init
Browse files- inference.py +254 -0
- vocab_v20230424.txt +0 -0
inference.py
ADDED
@@ -0,0 +1,254 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import numpy as np
|
2 |
+
|
3 |
+
np.set_printoptions(precision=4, suppress=True, linewidth=200)
|
4 |
+
import types, torch
|
5 |
+
from torch.nn import functional as F
|
6 |
+
|
7 |
+
MyModule = torch.jit.ScriptModule
|
8 |
+
MyFunction = torch.jit.script_method
|
9 |
+
|
10 |
+
|
11 |
+
class RWKV_TOKENIZER():
|
12 |
+
table: list[list[list[bytes]]]
|
13 |
+
good: list[set[int]]
|
14 |
+
wlen: list[int]
|
15 |
+
|
16 |
+
def __init__(self, file_name):
|
17 |
+
self.idx2token = {}
|
18 |
+
sorted = [] # must be already sorted
|
19 |
+
lines = open(file_name, "r", encoding="utf-8").readlines()
|
20 |
+
for l in lines:
|
21 |
+
idx = int(l[:l.index(' ')])
|
22 |
+
x = eval(l[l.index(' '):l.rindex(' ')])
|
23 |
+
x = x.encode("utf-8") if isinstance(x, str) else x
|
24 |
+
assert isinstance(x, bytes)
|
25 |
+
assert len(x) == int(l[l.rindex(' '):])
|
26 |
+
sorted += [x]
|
27 |
+
self.idx2token[idx] = x
|
28 |
+
|
29 |
+
self.token2idx = {}
|
30 |
+
for k, v in self.idx2token.items():
|
31 |
+
self.token2idx[v] = int(k)
|
32 |
+
|
33 |
+
# precompute some tables for fast matching
|
34 |
+
self.table = [[[] for j in range(256)] for i in range(256)]
|
35 |
+
self.good = [set() for i in range(256)]
|
36 |
+
self.wlen = [0 for i in range(256)]
|
37 |
+
|
38 |
+
for i in reversed(range(len(sorted))): # reverse order - match longer tokens first
|
39 |
+
s = sorted[i]
|
40 |
+
if len(s) >= 2:
|
41 |
+
s0 = int(s[0])
|
42 |
+
s1 = int(s[1])
|
43 |
+
self.table[s0][s1] += [s]
|
44 |
+
self.wlen[s0] = max(self.wlen[s0], len(s))
|
45 |
+
self.good[s0].add(s1)
|
46 |
+
|
47 |
+
def encodeBytes(self, src: bytes) -> list[int]:
|
48 |
+
src_len: int = len(src)
|
49 |
+
tokens: list[int] = []
|
50 |
+
i: int = 0
|
51 |
+
while i < src_len:
|
52 |
+
s: bytes = src[i: i + 1]
|
53 |
+
|
54 |
+
if i < src_len - 1:
|
55 |
+
s1: int = int(src[i + 1])
|
56 |
+
s0: int = int(src[i])
|
57 |
+
if s1 in self.good[s0]:
|
58 |
+
sss: bytes = src[i: i + self.wlen[s0]]
|
59 |
+
try:
|
60 |
+
s = next(filter(sss.startswith, self.table[s0][s1]))
|
61 |
+
except:
|
62 |
+
pass
|
63 |
+
tokens.append(self.token2idx[s])
|
64 |
+
i += len(s)
|
65 |
+
|
66 |
+
return tokens
|
67 |
+
|
68 |
+
def decodeBytes(self, tokens):
|
69 |
+
return b''.join(map(lambda i: self.idx2token[i], tokens))
|
70 |
+
|
71 |
+
def encode(self, src: str):
|
72 |
+
return self.encodeBytes(src.encode("utf-8"))
|
73 |
+
|
74 |
+
def decode(self, tokens):
|
75 |
+
return self.decodeBytes(tokens).decode('utf-8')
|
76 |
+
|
77 |
+
def printTokens(self, tokens):
|
78 |
+
for i in tokens:
|
79 |
+
s = self.idx2token[i]
|
80 |
+
try:
|
81 |
+
s = s.decode('utf-8')
|
82 |
+
except:
|
83 |
+
pass
|
84 |
+
print(f'{repr(s)}{i}', end=' ')
|
85 |
+
# print(repr(s), i)
|
86 |
+
print()
|
87 |
+
|
88 |
+
|
89 |
+
########################################################################################################
|
90 |
+
|
91 |
+
def sample_logits(out, temperature=1.0, top_p=0.8):
|
92 |
+
probs = F.softmax(out, dim=-1).numpy()
|
93 |
+
sorted_probs = np.sort(probs)[::-1]
|
94 |
+
cumulative_probs = np.cumsum(sorted_probs)
