Upload model.py
Browse files
model.py
ADDED
@@ -0,0 +1,406 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import math
|
5 |
+
|
6 |
+
from dataclasses import dataclass
|
7 |
+
from typing import Optional, Tuple
|
8 |
+
|
9 |
+
import tiktoken
|
10 |
+
import os
|
11 |
+
import numpy as np
|
12 |
+
|
13 |
+
torch.set_float32_matmul_precision('high')
|
14 |
+
|
15 |
+
def load_tokens(filename):
|
16 |
+
npt = np.load(filename)
|
17 |
+
npt = npt.astype(np.int32) # added after video
|
18 |
+
ptt = torch.tensor(npt, dtype=torch.long)
|
19 |
+
return ptt
|
20 |
+
|
21 |
+
class DataLoaderLite:
|
22 |
+
def __init__(self, B, T, split):
|
23 |
+
self.B = B
|
24 |
+
self.T = T
|
25 |
+
assert split in {'train', 'val'}
|
26 |
+
|
27 |
+
# get the shard filenames
|
28 |
+
data_root = "tinystories"
|
29 |
+
shards = os.listdir(data_root)
|
30 |
+
shards = [s for s in shards if split in s]
|
31 |
+
shards = sorted(shards)
|
32 |
+
shards = [os.path.join(data_root, s) for s in shards]
|
33 |
+
self.shards = shards
|
34 |
+
assert len(shards) > 0, f"no shards found for split {split}"
|
35 |
+
print(f"found {len(shards)} shards for split {split}")
|
36 |
+
self.reset()
|
37 |
+
|
38 |
+
def reset(self):
|
39 |
+
# state, init at shard zero
|
40 |
+
self.current_shard = 0
|
41 |
+
self.tokens = load_tokens(self.shards[self.current_shard])
|
42 |
+
self.current_position = self.B * self.T
|
43 |
+
|
44 |
+
def next_batch(self):
|
45 |
+
B, T = self.B, self.T
|
46 |
+
buf = self.tokens[self.current_position : self.current_position+B*T+1]
|
47 |
+
x = (buf[:-1]).view(B, T) # inputs
|
48 |
+
y = (buf[1:]).view(B, T) # targets
|
49 |
+
# advance the position in the tensor
|
50 |
+
self.current_position += B * T
|
51 |
+
# if loading the next batch would be out of bounds, advance to next shard
|
52 |
+
if self.current_position + (B * T ) > len(self.tokens):
|
53 |
+
self.current_shard = (self.current_shard + 1) % len(self.shards)
|
54 |
+
self.tokens = load_tokens(self.shards[self.current_shard])
|
55 |
+
self.current_position = B * T
|
56 |
+
return x, y
|
57 |
+
|
58 |
+
@dataclass # the hyperparameters
|
59 |
+
class ModelArgs:
|
60 |
+
dim: int = 768
|
61 |
+
n_layers: int = 16
|
62 |
+
n_heads: int = 16
|
63 |
+
n_kv_heads: Optional[int] = 4
|
64 |
+
vocab_size: int = 50304
|
65 |
+
multiple_of: int = 256
|
66 |
+
ffn_dim_multiplier: Optional[float] = None
|
67 |
+
norm_eps: float = 1e-5
|
68 |
+
rope_theta: float = 50000
|
69 |
+
max_batch_size: int = 4
|
70 |
+
max_seq_len: int = 1024
|
71 |
+
device: str = 'cuda' if torch.cuda.is_available() else 'cpu'
|
72 |
+
dropout_rate: float = 0.1
|
73 |
+
|
74 |
+
params = ModelArgs()
|
75 |
+
|
76 |
+
class RMSNorm(torch.nn.Module):
|
77 |
+
def __init__(self, dim: int, eps: float = 1e-6):
|
78 |
+
super().__init__()
|
79 |
+
self.eps = eps
|
80 |
+
self.weight = nn.Parameter(torch.ones(dim))
|
81 |
+
|
82 |
+
def _norm(self, x):
|
83 |
+
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
84 |
+
|
85 |
+
def forward(self, x):
|
86 |
+
output = self._norm(x.float()).type_as(x)
