|
72 |
ch = key[idx]
|
73 |
+
if self.to[ch] is None:
|
74 |
self.to[ch] = TRIE(front=self, ch=ch)
|
75 |
+
return self.to[ch].add(key, idx=idx + 1, val=val)
|
76 |
+
|
77 |
+
def find_longest(self, key: bytes, idx: int = 0):
|
78 |
+
u: TRIE = self
|
79 |
+
ch: int = key[idx]
|
80 |
+
|
81 |
+
while u.to[ch] is not None:
|
82 |
u = u.to[ch]
|
83 |
idx += 1
|
84 |
+
if u.values:
|
85 |
ret = idx, u, u.values
|
86 |
+
if idx == len(key):
|
87 |
break
|
88 |
ch = key[idx]
|
89 |
return ret
|
90 |
|
91 |
+
|
92 |
class RWKVWorldTokenizer(PreTrainedTokenizer):
|
93 |
vocab_files_names = VOCAB_FILES_NAMES
|
94 |
model_input_names = ["input_ids", "attention_mask"]
|
95 |
|
96 |
+
def __init__(self, vocab_file, errors="replace", pad_token="0", **kwargs):
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
self.add_bos_token = False
|
98 |
self.encoder = {}
|
99 |
+
sorted = [] # must be already sorted
|
100 |
with open(vocab_file, "r", encoding="utf-8") as f:
|
101 |
lines = f.readlines()
|
102 |
for l in lines:
|
103 |
+
idx = int(l[: l.index(" ")])
|
104 |
+
x = eval(l[l.index(" ") : l.rindex(" ")])
|
105 |
x = x.encode("utf-8") if isinstance(x, str) else x
|
106 |
assert isinstance(x, bytes)
|
107 |
+
assert len(x) == int(l[l.rindex(" ") :])
|
108 |
sorted += [x]
|
109 |
self.encoder[idx] = x
|
110 |
+
|
111 |
self.decoder = {}
|
112 |
+
for k, v in self.encoder.items():
|
113 |
self.decoder[v] = int(k)
|
114 |
|
115 |
self.trie = TRIE()
|
|
|
118 |
self.errors = errors # how to handle errors in decoding
|
119 |
self.cache = {}
|
120 |
self.first_max_length = 0
|
|
|
|
|
121 |
super().__init__(
|
122 |
errors=errors,
|
|
|
123 |
**kwargs,
|
124 |
)
|
125 |
+
|
126 |
+
@property
|
127 |
+
def eos_token_id(self) -> Optional[int]:
|
128 |
+
return 0
|
129 |
+
|
130 |
+
@property
|
131 |
+
def eot_token_id(self) -> Optional[int]:
|
132 |
+
return 0
|
133 |
|
134 |
@property
|
135 |
def vocab_size(self):
|
|
|
137 |
|
138 |
def get_vocab(self):
|
139 |
return dict(self.encoder, **self.added_tokens_encoder)
|
140 |
+
|
141 |
def add_tokens(self, new_tokens, special_tokens: bool = False):
|
142 |
for token in new_tokens:
|
143 |
token_id = self.convert_tokens_to_ids(token)
|
144 |
self.added_tokens_decoder[token_id] = token
|
145 |
+
|
146 |
def convert_ids_to_tokens(self, ids, skip_special_tokens=False):
|
147 |
if isinstance(ids, int):
|
148 |
ids = [ids]
|
|
|
168 |
return output + bos_token_ids + token_ids_1
|
169 |
|
170 |
def get_special_tokens_mask(
|
171 |
+
self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
|
|
|
172 |
) -> List[int]:
|
173 |
"""
|
174 |
Retrieves sequence ids from a token list that has no special tokens added. This method is called when adding
|
|
|
199 |
return [1] + ([0] * len(token_ids_0))
|
200 |
return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1))
|
201 |
|
202 |
+
def encodeBytes(self, src: bytes):
|
203 |
+
idx: int = 0
|
204 |
tokens = []
|
205 |
+
while idx < len(src):
|
206 |
+
_idx: int = idx
|
207 |
idx, _, values = self.trie.find_longest(src, idx)
|
208 |
+
assert idx != _idx
|
209 |
+
_, token = next(iter(values))
|
210 |
tokens.append(token)
|
211 |
return tokens
|
212 |
+
|
213 |
def decodeBytes(self, tokens):
|
214 |
+
return b''.join(map(lambda i: self.encoder[i], tokens)) # noqa
|
215 |
|
216 |
def _tokenize(self, text, **kwargs):
|
217 |
"""Tokenize a string."""
