wip
Browse files
app.py
CHANGED
@@ -13,17 +13,19 @@ from gated_state_spaces_pytorch import GatedStateSpacesLM
|
|
13 |
from gated_state_spaces_pytorch.autoregressive_wrapper import AutoregressiveWrapper
|
14 |
|
15 |
from c4x import C4X
|
|
|
16 |
|
17 |
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
entity="naxalpha",
|
22 |
)
|
23 |
|
24 |
-
|
25 |
-
|
26 |
-
|
|
|
|
|
27 |
|
28 |
f_emb = 1600
|
29 |
model = AutoregressiveWrapper(
|
@@ -32,56 +34,57 @@ if __name__ == '__main__':
|
|
32 |
dim=f_emb,
|
33 |
depth=24,
|
34 |
),
|
35 |
-
)
|
36 |
-
wandb.watch(model)
|
37 |
-
|
38 |
-
# emb = gpt_2.state_dict()['transformer.wte.weight']
|
39 |
-
|
40 |
model.net.token_emb.weight.requires_grad_(False)
|
41 |
-
# model.net.token_emb.weight.copy_(emb)
|
42 |
-
|
43 |
model.net.to_logits.weight.requires_grad_(False)
|
44 |
-
# model.net.to_logits.weight.copy_(emb)
|
45 |
-
|
46 |
model.net.to_logits = nn.Sequential(
|
47 |
nn.LayerNorm(f_emb),
|
48 |
model.net.to_logits,
|
49 |
)
|
|
|
|
|
|
|
|
|
50 |
|
51 |
model.load_state_dict(torch.load('model.pt'))
|
52 |
-
model = model.cuda()
|
53 |
optim = AdamW(model.parameters(), 2e-5)
|
54 |
|
55 |
-
bs =
|
56 |
kk = 128
|
57 |
dsx = C4X(kk+1)
|
58 |
dlx = DataLoader(
|
59 |
dsx,
|
60 |
batch_size=bs,
|
61 |
-
num_workers=
|
62 |
)
|
63 |
|
64 |
k = 4
|
65 |
-
prog = tqdm(dlx)
|
|
|
|
|
|
|
66 |
optim.zero_grad()
|
67 |
-
|
68 |
for i, batch in enumerate(prog):
|
69 |
-
batch = batch.
|
70 |
-
|
71 |
-
|
72 |
-
|
73 |
-
|
74 |
-
|
75 |
-
|
76 |
-
|
77 |
-
|
|
|
78 |
optim.step()
|
79 |
optim.zero_grad()
|
80 |
|
81 |
-
if i % 1000 == 0:
|
|
|
|
|
|
|
82 |
b, n = 4, 512
|
83 |
-
init = torch.tensor([[50256]]*b).
|
84 |
-
prd =
|
85 |
prd = [dsx.decode(p) for p in prd]
|
86 |
try:
|
87 |
wandb.log(dict(
|
@@ -92,9 +95,14 @@ if __name__ == '__main__':
|
|
92 |
)), step=i)
|
93 |
except Exception as ex:
|
94 |
print('Failed to log to W&B...', ex)
|
95 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
96 |
|
97 |
-
|
98 |
-
|
99 |
-
), step=i)
|
100 |
-
prog.set_postfix(loss=los.item())
|
|
|
13 |
from gated_state_spaces_pytorch.autoregressive_wrapper import AutoregressiveWrapper
|
14 |
|
15 |
from c4x import C4X
|
16 |
+
from accelerate import Accelerator
|
17 |
|
18 |
|
19 |
+
def main():
|
20 |
+
accelerator = Accelerator(
|
21 |
+
gradient_accumulation_steps=4,
|
|
|
22 |
)
|
23 |
|
24 |
+
if accelerator.is_main_process:
|
25 |
+
wandb.init(
|
26 |
+
project="gated-state-space",
|
27 |
+
entity="naxalpha",
|
28 |
+
)
|
29 |
|
30 |
f_emb = 1600
|
31 |
model = AutoregressiveWrapper(
|
|
|
34 |
dim=f_emb,
|
35 |
depth=24,
|
36 |
),
|
37 |
+
)
|
|
|
|
|
|
|
|
|
38 |
model.net.token_emb.weight.requires_grad_(False)
|
|
|
|
|
39 |
model.net.to_logits.weight.requires_grad_(False)
|
|
|
|
|
40 |
model.net.to_logits = nn.Sequential(
|
41 |
nn.LayerNorm(f_emb),
|
42 |
model.net.to_logits,
|
43 |
)
|
44 |
+
model = model.to(accelerator.device)
|
45 |
+
|
46 |
+
if accelerator.is_main_process:
|
47 |
+
wandb.watch(model)
|
48 |
|
49 |
model.load_state_dict(torch.load('model.pt'))
|
|
|
50 |
optim = AdamW(model.parameters(), 2e-5)
|
51 |
|
52 |
+
bs = 16
|
53 |
kk = 128
|
54 |
dsx = C4X(kk+1)
|
55 |
dlx = DataLoader(
|
56 |
dsx,
|
57 |
batch_size=bs,
|
58 |
+
num_workers=8,
|
59 |
)
|
60 |
|
61 |
k = 4
|
62 |
+
prog = tqdm(dlx, disable=not accelerator.is_main_process)
|
63 |
+
|
64 |
+
model, optim, dlx = accelerator.prepare(model, optim, dlx)
|
65 |
+
|
66 |
optim.zero_grad()
|
|
|
67 |
for i, batch in enumerate(prog):
|
68 |
+
batch = batch.to(accelerator.device)
|
69 |
+
with accelerator.accumulate(model):
|
70 |
+
with accelerator.autocast():
|
71 |
+
los = model(batch)
|
72 |
+
accelerator.backward(los)
|
73 |
+
if accelerator.sync_gradients:
|
74 |
+
accelerator.clip_grad_norm_(
|
75 |
+
model.parameters(),
|
76 |
+
1.0,
|
77 |
+
)
|
78 |
optim.step()
|
79 |
optim.zero_grad()
|
80 |
|
81 |
+
if i % 1000 == 0 and accelerator.is_main_process:
|
82 |
+
print('generating...')
|
83 |
+
accelerator.wait_for_everyone()
|
84 |
+
unwrapped_model = accelerator.unwrap_model(model)
|
85 |
b, n = 4, 512
|
86 |
+
init = torch.tensor([[50256]]*b).to(accelerator.device)
|
87 |
+
prd = unwrapped_model.generate(init, n)
|
88 |
prd = [dsx.decode(p) for p in prd]
|
89 |
try:
|
90 |
wandb.log(dict(
|
|
|
95 |
)), step=i)
|
96 |
except Exception as ex:
|
97 |
print('Failed to log to W&B...', ex)
|
98 |
+
accelerator.save(unwrapped_model.state_dict(), 'model.pt')
|
99 |
+
|
100 |
+
if i % 10 == 0 and accelerator.is_main_process:
|
101 |
+
print('logging...')
|
102 |
+
wandb.log(dict(
|
103 |
+
loss=los.item(),
|
104 |
+
), step=i)
|
105 |
+
prog.set_postfix(loss=los.item())
|
106 |
|
107 |
+
if __name__ == '__main__':
|
108 |
+
main()
|
|
|
|