Upload 3 files
Browse files- architecture.py +129 -0
- config.json +7 -0
- generate.py +138 -0
architecture.py
ADDED
@@ -0,0 +1,129 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import torch.nn as nn
|
3 |
+
import torch.nn.functional as F
|
4 |
+
import xformers.ops as xops
|
5 |
+
|
6 |
+
class SmallGPT(nn.Module):
|
7 |
+
def __init__(self, vocab_size, d_model=256, n_heads=8, n_layers=6, max_length=128, pad_idx=0):
|
8 |
+
super().__init__()
|
9 |
+
self.d_model = d_model
|
10 |
+
self.max_length = max_length
|
11 |
+
|
12 |
+
# Embeddings
|
13 |
+
self.token_embedding = nn.Embedding(vocab_size, d_model, padding_idx=pad_idx)
|
14 |
+
self.position_embedding = nn.Embedding(max_length, d_model)
|
15 |
+
|
16 |
+
# Transformer blocks
|
17 |
+
self.blocks = nn.ModuleList([
|
18 |
+
TransformerBlock(d_model, n_heads) for _ in range(n_layers)
|
19 |
+
])
|
20 |
+
|
21 |
+
# Output
|
22 |
+
self.ln_f = nn.LayerNorm(d_model)
|
23 |
+
self.head = nn.Linear(d_model, vocab_size, bias=False)
|
24 |
+
|
25 |
+
# Init weights
|
26 |
+
self.apply(self._init_weights)
|
27 |
+
|
28 |
+
def _init_weights(self, module):
|
29 |
+
if isinstance(module, nn.Linear):
|
30 |
+
torch.nn.init.normal_(module.weight, mean=0.0, std=0.03)
|
31 |
+
if module.bias is not None:
|
32 |
+
torch.nn.init.zeros_(module.bias)
|
33 |
+
elif isinstance(module, nn.Embedding):
|
34 |
+
torch.nn.init.normal_(module.weight, mean=0.0, std=0.03)
|
35 |
+
|
36 |
+
def forward(self, x):
|
37 |
+
batch_size, seq_len = x.size()
|
38 |
+
|
39 |
+
# position indices
|
40 |
+
pos = torch.arange(0, seq_len, dtype=torch.long, device=x.device)
|
41 |
+
pos = pos.unsqueeze(0).expand(batch_size, seq_len)
|
42 |
+
|
43 |
+
# Embeddings
|
44 |
+
tok_emb = self.token_embedding(x)
|
45 |
+
pos_emb = self.position_embedding(pos)
|
46 |
+
x = tok_emb + pos_emb
|
47 |
+
|
48 |
+
# Transformer blocks
|
49 |
+
for block in self.blocks:
|
50 |
+
x = block(x)
|
51 |
+
|
52 |
+
# Final layer norm and projection
|
53 |
+
x = self.ln_f(x)
|
54 |
+
logits = self.head(x)
|
55 |
+
|
56 |
+
return logits
|
57 |
+
|
58 |
+
|
59 |
+
class TransformerBlock(nn.Module):
|
60 |
+
def __init__(self, d_model, n_heads):
|
61 |
+
super().__init__()
|
62 |
+
self.ln1 = nn.LayerNorm(d_model)
|
63 |
+
self.attn = CausalSelfAttention(d_model, n_heads)
|
64 |
+
self.ln2 = nn.LayerNorm(d_model)
|
65 |
+
self.mlp = MLP(d_model)
|
66 |
+
|
67 |
+
def forward(self, x):
|
68 |
+
x = x + self.attn(self.ln1(x))
|
69 |
+
x = x + self.mlp(self.ln2(x))
|
70 |
+
return x
|
71 |
+
|
72 |
+
|
73 |
+
class CausalSelfAttention(nn.Module):
|
74 |
+
def __init__(self, d_model, n_heads):
|
75 |
+
super().__init__()
|
76 |
+
assert d_model % n_heads == 0
|
77 |
+
self.n_heads = n_heads