|
|
|
219 |
|
220 |
def _decode_tokens(self, tokens):
|
221 |
try:
|
222 |
+
return self.decodeBytes(tokens).decode("utf-8")
|
223 |
+
except Exception:
|
224 |
+
return "\ufffd" # bad utf-8
|
225 |
+
|
226 |
+
def _decode(
|
227 |
+
self,
|
228 |
+
token_ids: Union[int, List[int]],
|
229 |
+
skip_special_tokens: bool = False,
|
230 |
+
**kwargs,
|
231 |
+
) -> str:
|
232 |
def remove_zeros_from_first_segment(token_ids, first_max_length):
|
233 |
first_segment = token_ids[:first_max_length]
|
234 |
first_segment_cleaned = [token for token in first_segment if token != 0]
|
235 |
return first_segment_cleaned + token_ids[first_max_length:]
|
236 |
+
|
237 |
# Convert inputs to python lists
|
238 |
token_ids = to_py_obj(token_ids)
|
239 |
token_ids = remove_zeros_from_first_segment(token_ids, self.first_max_length)
|
|
|
251 |
break
|
252 |
out_tokens += [token]
|
253 |
tmp = self._decode_tokens(out_tokens[out_last:])
|
254 |
+
if "\ufffd" not in tmp:
|
255 |
out_str += tmp
|
256 |
out_last = i + 1
|
257 |
return out_str
|
|
|
306 |
return_offsets_mapping: bool = False,
|
307 |
return_length: bool = False,
|
308 |
verbose: bool = True,
|
309 |
+
**kwargs,
|
310 |
) -> BatchEncoding:
|
311 |
+
def get_input_ids(text, max_length=None, pad_token_id=0):
|
312 |
+
def pad_sequence(seq, max_len, pad_tok):
|
313 |
+
return [pad_tok] * (max_len - len(seq)) + seq
|
314 |
+
|
315 |
if isinstance(text, str):
|
316 |
+
tokens = self._tokenize(text)
|
317 |
+
if max_length is not None:
|
318 |
+
tokens = pad_sequence(tokens, max_length, pad_token_id)
|
319 |
+
return tokens
|
320 |
+
|
321 |
elif isinstance(text, list) and len(text) > 0 and isinstance(text[0], str):
|
322 |
+
tokenized_texts = [self._tokenize(t) for t in text]
|
323 |
+
if max_length is None:
|
324 |
+
max_length = max(len(t) for t in tokenized_texts)
|
325 |
+
return [pad_sequence(t, max_length, pad_token_id) for t in tokenized_texts]
|
326 |
+
|
327 |
elif isinstance(text, (list, tuple)) and len(text) > 0 and isinstance(text[0], int):
|
328 |
+
if max_length is not None and len(text) < max_length:
|
329 |
+
return pad_sequence(text, max_length, pad_token_id)
|
330 |
return text
|
331 |
+
|
332 |
else:
|
333 |
raise ValueError(
|
334 |
"Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
|
|
|
384 |
return_offsets_mapping: bool = False,
|
385 |
return_length: bool = False,
|
386 |
verbose: bool = True,
|
387 |
+
**kwargs,
|
388 |
) -> BatchEncoding:
|
389 |
def get_input_ids(text, max_length=None, pad_token_id=0):
|
390 |
def pad_sequence(seq, max_len, pad_tok):
|
|
|
412 |
"Input is not valid. Should be a string, a list/tuple of strings or a list/tuple of integers."
|
413 |
)
|
414 |
|
|
|
415 |
if return_offsets_mapping:
|
416 |
raise NotImplementedError(
|
417 |
"return_offset_mapping is not available when using Python tokenizers. "
|
|
|
462 |
)
|
463 |
|
464 |
return BatchEncoding(batch_outputs)
|
465 |
+
|
466 |
def decode(
|
467 |
self,
|
468 |
+
token_ids: Union[int, List[int]],
|
469 |
skip_special_tokens: bool = False,
|
470 |
clean_up_tokenization_spaces: bool = None,
|
471 |
**kwargs,
|
|
|
500 |
|
501 |
def batch_decode(
|
502 |
self,
|
503 |
+
sequences: Union[List[int], List[List[int]]],
|
504 |
skip_special_tokens: bool = False,
|
505 |
clean_up_tokenization_spaces: bool = None,
|
506 |
**kwargs,
|
|
|
537 |
for is_user, text in conversation.iter_texts():
|
538 |
input_ids.extend(self.encode(text, add_special_tokens=False) + [self.eos_token_id])
|
539 |
if len(input_ids) > self.model_max_length:
|
540 |
+
input_ids = input_ids[-self.model_max_length :]
|
541 |
return input_ids
|