add files
Browse files- config.json +24 -0
- dnaflash.py +414 -0
- special_tokens_map.json +37 -0
- tokenizer.json +0 -0
- tokenizer_config.json +52 -0
config.json
ADDED
@@ -0,0 +1,24 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"architectures": [
|
3 |
+
"FLASHTransformerForPretrained"
|
4 |
+
],
|
5 |
+
"auto_map": {
|
6 |
+
"AutoConfig": "dnaflash.FLASHTransformerConfig",
|
7 |
+
"AutoModel": "dnaflash.FLASHTransformerForPretrained"
|
8 |
+
},
|
9 |
+
"attn_dropout": 0.0,
|
10 |
+
"causal": false,
|
11 |
+
"expansion_factor": 2.0,
|
12 |
+
"group_size": 256,
|
13 |
+
"hidden_size": 1024,
|
14 |
+
"laplace_attn_fn": false,
|
15 |
+
"model_type": "flash_transformer",
|
16 |
+
"norm_type": "scalenorm",
|
17 |
+
"num_layers": 36,
|
18 |
+
"query_key_dim": 128,
|
19 |
+
"reduce_group_non_causal_attn": true,
|
20 |
+
"shift_tokens": true,
|
21 |
+
"torch_dtype": "float32",
|
22 |
+
"transformers_version": "4.39.3",
|
23 |
+
"vocab_size": 4096
|
24 |
+
}
|
dnaflash.py
ADDED
@@ -0,0 +1,414 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
import torch
|
3 |
+
import torch.nn.functional as F
|
4 |
+
from torch import nn, einsum
|
5 |
+
|
6 |
+
from einops import rearrange
|
7 |
+
from rotary_embedding_torch import RotaryEmbedding
|
8 |
+
|
9 |
+
from transformers import PreTrainedModel, PretrainedConfig
|
10 |
+
from transformers.modeling_outputs import MaskedLMOutput
|
11 |
+
|
12 |
+
# helper functions
|
13 |
+
|
14 |
+
def exists(val):
|
15 |
+
return val is not None
|
16 |
+
|
17 |
+
def default(val, d):
|
18 |
+
return val if exists(val) else d
|
19 |
+
|
20 |
+
def padding_to_multiple_of(n, mult):
|
21 |
+
remainder = n % mult
|
22 |
+
if remainder == 0:
|
23 |
+
return 0
|
24 |
+
return mult - remainder
|
25 |
+
|
26 |
+
# scalenorm
|
27 |
+
|
28 |
+
class ScaleNorm(nn.Module):
|
29 |
+
def __init__(self, dim, eps = 1e-5):
|
30 |
+
super().__init__()
|
31 |
+
self.scale = dim ** -0.5
|
32 |
+
self.eps = eps
|
33 |
+
self.g = nn.Parameter(torch.ones(1))
|
34 |
+
|
35 |
+
def forward(self, x):
|
36 |
+
norm = torch.norm(x, dim = -1, keepdim = True) * self.scale
|
37 |
+
return x / norm.clamp(min = self.eps) * self.g
|
38 |
+
|
39 |
+
# absolute positional encodings
|
40 |
+
|
41 |
+
class ScaledSinuEmbedding(nn.Module):
|
42 |
+
def __init__(self, dim):
|
43 |
+
super().__init__()
|
44 |
+
self.scale = nn.Parameter(torch.ones(1,))
|
45 |
+
inv_freq = 1. / (10000 ** (torch.arange(0, dim, 2).float() / dim))
|
46 |
+
self.register_buffer('inv_freq', inv_freq)
|
47 |
+
|
48 |
+
def forward(self, x):
|
49 |
+
n, device = x.shape[1], x.device
|
50 |
+
t = torch.arange(n, device = device).type_as(self.inv_freq)
|
51 |
+
sinu = einsum('i , j -> i j', t, self.inv_freq)
|
52 |
+
emb = torch.cat((sinu.sin(), sinu.cos()), dim = -1)
|
53 |
+
return emb * self.scale
|
54 |
+
|
55 |
+
# T5 relative positional bias
|
56 |
+
|
57 |
+
class T5RelativePositionBias(nn.Module):