|
87 |
+
return output * self.weight
|
88 |
+
|
89 |
+
|
90 |
+
def precompute_freqs_cis(dim: int, end: int, theta: float = 10000.0):
|
91 |
+
freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))
|
92 |
+
t = torch.arange(end, device=freqs.device, dtype=torch.float32)
|
93 |
+
freqs = torch.outer(t, freqs)
|
94 |
+
freqs_cis = torch.polar(torch.ones_like(freqs), freqs) # complex64
|
95 |
+
return freqs_cis.to(params.device)
|
96 |
+
|
97 |
+
def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
|
98 |
+
ndim = x.ndim
|
99 |
+
assert 0 <= 1 < ndim
|
100 |
+
assert freqs_cis.shape == (x.shape[1], x.shape[-1]), f'freqs_cis.shape {freqs_cis.shape} != (x.shape[1], x.shape[-1]) {(x.shape[1], x.shape[-1])}'
|
101 |
+
shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
|
102 |
+
return freqs_cis.view(*shape)
|
103 |
+
|
104 |
+
def apply_rotary_emb(
|
105 |
+
xq: torch.Tensor,
|
106 |
+
xk: torch.Tensor,
|
107 |
+
freqs_cis: torch.Tensor,
|
108 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
109 |
+
xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
|
110 |
+
xk_ = torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
|
111 |
+
freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
|
112 |
+
xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
|
113 |
+
xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
|
114 |
+
return xq_out.type_as(xq), xk_out.type_as(xk)
|
115 |
+
|
116 |
+
def repeat_kv(x: torch.Tensor, n_rep: int) -> torch.Tensor:
|
117 |
+
"""torch.repeat_interleave(x, dim=2, repeats=n_rep)"""
|
118 |
+
bs, seqlen, n_kv_heads, head_dim = x.shape
|
119 |
+
if n_rep == 1:
|
120 |
+
return x
|
121 |
+
return (
|
122 |
+
x[:, :, :, None, :]
|
123 |
+
.expand(bs, seqlen, n_kv_heads, n_rep, head_dim)
|
124 |
+
.reshape(bs, seqlen, n_kv_heads * n_rep, head_dim)
|
125 |
+
)
|
126 |
+
|
127 |
+
class Attention(nn.Module):
|
128 |
+
def __init__(self, args: ModelArgs):
|
129 |
+
super().__init__()
|
130 |
+
self.n_heads = args.n_heads
|
131 |
+
self.n_kv_heads = args.n_heads if args.n_kv_heads is None else args.n_kv_heads
|
132 |
+
self.n_rep = args.n_heads // self.n_kv_heads
|
133 |
+
self.head_dim = args.dim // args.n_heads
|
134 |
+
|
135 |
+
self.wq = nn.Linear(args.dim, args.n_heads * self.head_dim, bias=False)
|
136 |
+
self.wk = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False)
|
137 |
+
self.wv = nn.Linear(args.dim, self.n_kv_heads * self.head_dim, bias=False)
|
138 |
+
self.wo = nn.Linear(args.n_heads * self.head_dim, args.dim, bias=False)
|
139 |
+
|
140 |
+
self.cache_k = torch.zeros(
|
141 |
+
(args.max_batch_size, args.max_seq_len, self.n_kv_heads, self.head_dim),
|
142 |
+
requires_grad = False
|
143 |
+
).to(args.device)
|
144 |
+
self.cache_v = torch.zeros(
|
145 |
+
(args.max_batch_size, args.max_seq_len, self.n_kv_heads, self.head_dim),
|
146 |
+
requires_grad = False
|
147 |
+
).to(args.device)
|
148 |
+
|
149 |
+
def forward(
|
150 |
+
self,
|
151 |
+
x: torch.Tensor,
|
152 |
+
freqs_cis: torch.Tensor,
|
153 |
+
mask: Optional[torch.Tensor],