|
95 |
+
cutoff = float(sorted_probs[np.argmax(cumulative_probs > top_p)])
|
96 |
+
probs[probs < cutoff] = 0
|
97 |
+
if temperature != 1.0:
|
98 |
+
probs = probs.pow(1.0 / temperature)
|
99 |
+
probs = probs / np.sum(probs)
|
100 |
+
out = np.random.choice(a=len(probs), p=probs)
|
101 |
+
return out
|
102 |
+
|
103 |
+
|
104 |
+
########################################################################################################
|
105 |
+
class RWKV_RNN(MyModule):
|
106 |
+
def __init__(self, args):
|
107 |
+
super().__init__()
|
108 |
+
self.args = args
|
109 |
+
self.eval() # set torch to inference mode
|
110 |
+
|
111 |
+
w = torch.load(args.MODEL_NAME + '.pth', map_location='cpu')
|
112 |
+
|
113 |
+
for k in w.keys():
|
114 |
+
w[k] = w[k].float() # convert to f32 type
|
115 |
+
if '.time_' in k: w[k] = w[k].squeeze()
|
116 |
+
if '.time_faaaa' in k: w[k] = w[k].unsqueeze(-1)
|
117 |
+
|
118 |
+
self.n_head = w['blocks.0.att.time_faaaa'].shape[0]
|
119 |
+
self.head_size = w['blocks.0.ln1.weight'].shape[0] // self.n_head
|
120 |
+
|
121 |
+
self.w = types.SimpleNamespace() # set self.w from w
|
122 |
+
self.w.blocks = {}
|
123 |
+
for k in w.keys(): # example: "blocks.0.att.time_first" => self.w.blocks[0].att.time_first
|
124 |
+
parts = k.split('.')
|
125 |
+
last = parts.pop()
|
126 |
+
here = self.w
|
127 |
+
for p in parts:
|
128 |
+
if p.isdigit():
|
129 |
+
p = int(p)
|
130 |
+
if p not in here: here[p] = types.SimpleNamespace()
|
131 |
+
here = here[p]
|
132 |
+
else:
|
133 |
+
if not hasattr(here, p): setattr(here, p, types.SimpleNamespace())
|
134 |
+
here = getattr(here, p)
|
135 |
+
setattr(here, last, w[k])
|
136 |
+
|
137 |
+
def layer_norm(self, x, w):
|
138 |
+
return F.layer_norm(x, (self.args.n_embd,), weight=w.weight, bias=w.bias)
|
139 |
+
|
140 |
+
@MyFunction
|
141 |
+
def channel_mixing(self, x, state, i: int, time_maa_k, time_maa_r, kw, vw, rw):
|
142 |
+
i0 = (2 + self.head_size) * i + 0
|
143 |
+
sx = state[i0] - x
|
144 |
+
xk = x + sx * time_maa_k
|
145 |
+
xr = x + sx * time_maa_r
|
146 |
+
state[i0] = x
|
147 |
+
r = torch.sigmoid(rw @ xr)
|
148 |
+
k = torch.square(torch.relu(kw @ xk)) # square relu, primer paper
|
149 |
+
return r * (vw @ k)
|
150 |
+
|
151 |
+
@MyFunction
|
152 |
+
def time_mixing(self, x, state, i: int, x_maa, w_maa, k_maa, v_maa, r_maa, g_maa, tm_w1, tm_w2, td_w1, td_w2,
|
153 |
+
time_first, time_decay, kw, vw, rw, gw, ow, ln_w, ln_b):
|
154 |
+
H = self.n_head
|
155 |
+
S = self.head_size
|
156 |
+
|
157 |
+
i1 = (2 + S) * i + 1
|
158 |
+
sx = state[i1] - x
|
159 |
+
state[i1] = x
|
160 |
+
xxx = x + sx * x_maa
|
161 |
+
xxx = torch.tanh(xxx @ tm_w1).view(5, 1, -1)
|
162 |
+
xxx = torch.bmm(xxx, tm_w2).view(5, -1)
|
163 |
+
mw, mk, mv, mr, mg = xxx.unbind(dim=0)
|
164 |
+
|
165 |
+
xw = x + sx * (w_maa + mw)
|
166 |
+
xk = x + sx * (k_maa + mk)
|
167 |
+
xv = x + sx * (v_maa + mv)
|
168 |
+
xr = x + sx * (r_maa + mr)
|
169 |
+
xg = x + sx * (g_maa + mg)
|
170 |
+
|
171 |
+
w = (time_decay + (torch.tanh(xw @ td_w1) @ td_w2).float()).view(H, S, 1)
|
172 |
+
w = torch.exp(-torch.exp(w.float()))
|
173 |
+
|
174 |
+
r = (rw @ xr).view(H, 1, S)
|
175 |
+
k = (kw @ xk).view(H, S, 1)
|
176 |
+
v = (vw @ xv).view(H, 1, S)