|
78 |
+
self.d_model = d_model
|
79 |
+
self.head_dim = d_model // n_heads
|
80 |
+
|
81 |
+
self.qkv = nn.Linear(d_model, 3 * d_model)
|
82 |
+
self.proj = nn.Linear(d_model, d_model)
|
83 |
+
|
84 |
+
def forward(self, x):
|
85 |
+
batch, seq_len, d_model = x.size()
|
86 |
+
qkv = self.qkv(x) # [B, S, 3*D]
|
87 |
+
q, k, v = qkv.chunk(3, dim=-1)
|
88 |
+
|
89 |
+
# reshape into heads
|
90 |
+
q = q.view(batch, seq_len, self.n_heads, self.head_dim).transpose(1, 2) # [B, H, S, Hd]
|
91 |
+
k = k.view(batch, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
|
92 |
+
v = v.view(batch, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
|
93 |
+
|
94 |
+
# flatten for xformers: [B*H, S, Hd]
|
95 |
+
q = q.reshape(batch * self.n_heads, seq_len, self.head_dim)
|
96 |
+
k = k.reshape(batch * self.n_heads, seq_len, self.head_dim)
|
97 |
+
v = v.reshape(batch * self.n_heads, seq_len, self.head_dim)
|
98 |
+
|
99 |
+
# apply memory-efficient attention with causal mask
|
100 |
+
out = xops.memory_efficient_attention(q, k, v, attn_bias=xops.LowerTriangularMask())
|
101 |
+
# out: [B*H, S, Hd]
|
102 |
+
|
103 |
+
# reshape back
|
104 |
+
out = out.view(batch, self.n_heads, seq_len, self.head_dim)
|
105 |
+
out = out.transpose(1, 2).contiguous().view(batch, seq_len, d_model)
|
106 |
+
|
107 |
+
return self.proj(out)
|
108 |
+
|
109 |
+
class MLP(nn.Module):
|
110 |
+
def __init__(self, d_model):
|
111 |
+
super().__init__()
|
112 |
+
self.fc1 = nn.Linear(d_model, 4 * d_model)
|
113 |
+
self.fc2 = nn.Linear(4 * d_model, d_model)
|
114 |
+
self.silu = nn.SiLU()
|
115 |
+
|
116 |
+
def forward(self, x):
|
117 |
+
x = self.fc1(x)
|
118 |
+
x = self.silu(x)
|
119 |
+
x = self.fc2(x)
|
120 |
+
return x
|
121 |
+
|
122 |
+
|
123 |
+
DEFAULT_CONFIG = {
|
124 |
+
"vocab_size": 24_005,
|
125 |
+
"d_model": 256,
|
126 |
+
"n_heads": 8,
|
127 |
+
"n_layers": 6,
|
128 |
+
"max_length": 128,
|
129 |
+
}
|
config.json
ADDED
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
{
|
2 |
+
"vocab_size": 24005,
|
3 |
+
"d_model": 256,
|
4 |
+
"n_heads": 8,
|
5 |
+
"n_layers": 6,
|
6 |
+
"max_length": 128
|
7 |
+
}
|
generate.py
ADDED
@@ -0,0 +1,138 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import torch
|
2 |
+
import sys
|
3 |
+
import time
|
4 |
+
from architecture import SmallGPT
|
5 |
+
from tokenizers import Tokenizer
|
6 |
+
|
7 |
+
def load_tokenizer(path="smptokenizer/tokenizer.json"):
|
8 |
+
tokenizer = Tokenizer.from_file(path)
|
9 |
+
return tokenizer
|
10 |
+
|
11 |
+
def generate_text_streaming(model, tokenizer, start_text, device, max_length=64, temperature=1.0, max_new_tokens=20, repetition_penalty=1.2):
|
12 |
+
"""
|
13 |
+
Generates text token by token, yielding each new token.