|
154 |
+
start_pos: int = None,
|
155 |
+
):
|
156 |
+
bsz, seqlen, _ = x.shape
|
157 |
+
xq, xk, xv = self.wq(x), self.wk(x), self.wv(x)
|
158 |
+
|
159 |
+
xq = xq.view(bsz, seqlen, self.n_heads, self.head_dim)
|
160 |
+
xk = xk.view(bsz, seqlen, self.n_kv_heads, self.head_dim)
|
161 |
+
xv = xv.view(bsz, seqlen, self.n_kv_heads, self.head_dim)
|
162 |
+
|
163 |
+
xq, xk = apply_rotary_emb(xq, xk, freqs_cis=freqs_cis)
|
164 |
+
|
165 |
+
if start_pos is not None: # if we're performing inference, use kv caching
|
166 |
+
self.cache_k = self.cache_k.to(xq)
|
167 |
+
self.cache_v = self.cache_v.to(xq)
|
168 |
+
|
169 |
+
# set the values in our cache according to the current input
|
170 |
+
self.cache_k[:bsz, start_pos : start_pos + seqlen] = xk
|
171 |
+
self.cache_v[:bsz, start_pos : start_pos + seqlen] = xv
|
172 |
+
|
173 |
+
# grab our key and value matrixes which have a longer sequence length than our queries
|
174 |
+
keys = self.cache_k[:bsz, : start_pos + seqlen]
|
175 |
+
values = self.cache_v[:bsz, : start_pos + seqlen]
|
176 |
+
else:
|
177 |
+
# if we're training, do full sequence length
|
178 |
+
keys, values = xk, xv
|
179 |
+
|
180 |
+
keys = repeat_kv(keys, self.n_rep)
|
181 |
+
values = repeat_kv(values, self.n_rep)
|
182 |
+
|
183 |
+
xq = xq.transpose(1, 2)
|
184 |
+
keys = keys.transpose(1, 2)
|
185 |
+
values = values.transpose(1, 2)
|
186 |
+
|
187 |
+
scores = torch.matmul(xq, keys.transpose(2, 3)) / math.sqrt(self.head_dim)
|
188 |
+
if mask is not None:
|
189 |
+
scores = scores + mask
|
190 |
+
scores = F.softmax(scores.float(), dim=-1).type_as(xq)
|
191 |
+
|
192 |
+
output = torch.matmul(scores, values)
|
193 |
+
output = output.transpose(1, 2).contiguous().view(bsz, seqlen, -1)
|
194 |
+
return self.wo(output)
|
195 |
+
|
196 |
+
class FeedForward(nn.Module):
|
197 |
+
def __init__(
|
198 |
+
self,
|
199 |
+
dim: int,
|
200 |
+
hidden_dim: int,
|
201 |
+
multiple_of: int,
|
202 |
+
ffn_dim_multiplier: Optional[float],
|
203 |
+
):
|
204 |
+
super().__init__()
|
205 |
+
# custom dim factor multiplier that ensures we're using a multiple of "multiple_of" for hardware efficiency reasons
|
206 |
+
hidden_dim = int(2 * hidden_dim / 3)
|
207 |
+
if ffn_dim_multiplier is not None:
|
208 |
+
hidden_dim = int(ffn_dim_multiplier * hidden_dim)
|
209 |
+
hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
|
210 |
+
|
211 |
+
self.w1 = nn.Linear(dim, hidden_dim, bias=False)
|
212 |
+
self.w2 = nn.Linear(hidden_dim, dim, bias=False)
|
213 |
+
self.w3 = nn.Linear(dim, hidden_dim, bias=False)
|
214 |
+
|
215 |
+
def forward(self, x):
|
216 |
+
return self.w2(F.silu(self.w1(x)) * self.w3(x))
|
217 |
+
|
218 |
+
class TransformerBlock(nn.Module):
|
219 |
+
def __init__(self, args: ModelArgs):
|
220 |
+
super().__init__()
|
221 |
+
self.n_heads = args.n_heads
|
222 |
+
self.dim = args.dim
|
223 |
+
self.head_dim = args.dim // args.n_heads
|
224 |
+
self.attention = Attention(args)
|
225 |
+
self.feed_forward = FeedForward(
|
226 |
+
dim=args.dim,
|
227 |
+
hidden_dim=4 * args.dim,