|
58 |
+
def __init__(
|
59 |
+
self,
|
60 |
+
scale,
|
61 |
+
causal = False,
|
62 |
+
num_buckets = 32,
|
63 |
+
max_distance = 128
|
64 |
+
):
|
65 |
+
super().__init__()
|
66 |
+
self.scale = scale
|
67 |
+
self.causal = causal
|
68 |
+
self.num_buckets = num_buckets
|
69 |
+
self.max_distance = max_distance
|
70 |
+
self.relative_attention_bias = nn.Embedding(num_buckets, 1)
|
71 |
+
|
72 |
+
@staticmethod
|
73 |
+
def _relative_position_bucket(
|
74 |
+
relative_position,
|
75 |
+
causal = True,
|
76 |
+
num_buckets = 32,
|
77 |
+
max_distance = 128
|
78 |
+
):
|
79 |
+
ret = 0
|
80 |
+
n = -relative_position
|
81 |
+
if not causal:
|
82 |
+
num_buckets //= 2
|
83 |
+
ret += (n < 0).long() * num_buckets
|
84 |
+
n = torch.abs(n)
|
85 |
+
else:
|
86 |
+
n = torch.max(n, torch.zeros_like(n))
|
87 |
+
|
88 |
+
max_exact = num_buckets // 2
|
89 |
+
is_small = n < max_exact
|
90 |
+
|
91 |
+
val_if_large = max_exact + (
|
92 |
+
torch.log(n.float() / max_exact) / math.log(max_distance / max_exact) * (num_buckets - max_exact)
|
93 |
+
).long()
|
94 |
+
val_if_large = torch.min(val_if_large, torch.full_like(val_if_large, num_buckets - 1))
|
95 |
+
|
96 |
+
ret += torch.where(is_small, n, val_if_large)
|
97 |
+
return ret
|
98 |
+
|
99 |
+
def forward(self, x):
|
100 |
+
i, j, device = *x.shape[-2:], x.device
|
101 |
+
q_pos = torch.arange(i, dtype = torch.long, device = device)
|
102 |
+
k_pos = torch.arange(j, dtype = torch.long, device = device)
|
103 |
+
rel_pos = rearrange(k_pos, 'j -> 1 j') - rearrange(q_pos, 'i -> i 1')
|
104 |
+
rp_bucket = self._relative_position_bucket(rel_pos, causal = self.causal, num_buckets = self.num_buckets, max_distance = self.max_distance)
|
105 |
+
values = self.relative_attention_bias(rp_bucket)
|
106 |
+
bias = rearrange(values, 'i j 1 -> i j')
|
107 |
+
return bias * self.scale
|
108 |
+
|
109 |
+
# class
|
110 |
+
|
111 |
+
class OffsetScale(nn.Module):
|
112 |
+
def __init__(self, dim, heads = 1):
|
113 |
+
super().__init__()
|
114 |
+
self.weight = nn.Parameter(torch.ones(heads, dim))
|
115 |
+
self.bias = nn.Parameter(torch.zeros(heads, dim))
|
116 |
+
nn.init.normal_(self.weight, std = 0.02)
|
117 |
+
|
118 |
+
def forward(self, x):
|
119 |
+
out = einsum('... d, h d -> ... h d', x, self.weight) + self.bias
|
120 |
+
return out.unbind(dim = -2)
|
121 |
+
|
122 |
+
# activation functions
|
123 |
+
|
124 |
+
class ReLUSquared(nn.Module):
|
125 |
+
def forward(self, x):
|
126 |
+
return F.relu(x) ** 2
|
127 |
+
|
128 |
+
class LaplacianAttnFn(nn.Module):
|
129 |
+
""" https://arxiv.org/abs/2209.10655 claims this is more stable than Relu squared """
|
130 |
+
|
131 |
+
def forward(self, x):
|
132 |
+
mu = math.sqrt(0.5)
|
133 |
+
std = math.sqrt((4 * math.pi) ** -1)
|
134 |
+
return (1 + torch.special.erf((x - mu) / (std * math.sqrt(2)))) * 0.5
|
135 |
+
|
136 |
+
|
137 |
+
class FLASH(nn.Module):
|
138 |
+
def __init__(