|
14 |
+
"""
|
15 |
+
model.eval()
|
16 |
+
|
17 |
+
# Encode start text
|
18 |
+
input_ids = tokenizer.encode(start_text).ids
|
19 |
+
generated_ids = []
|
20 |
+
|
21 |
+
# Print the starting text, and wait for the model to continue
|
22 |
+
print("Generated Sentence:")
|
23 |
+
print(start_text, end="", flush=True)
|
24 |
+
|
25 |
+
current_ids = input_ids
|
26 |
+
|
27 |
+
with torch.no_grad():
|
28 |
+
for _ in range(max_new_tokens):
|
29 |
+
# Prepare input (truncate if too long)
|
30 |
+
current_input = current_ids[-max_length+1:] if len(current_ids) >= max_length else current_ids
|
31 |
+
input_tensor = torch.tensor([current_input], dtype=torch.long, device=device)
|
32 |
+
|
33 |
+
# Get output
|
34 |
+
logits = model(input_tensor)
|
35 |
+
|
36 |
+
# Get logits for last position
|
37 |
+
next_token_logits = logits[0, -1, :] / temperature
|
38 |
+
|
39 |
+
# Apply repetition penalty, if needed
|
40 |
+
if repetition_penalty > 1.0:
|
41 |
+
for token_id in set(current_ids):
|
42 |
+
next_token_logits[token_id] /= repetition_penalty
|
43 |
+
|
44 |
+
# Sample next token
|
45 |
+
probs = torch.softmax(next_token_logits, dim=-1)
|
46 |
+
next_token_id = torch.multinomial(probs, 1).item()
|
47 |
+
|
48 |
+
# Check for EOS
|
49 |
+
if next_token_id == tokenizer.token_to_id("<eos>"):
|
50 |
+
break
|
51 |
+
|
52 |
+
generated_ids.append(next_token_id)
|
53 |
+
current_ids.append(next_token_id)
|
54 |
+
|
55 |
+
# Decode and yield the new token
|
56 |
+
new_token = tokenizer.decode([next_token_id])
|
57 |
+
yield new_token
|
58 |
+
|
59 |
+
def main(seed):
|
60 |
+
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
61 |
+
print(f"Using device: {device}")
|
62 |
+
|
63 |
+
# Load tokenizer
|
64 |
+
tokenizer_path = "smptokenizer/tokenizer.json"
|
65 |
+
tokenizer = load_tokenizer(tokenizer_path)
|
66 |
+
vocab_size = tokenizer.get_vocab_size()
|
67 |
+
pad_id = tokenizer.token_to_id("<pad>") or 0
|
68 |
+
|
69 |
+
# Model parameters from training
|
70 |
+
d_model = 256
|
71 |
+
n_heads = 8
|
72 |
+
n_layers = 6
|
73 |
+
max_length = 172
|
74 |
+
|
75 |
+
# Instantiate the model
|
76 |
+
model = SmallGPT(
|
77 |
+
vocab_size=vocab_size,
|
78 |
+
d_model=d_model,
|
79 |
+
n_heads=n_heads,
|
80 |
+
n_layers=n_layers,
|
81 |
+
max_length=max_length,
|
82 |
+
pad_idx=pad_id,
|
83 |
+
).to(device)
|
84 |
+
|
85 |
+
# Load the trained model weights
|
86 |
+
model_path = "models/pytorch_model.bin" # idk if safetensor works
|
87 |
+
try:
|
88 |
+
model.load_state_dict(torch.load(model_path, map_location=device))
|
89 |
+
model.eval()
|
90 |
+
print(f"Model loaded from {model_path}")
|
91 |
+
except FileNotFoundError:
|
92 |
+
print(f"Error: Model file not found at {model_path}")
|
93 |
+
print("Please ensure the model is trained and the path is correct.")
|
94 |
+
return
|
95 |
+
|
96 |
+
while True:
|
97 |
+
# Reset seed
|
98 |
+
torch.manual_seed(seed)
|
99 |
+
torch.cuda.manual_seed(seed)
|
100 |
+
|
101 |
+
start_text = input("Enter a starting word or phrase (or 'quit' to exit): ")
|
102 |
+
if start_text.lower() == 'quit':
|
103 |
+
break
|
104 |
+
|
105 |
+
if not start_text.strip():
|
106 |
+
print("Please enter some text. We are using a random character as a starting point.")
|
107 |
+
start_text = str(time.time())
|
108 |
+
|
109 |
+
print("Generating...")
|
110 |
+
|
111 |
+
token_count = 0
|
112 |
+
start_time = time.time()
|
113 |
+
|
114 |
+
for token in generate_text_streaming(
|
115 |
+
model=model,
|
116 |
+
tokenizer=tokenizer,
|
117 |
+
start_text=start_text,
|
118 |
+
device=device,
|
119 |
+
max_new_tokens=1000,
|
120 |
+
temperature=0.7,
|
121 |
+
max_length=max_length,
|
122 |
+
repetition_penalty=1.2
|
123 |
+
):
|
124 |
+
print(token, end="", flush=True)
|
125 |
+
token_count += 1
|
126 |
+
|
127 |
+
end_time = time.time()
|
128 |
+
elapsed_time = end_time - start_time
|
129 |
+
tokens_per_sec = token_count / elapsed_time if elapsed_time > 0 else 0
|
130 |
+
|
131 |
+
print(f"\n\nPerformance: {tokens_per_sec:.2f} tokens/sec")
|
132 |
+
print("-" * 30)
|
133 |
+
|
134 |
+
if __name__ == "__main__":
|
135 |
+
seed = 42
|
136 |
+
torch.manual_seed(seed)
|
137 |
+
torch.cuda.manual_seed(seed)
|
138 |
+
main(seed)
|