|
228 |
+
multiple_of=args.multiple_of,
|
229 |
+
ffn_dim_multiplier=args.ffn_dim_multiplier,
|
230 |
+
)
|
231 |
+
self.attention_norm = RMSNorm(args.dim, eps=args.norm_eps)
|
232 |
+
self.ffn_norm = RMSNorm(args.dim, eps=args.norm_eps)
|
233 |
+
self.dropout_rate = args.dropout_rate
|
234 |
+
|
235 |
+
def forward(
|
236 |
+
self,
|
237 |
+
x: torch.Tensor,
|
238 |
+
freqs_cis: torch.Tensor,
|
239 |
+
mask: Optional[torch.Tensor],
|
240 |
+
start_pos: int = None,
|
241 |
+
training = False,
|
242 |
+
):
|
243 |
+
h = x + F.dropout(self.attention(self.attention_norm(x), freqs_cis, mask, start_pos), p=self.dropout_rate, training=training)
|
244 |
+
out = h + F.dropout(self.feed_forward(self.ffn_norm(h)), p=self.dropout_rate, training=training)
|
245 |
+
return out
|
246 |
+
|
247 |
+
class Moose(nn.Module):
|
248 |
+
def __init__(self, params: ModelArgs):
|
249 |
+
super().__init__()
|
250 |
+
self.params = params
|
251 |
+
self.vocab_size = params.vocab_size
|
252 |
+
self.n_layers = params.n_layers
|
253 |
+
self.max_seq_len = params.max_seq_len
|
254 |
+
|
255 |
+
self.tok_embeddings = nn.Embedding(params.vocab_size, params.dim)
|
256 |
+
|
257 |
+
self.layers = torch.nn.ModuleList()
|
258 |
+
for _ in range(params.n_layers):
|
259 |
+
self.layers.append(TransformerBlock(params))
|
260 |
+
|
261 |
+
self.norm = RMSNorm(params.dim, eps=params.norm_eps)
|
262 |
+
self.output = nn.Linear(
|
263 |
+
params.dim,
|
264 |
+
params.vocab_size,
|
265 |
+
bias=False)
|
266 |
+
|
267 |
+
self.freqs_cis = precompute_freqs_cis(
|
268 |
+
params.dim // params.n_heads,
|
269 |
+
params.max_seq_len * 2,
|
270 |
+
params.rope_theta,)
|
271 |
+
|
272 |
+
# precompute the causal attention mask
|
273 |
+
mask = torch.full((params.max_seq_len, params.max_seq_len),
|
274 |
+
float("-inf"),
|
275 |
+
device=params.device)
|
276 |
+
mask = torch.triu(mask, diagonal=1)
|
277 |
+
self.register_buffer('mask', mask)
|
278 |
+
|
279 |
+
self.criterion = nn.CrossEntropyLoss()
|
280 |
+
|
281 |
+
def forward(self,
|
282 |
+
tokens: torch.Tensor,
|
283 |
+
targets: torch.Tensor):
|
284 |
+
bsz, seqlen = tokens.shape
|
285 |
+
assert tokens.shape == targets.shape
|
286 |
+
assert seqlen == self.max_seq_len
|
287 |
+
|
288 |
+
h = self.tok_embeddings(tokens)
|
289 |
+
|
290 |
+
freqs_cis = self.freqs_cis.to(h.device)
|
291 |
+
freqs_cis = self.freqs_cis[:seqlen]
|
292 |
+
|
293 |
+
for layer in self.layers:
|
294 |
+
h = layer(
|
295 |
+
h,
|
296 |
+
freqs_cis,
|
297 |
+
self.mask,
|
298 |
+
start_pos = None,
|
299 |
+
training = True
|
300 |
+
)
|
301 |
+
|
302 |
+
h = self.norm(h)
|
303 |
+
logits = self.output(h).float()
|
304 |
+
|
305 |
+
loss = self.criterion(
|
306 |
+
logits.view(bsz * seqlen, self.vocab_size),
|
307 |
+
targets.reshape(bsz * seqlen))
|
308 |
+
|
309 |
+
return logits, loss
|
310 |
+
|
311 |
+
@torch.inference_mode()
|
312 |
+
def forward_inference(self,
|
313 |
+
tokens: torch.Tensor,
|
314 |
+
start_pos: int,
|
315 |
+
max_context_window: int,
|
316 |
+
):
|
317 |
+
_bsz, seqlen = tokens.shape
|