|
139 |
+
self,
|
140 |
+
*,
|
141 |
+
dim,
|
142 |
+
group_size = 256,
|
143 |
+
query_key_dim = 128,
|
144 |
+
expansion_factor = 2.,
|
145 |
+
causal = False,
|
146 |
+
dropout = 0.,
|
147 |
+
rotary_pos_emb = None,
|
148 |
+
norm_klass = nn.LayerNorm,
|
149 |
+
shift_tokens = False,
|
150 |
+
laplace_attn_fn = False,
|
151 |
+
reduce_group_non_causal_attn = True
|
152 |
+
):
|
153 |
+
super().__init__()
|
154 |
+
hidden_dim = int(dim * expansion_factor)
|
155 |
+
self.group_size = group_size
|
156 |
+
self.causal = causal
|
157 |
+
self.shift_tokens = shift_tokens
|
158 |
+
|
159 |
+
self.attn_fn = ReLUSquared() if not laplace_attn_fn else LaplacianAttnFn()
|
160 |
+
|
161 |
+
# positional embeddings
|
162 |
+
|
163 |
+
self.rotary_pos_emb = rotary_pos_emb
|
164 |
+
self.rel_pos_bias = T5RelativePositionBias(query_key_dim ** 0.5, causal = causal)
|
165 |
+
|
166 |
+
# norm
|
167 |
+
|
168 |
+
self.norm = norm_klass(dim)
|
169 |
+
self.dropout = nn.Dropout(dropout)
|
170 |
+
|
171 |
+
# whether to reduce groups in non causal linear attention
|
172 |
+
|
173 |
+
self.reduce_group_non_causal_attn = reduce_group_non_causal_attn
|
174 |
+
|
175 |
+
# projections
|
176 |
+
|
177 |
+
self.to_hidden = nn.Sequential(
|
178 |
+
nn.Linear(dim, hidden_dim * 2),
|
179 |
+
nn.SiLU()
|
180 |
+
)
|
181 |
+
|
182 |
+
self.to_qk = nn.Sequential(
|
183 |
+
nn.Linear(dim, query_key_dim),
|
184 |
+
nn.SiLU()
|
185 |
+
)
|
186 |
+
|
187 |
+
self.qk_offset_scale = OffsetScale(query_key_dim, heads = 4)
|
188 |
+
self.to_out = nn.Linear(hidden_dim, dim)
|
189 |
+
|
190 |
+
def forward(
|
191 |
+
self,
|
192 |
+
x,
|
193 |
+
*,
|
194 |
+
mask = None
|
195 |
+
):
|
196 |
+
"""
|
197 |
+
b - batch
|
198 |
+
n - sequence length (within groups)
|
199 |
+
g - group dimension
|
200 |
+
d - feature dimension (keys)
|
201 |
+
e - feature dimension (values)
|
202 |
+
i - sequence dimension (source)
|
203 |
+
j - sequence dimension (target)
|
204 |
+
"""
|
205 |
+
|
206 |
+
b, n, device, g = x.shape[0], x.shape[-2], x.device, self.group_size
|
207 |
+
|
208 |
+
# prenorm
|
209 |
+
|
210 |
+
normed_x = self.norm(x)
|
211 |
+
|
212 |
+
# do token shift - a great, costless trick from an independent AI researcher in Shenzhen
|
213 |
+
|
214 |
+
if self.shift_tokens:
|
215 |
+
x_shift, x_pass = normed_x.chunk(2, dim = -1)
|
216 |
+
x_shift = F.pad(x_shift, (0, 0, 1, -1), value = 0.)
|
217 |
+
normed_x = torch.cat((x_shift, x_pass), dim = -1)
|
218 |
+
|
219 |
+
# initial projections
|
220 |
+
|
221 |
+
v, gate = self.to_hidden(normed_x).chunk(2, dim = -1)
|
222 |
+
qk = self.to_qk(normed_x)
|
223 |
+
|
224 |
+
# offset and scale
|
225 |
+
|
226 |
+
quad_q, lin_q, quad_k, lin_k = self.qk_offset_scale(qk)
|
227 |
+
|
228 |
+
# mask out linear attention keys
|
229 |
+
|
230 |
+
if exists(mask):
|
231 |
+
lin_mask = rearrange(mask, '... -> ... 1')
|
232 |
+
lin_k = lin_k.masked_fill(~lin_mask.bool(), 0.)