|
177 |
+
g = F.silu(gw @ xg)
|
178 |
+
|
179 |
+
s = state[(2 + S) * i + 2:(2 + S) * (i + 1), :].reshape(H, S, S)
|
180 |
+
|
181 |
+
x = torch.zeros(H, S)
|
182 |
+
a = k @ v
|
183 |
+
x = r @ (time_first * a + s)
|
184 |
+
s = a + w * s
|
185 |
+
|
186 |
+
state[(2 + S) * i + 2:(2 + S) * (i + 1), :] = s.reshape(S, -1)
|
187 |
+
x = x.flatten()
|
188 |
+
|
189 |
+
x = F.group_norm(x.unsqueeze(0), num_groups=H, weight=ln_w, bias=ln_b, eps=64e-5).squeeze(
|
190 |
+
0) * g # same as gn(x/8, eps=1e-5)
|
191 |
+
return ow @ x
|
192 |
+
|
193 |
+
def forward(self, token, state):
|
194 |
+
with torch.no_grad():
|
195 |
+
if state == None:
|
196 |
+
state = torch.zeros(self.args.n_layer * (2 + self.head_size), self.args.n_embd)
|
197 |
+
|
198 |
+
x = self.w.emb.weight[token]
|
199 |
+
x = self.layer_norm(x, self.w.blocks[0].ln0)
|
200 |
+
for i in range(self.args.n_layer):
|
201 |
+
att = self.w.blocks[i].att
|
202 |
+
x = x + self.time_mixing(self.layer_norm(x, self.w.blocks[i].ln1), state, i,
|
203 |
+
att.time_maa_x, att.time_maa_w, att.time_maa_k, att.time_maa_v, att.time_maa_r,
|
204 |
+
att.time_maa_g, att.time_maa_w1, att.time_maa_w2,
|
205 |
+
att.time_decay_w1, att.time_decay_w2, att.time_faaaa, att.time_decay,
|
206 |
+
att.key.weight, att.value.weight, att.receptance.weight, att.gate.weight,
|
207 |
+
att.output.weight,
|
208 |
+
att.ln_x.weight, att.ln_x.bias)
|
209 |
+
ffn = self.w.blocks[i].ffn
|
210 |
+
x = x + self.channel_mixing(self.layer_norm(x, self.w.blocks[i].ln2), state, i,
|
211 |
+
ffn.time_maa_k, ffn.time_maa_r,
|
212 |
+
ffn.key.weight, ffn.value.weight, ffn.receptance.weight)
|
213 |
+
|
214 |
+
x = self.w.head.weight @ self.layer_norm(x, self.w.ln_out)
|
215 |
+
return x.float(), state
|
216 |
+
|
217 |
+
|
218 |
+
tokenizer = RWKV_TOKENIZER("vocab_v20230424.txt")
|
219 |
+
|
220 |
+
args = types.SimpleNamespace()
|
221 |
+
args.MODEL_NAME = 'rwkv-30'
|
222 |
+
args.n_layer = 12
|
223 |
+
args.n_embd = 768
|
224 |
+
args.vocab_size = 65536
|
225 |
+
|
226 |
+
context = "Today is a beautiful"
|
227 |
+
NUM_TRIALS = 3
|
228 |
+
LENGTH_PER_TRIAL = 50
|
229 |
+
TEMPERATURE = 1.0
|
230 |
+
TOP_P = 0.7
|
231 |
+
|
232 |
+
print(f'model= {args.MODEL_NAME}\n')
|
233 |
+
model = RWKV_RNN(args)
|
234 |
+
init_state = None
|
235 |
+
for token in tokenizer.encode(context):
|
236 |
+
init_out, init_state = model.forward(token, init_state)
|
237 |
+
|
238 |
+
for TRIAL in range(NUM_TRIALS):
|
239 |
+
print(f'Trial {TRIAL + 1}=', context, end="")
|
240 |
+
all_tokens = []
|
241 |
+
n_sampled = 0
|
242 |
+
out, state = init_out.clone(), init_state.clone()
|
243 |
+
for i in range(LENGTH_PER_TRIAL):
|
244 |
+
token = sample_logits(out, TEMPERATURE, TOP_P)
|
245 |
+
all_tokens += [token]
|
246 |
+
try:
|
247 |
+
tmp = tokenizer.decode(all_tokens[n_sampled:])
|
248 |
+
if '\ufffd' not in tmp: # only print when we have a valid utf-8 string
|
249 |
+
print(tmp, end="", flush=True)
|
250 |
+
n_sampled = i + 1
|
251 |
+
except:
|
252 |
+
pass
|
253 |
+
out, state = model.forward(token, state)
|
254 |
+
print("\nSampled tokens=", n_sampled, "out of", LENGTH_PER_TRIAL, "tokens\n")
|
vocab_v20230424.txt
ADDED
The diff for this file is too large to render.
See raw diff
|
|