318 |
+
h = self.tok_embeddings(tokens)
|
319 |
+
self.freqs_cis = self.freqs_cis.to(h.device)
|
320 |
+
freqs_cis = self.freqs_cis[start_pos : start_pos + seqlen]
|
321 |
+
|
322 |
+
mask = self.mask[:seqlen, :seqlen]
|
323 |
+
mask = torch.hstack(
|
324 |
+
[torch.zeros((seqlen, start_pos), device=tokens.device), mask]
|
325 |
+
).type_as(h)
|
326 |
+
|
327 |
+
for layer in self.layers:
|
328 |
+
h = layer(
|
329 |
+
h,
|
330 |
+
freqs_cis,
|
331 |
+
mask,
|
332 |
+
start_pos = start_pos
|
333 |
+
)
|
334 |
+
h = self.norm(h)
|
335 |
+
logits = self.output(h).float()
|
336 |
+
return logits
|
337 |
+
|
338 |
+
@torch.inference_mode()
|
339 |
+
def Sampler(
|
340 |
+
self,
|
341 |
+
logits: torch.Tensor,
|
342 |
+
temperature: float,
|
343 |
+
top_p: float
|
344 |
+
) -> torch.Tensor:
|
345 |
+
logits = logits[:,-1,:]
|
346 |
+
|
347 |
+
logits.div_(temperature)
|
348 |
+
|
349 |
+
probs = torch.softmax(logits, dim=-1, dtype=torch.float)
|
350 |
+
|
351 |
+
probs_sort, probs_idx = torch.sort(probs, dim=-1, descending=True)
|
352 |
+
|
353 |
+
probs_sum = torch.cumsum(probs_sort, dim=-1)
|
354 |
+
top_ps_mask = (probs_sum - probs_sort) > top_p
|
355 |
+
probs_sort = torch.where(top_ps_mask, 0, probs_sort)
|
356 |
+
probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
|
357 |
+
probs = torch.gather(probs_sort, dim=-1, index=torch.argsort(probs_idx, dim=-1))
|
358 |
+
next_token_id = torch.multinomial(probs, num_samples=1)
|
359 |
+
|
360 |
+
return next_token_id
|
361 |
+
|
362 |
+
@torch.inference_mode()
|
363 |
+
def generate(
|
364 |
+
self,
|
365 |
+
prompt: str,
|
366 |
+
max_gen_len: int = 100,
|
367 |
+
memory_saver_div: int = 1,
|
368 |
+
temperature: float = 0.6,
|
369 |
+
top_p: float = 0.9,
|
370 |
+
) -> str:
|
371 |
+
assert ((memory_saver_div & (memory_saver_div-1)) == 0) & (memory_saver_div > 0), f'memory_saver_div {memory_saver_div} must be power of 2'
|
372 |
+
max_context_window = self.max_seq_len // memory_saver_div
|
373 |
+
|
374 |
+
enc = tiktoken.get_encoding('gpt2')
|
375 |
+
tokens = enc.encode(prompt)
|
376 |
+
|
377 |
+
tokens = torch.tensor(tokens, device=self.params.device)
|
378 |
+
tokens = tokens.unsqueeze(0) if len(tokens.shape)==1 else tokens
|
379 |
+
|
380 |
+
start_pos = max(tokens.shape[1] - max_context_window, 0)
|
381 |
+
eot = enc._special_tokens['<|endoftext|>'] # end of text token
|
382 |
+
|
383 |
+
while True:
|
384 |
+
logits = self.forward_inference(
|
385 |
+
tokens[:,-max_context_window:],
|
386 |
+
start_pos = start_pos,
|
387 |
+
max_context_window = max_context_window
|
388 |
+
)
|
389 |
+
|
390 |
+
next_token = self.Sampler(
|
391 |
+
logits = logits,
|
392 |
+
temperature = temperature,
|
393 |
+
top_p = top_p,
|
394 |
+
)
|
395 |
+
|
396 |
+
tokens = torch.cat((tokens, next_token), dim=1)
|
397 |
+
|
398 |
+
if next_token.item() == eot:
|
399 |
+
break
|
400 |
+
|
401 |
+
if tokens.shape[1] >= max_context_window:
|
402 |
+
start_pos += 1
|
403 |
+
|
404 |
+
output = enc.decode(tokens.squeeze(0).tolist())
|
405 |
+
|
406 |
+
return output
|