|
233 |
+
|
234 |
+
# rotate queries and keys
|
235 |
+
|
236 |
+
if exists(self.rotary_pos_emb):
|
237 |
+
quad_q, lin_q, quad_k, lin_k = map(self.rotary_pos_emb.rotate_queries_or_keys, (quad_q, lin_q, quad_k, lin_k))
|
238 |
+
|
239 |
+
# padding for groups
|
240 |
+
|
241 |
+
padding = padding_to_multiple_of(n, g)
|
242 |
+
|
243 |
+
if padding > 0:
|
244 |
+
quad_q, quad_k, lin_q, lin_k, v = map(lambda t: F.pad(t, (0, 0, 0, padding), value = 0.), (quad_q, quad_k, lin_q, lin_k, v))
|
245 |
+
|
246 |
+
mask = default(mask, torch.ones((b, n), device = device, dtype = torch.bool))
|
247 |
+
mask = F.pad(mask, (0, padding), value = False)
|
248 |
+
|
249 |
+
# group along sequence
|
250 |
+
|
251 |
+
quad_q, quad_k, lin_q, lin_k, v = map(lambda t: rearrange(t, 'b (n g) d -> b n g d', g = self.group_size), (quad_q, quad_k, lin_q, lin_k, v))
|
252 |
+
|
253 |
+
if exists(mask):
|
254 |
+
mask = rearrange(mask, 'b (g j) -> b g 1 j', j = g)
|
255 |
+
|
256 |
+
# calculate quadratic attention output
|
257 |
+
|
258 |
+
sim = einsum('... i d, ... j d -> ... i j', quad_q, quad_k) / g
|
259 |
+
|
260 |
+
sim = sim + self.rel_pos_bias(sim)
|
261 |
+
|
262 |
+
attn = self.attn_fn(sim)
|
263 |
+
attn = self.dropout(attn)
|
264 |
+
|
265 |
+
if exists(mask):
|
266 |
+
attn = attn.masked_fill(~mask.bool(), 0.)
|
267 |
+
|
268 |
+
if self.causal:
|
269 |
+
causal_mask = torch.ones((g, g), dtype = torch.bool, device = device).triu(1)
|
270 |
+
attn = attn.masked_fill(causal_mask.bool(), 0.)
|
271 |
+
|
272 |
+
quad_out = einsum('... i j, ... j d -> ... i d', attn, v)
|
273 |
+
|
274 |
+
# calculate linear attention output
|
275 |
+
|
276 |
+
if self.causal:
|
277 |
+
lin_kv = einsum('b g n d, b g n e -> b g d e', lin_k, v) / g
|
278 |
+
|
279 |
+
# exclusive cumulative sum along group dimension
|
280 |
+
|
281 |
+
lin_kv = lin_kv.cumsum(dim = 1)
|
282 |
+
lin_kv = F.pad(lin_kv, (0, 0, 0, 0, 1, -1), value = 0.)
|
283 |
+
|
284 |
+
lin_out = einsum('b g d e, b g n d -> b g n e', lin_kv, lin_q)
|
285 |
+
else:
|
286 |
+
context_einsum_eq = 'b d e' if self.reduce_group_non_causal_attn else 'b g d e'
|
287 |
+
lin_kv = einsum(f'b g n d, b g n e -> {context_einsum_eq}', lin_k, v) / n
|
288 |
+
lin_out = einsum(f'b g n d, {context_einsum_eq} -> b g n e', lin_q, lin_kv)
|
289 |
+
|
290 |
+
# fold back groups into full sequence, and excise out padding
|
291 |
+
|
292 |
+
quad_attn_out, lin_attn_out = map(lambda t: rearrange(t, 'b g n d -> b (g n) d')[:, :n], (quad_out, lin_out))
|
293 |
+
|
294 |
+
# gate
|
295 |
+
|
296 |
+
out = gate * (quad_attn_out + lin_attn_out)
|
297 |
+
|
298 |
+
# projection out and residual
|
299 |
+
|
300 |
+
return self.to_out(out) + x
|
301 |
+
|
302 |
+
# FLASH Transformer
|
303 |
+
|
304 |
+
class FLASHTransformer(nn.Module):
|
305 |
+
def __init__(
|
306 |
+
self,
|
307 |
+
*,
|
308 |
+
dim,
|
309 |
+
num_tokens,
|
310 |
+
depth,
|
311 |
+
group_size = 256,
|
312 |
+
query_key_dim = 128,
|
313 |
+
expansion_factor = 2.,
|
314 |
+
causal = False,
|
315 |
+
attn_dropout = 0.,
|
316 |
+
norm_type = 'scalenorm',
|
317 |
+
shift_tokens = True,
|
318 |
+
laplace_attn_fn = False,
|
319 |
+
reduce_group_non_causal_attn = True
|
320 |
+
):
|
321 |
+
super().__init__()
|
322 |
+
assert norm_type in ('scalenorm', 'layernorm'), 'norm_type must be one of scalenorm or layernorm'
|
323 |
+
|
324 |
+
if norm_type == 'scalenorm':
|
325 |
+
norm_klass = ScaleNorm
|
326 |
+
elif norm_type == 'layernorm':
|
327 |
+
norm_klass = nn.LayerNorm
|
328 |
+
|
329 |
+
self.token_emb = nn.Embedding(num_tokens, dim)
|
330 |
+
self.abs_pos_emb = ScaledSinuEmbedding(dim)
|
331 |
+
self.group_size = group_size
|
332 |
+
|
333 |
+
rotary_pos_emb = RotaryEmbedding(dim = min(32, query_key_dim))
|
334 |
+
# max rotary embedding dimensions of 32, partial Rotary embeddings, from Wang et al - GPT-J
|
335 |
+
|
336 |
+
self.layers = nn.ModuleList([FLASH(dim = dim, group_size = group_size, query_key_dim = query_key_dim, expansion_factor = expansion_factor, causal = causal, dropout = attn_dropout, rotary_pos_emb = rotary_pos_emb, norm_klass = norm_klass, shift_tokens = shift_tokens, reduce_group_non_causal_attn = reduce_group_non_causal_attn, laplace_attn_fn = laplace_attn_fn) for _ in range(depth)])
|
337 |
+
|
338 |
+
self.to_logits = nn.Sequential(
|
339 |
+
nn.LayerNorm(dim),
|
340 |
+
nn.Linear(dim, num_tokens)
|
341 |
+
)
|
342 |
+
|
343 |
+
def forward(
|
344 |
+
self,
|
345 |
+
x,
|
346 |
+
*,
|
347 |
+
mask = None
|
348 |
+
):
|
349 |
+
x = self.token_emb(x)
|
350 |
+
x = self.abs_pos_emb(x) + x
|
351 |
+
|
352 |
+
for flash in self.layers:
|
353 |
+
x = flash(x, mask = mask)
|
354 |
+
|
355 |
+
return self.to_logits(x), x
|
356 |
+
|
357 |
+
class FLASHTransformerConfig(PretrainedConfig):
|
358 |
+
model_type = "flash_transformer"
|
359 |
+
|
360 |
+
def __init__(
|
361 |
+
self,
|
362 |
+
hidden_size=512,
|
363 |
+
vocab_size=4096,
|
364 |
+
num_layers=12,
|
365 |
+
group_size=256,
|
366 |
+
query_key_dim=128,
|
367 |
+
expansion_factor=2.0,
|
368 |
+
causal=False,
|
369 |
+
attn_dropout=0.1,
|
370 |
+
norm_type="scalenorm",
|
371 |
+
shift_tokens=True,
|
372 |
+
laplace_attn_fn=False,
|
373 |
+
reduce_group_non_causal_attn=True,
|
374 |
+
**kwargs
|
375 |
+
):
|
376 |
+
super().__init__(**kwargs)
|
377 |
+
self.hidden_size = hidden_size
|
378 |
+
self.vocab_size = vocab_size
|
379 |
+
self.num_layers = num_layers
|
380 |
+
self.group_size = group_size
|
381 |
+
self.query_key_dim = query_key_dim
|
382 |
+
self.expansion_factor = expansion_factor
|
383 |
+
self.causal = causal
|
384 |
+
self.attn_dropout = attn_dropout
|
385 |
+
self.norm_type = norm_type
|
386 |
+
self.shift_tokens = shift_tokens
|
387 |
+
self.laplace_attn_fn = laplace_attn_fn
|
388 |
+
self.reduce_group_non_causal_attn = reduce_group_non_causal_attn
|
389 |
+
|
390 |
+
|
391 |
+
class FLASHTransformerForPretrained(PreTrainedModel):
|
392 |
+
config_class = FLASHTransformerConfig
|
393 |
+
base_model_prefix = "flash_transformer"
|
394 |
+
def __init__(self, config):
|
395 |
+
super().__init__(config)
|
396 |
+
self.model = FLASHTransformer(
|
397 |
+
dim=config.hidden_size,
|
398 |
+
num_tokens=config.vocab_size,
|
399 |
+
depth=config.num_layers,
|
400 |
+
group_size=config.group_size,
|
401 |
+
query_key_dim=config.query_key_dim,
|
402 |
+
expansion_factor=config.expansion_factor,
|
403 |
+
causal=config.causal,
|
404 |
+
attn_dropout=config.attn_dropout,
|
405 |
+
norm_type=config.norm_type,
|
406 |
+
shift_tokens=config.shift_tokens,
|
407 |
+
laplace_attn_fn=config.laplace_attn_fn,
|
408 |
+
reduce_group_non_causal_attn=config.reduce_group_non_causal_attn
|
409 |
+
)
|
410 |
+
|
411 |
+
def forward(self, input_ids, mask=None):
|
412 |
+
logits, x = self.model(input_ids, mask=mask)
|
413 |
+
return MaskedLMOutput(logits=logits, hidden_states=x, loss=None, attentions=None)
|
414 |
+
|
special_tokens_map.json
ADDED
@@ -0,0 +1,37 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"cls_token": {
|
3 |
+
"content": "[CLS]",
|
4 |
+
"lstrip": false,
|
5 |
+
"normalized": false,
|
6 |
+
"rstrip": false,
|
7 |
+
"single_word": false
|
8 |
+
},
|
9 |
+
"mask_token": {
|
10 |
+
"content": "[MASK]",
|
11 |
+
"lstrip": false,
|
12 |
+
"normalized": false,
|
13 |
+
"rstrip": false,
|
14 |
+
"single_word": false
|
15 |
+
},
|
16 |
+
"pad_token": {
|
17 |
+
"content": "[PAD]",
|
18 |
+
"lstrip": false,
|
19 |
+
"normalized": false,
|
20 |
+
"rstrip": false,
|
21 |
+
"single_word": false
|
22 |
+
},
|
23 |
+
"sep_token": {
|
24 |
+
"content": "[SEP]",
|
25 |
+
"lstrip": false,
|
26 |
+
"normalized": false,
|
27 |
+
"rstrip": false,
|
28 |
+
"single_word": false
|
29 |
+
},
|
30 |
+
"unk_token": {
|
31 |
+
"content": "[UNK]",
|
32 |
+
"lstrip": false,
|
33 |
+
"normalized": false,
|
34 |
+
"rstrip": false,
|
35 |
+
"single_word": false
|
36 |
+
}
|
37 |
+
}
|
tokenizer.json
ADDED
The diff for this file is too large to render.
See raw diff
|
|
tokenizer_config.json
ADDED
@@ -0,0 +1,52 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"added_tokens_decoder": {
|
3 |
+
"0": {
|
4 |
+
"content": "[UNK]",
|
5 |
+
"lstrip": false,
|
6 |
+
"normalized": false,
|
7 |
+
"rstrip": false,
|
8 |
+
"single_word": false,
|
9 |
+
"special": true
|
10 |
+
},
|
11 |
+
"1": {
|
12 |
+
"content": "[CLS]",
|
13 |
+
"lstrip": false,
|
14 |
+
"normalized": false,
|
15 |
+
"rstrip": false,
|
16 |
+
"single_word": false,
|
17 |
+
"special": true
|
18 |
+
},
|
19 |
+
"2": {
|
20 |
+
"content": "[SEP]",
|
21 |
+
"lstrip": false,
|
22 |
+
"normalized": false,
|
23 |
+
"rstrip": false,
|
24 |
+
"single_word": false,
|
25 |
+
"special": true
|
26 |
+
},
|
27 |
+
"3": {
|
28 |
+
"content": "[PAD]",
|
29 |
+
"lstrip": false,
|
30 |
+
"normalized": false,
|
31 |
+
"rstrip": false,
|
32 |
+
"single_word": false,
|
33 |
+
"special": true
|
34 |
+
},
|
35 |
+
"4": {
|
36 |
+
"content": "[MASK]",
|
37 |
+
"lstrip": false,
|
38 |
+
"normalized": false,
|
39 |
+
"rstrip": false,
|
40 |
+
"single_word": false,
|
41 |
+
"special": true
|
42 |
+
}
|
43 |
+
},
|
44 |
+
"clean_up_tokenization_spaces": true,
|
45 |
+
"cls_token": "[CLS]",
|
46 |
+
"mask_token": "[MASK]",
|
47 |
+
"model_max_length": 1000000000000000019884624838656,
|
48 |
+
"pad_token": "[PAD]",
|
49 |
+
"sep_token": "[SEP]",
|
50 |
+
"tokenizer_class": "PreTrainedTokenizerFast",
|
51 |
+
"unk_token": "[UNK]"
|
52 |
+
}
|