Spaces:
Sleeping
Sleeping
File size: 5,103 Bytes
3905c4a |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 |
import torch
import torch.nn as nn
from torch.nn import functional as F
import math, time, os
from torch.utils.data import Dataset, DataLoader
import tiktoken
# from torch.cuda.amp import autocast, GradScaler
from torch.amp.autocast_mode import autocast
from torch.amp.grad_scaler import GradScaler
from tqdm import tqdm
from datasets import load_dataset
from components.model import GPTModel
from components.dataset import TextDataset
# Load dataset
dataset = load_dataset("starhopp3r/TinyChat")
print(
dataset["train"][100]["text"][:500]
) # Print the first 500 characters of the first article
print(dataset["train"][600000])
tokenizer = tiktoken.get_encoding("gpt2")
base_encoding = tiktoken.get_encoding("gpt2")
special_tokens = {
"[INST]": base_encoding.n_vocab, # next available token id
"[/INST]": base_encoding.n_vocab + 1,
}
# 3. Create a new encoding that merges GPT‑2’s tokens + your special tokens
tokenizer = tiktoken.Encoding(
name="gpt2_with_inst",
pat_str=base_encoding._pat_str,
mergeable_ranks=base_encoding._mergeable_ranks,
special_tokens={**base_encoding._special_tokens, **special_tokens},
)
def encode(text):
return tokenizer.encode(text, allowed_special={"[INST]", "[/INST]"})
def decode(tokens):
return tokenizer.decode(tokens)
print("testing encoding and decoding functions:")
print(encode("[INST] Hello, world! [/INST]"))
print(decode(encode("[INST] Hello, world! [/INST]")))
# hyperparameters
train_model = True
periodic_outputs = False
block_size = 128
n_layers = 16
n_heads = 8
dropout_p = 0.1
batch_size = 64
learning_rate = 3e-4
n_embedding = 256
max_iters = 400000
device = "cuda" if torch.cuda.is_available() else "cpu"
train_dataset = TextDataset(dataset, block_size=block_size)
train_dataloader = DataLoader(
train_dataset, batch_size=batch_size, shuffle=True, drop_last=True, num_workers=16
)
# define objects
vocab_size = tokenizer.n_vocab
model = GPTModel(vocab_size, n_embedding, n_layers, n_heads, dropout_p, block_size).to(
device
)
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)
loss_fn = nn.CrossEntropyLoss()
# training loop
torch.set_float32_matmul_precision("high")
scaler = GradScaler(device)
if train_model:
compiled_model = torch.compile(model)
pbar = tqdm(range(max_iters), desc="Training", ncols=100)
data_iter = iter(train_dataloader)
for count in pbar:
try:
xb, yb = next(data_iter)
except StopIteration:
# dataloader exhausted — restart it
data_iter = iter(train_dataloader)
xb, yb = next(data_iter)
if count % 100 == 0 and periodic_outputs:
# print out xb, yb, encoded too
print("xb decoded: ", decode(xb[0].tolist()))
print("yb decoded: ", decode(yb[0].tolist()))
print("---" * 10)
print("xb raw: ", xb[0].tolist())
print("yb raw: ", yb[0].tolist())
#
# except StopIteration:
# break # dataloader exhausted before max_iters
xb, yb = xb.to(device), yb.to(device)
# logits = compiled_model(xb)
# loss = loss_fn(logits.view(-1, vocab_size), yb.view(-1))
# optimizer.zero_grad()
# loss.backward()
# optimizer.step()
with autocast(device, dtype=torch.float16):
logits = compiled_model(xb)
loss = loss_fn(logits.view(-1, vocab_size), yb.view(-1))
# backward pass with gradient scaling
optimizer.zero_grad()
scaler.scale(loss).backward()
scaler.step(optimizer)
scaler.update()
# update bar text dynamically
pbar.set_postfix({"loss": f"{loss.item():.4f}"})
if train_model:
torch.save(model.state_dict(), "checkpoints/gpt_model-1.pth")
else:
model.load_state_dict(torch.load("checkpoints/gpt_model-1.pth"))
@torch.no_grad()
def generate_text(model, prompt, max_new_tokens, block_size, device):
model.eval()
# Encode the prompt text into token IDs using our custom encode function
tokens = torch.tensor(encode(prompt), dtype=torch.long).unsqueeze(0).to(device)
for _ in range(max_new_tokens):
# Only keep the last block_size tokens for context
input_tokens = tokens[:, -block_size:]
# Get logits and take the last token's distribution
logits = model(input_tokens)
logits = logits[:, -1, :] # (batch=1, vocab)
probs = F.softmax(logits, dim=-1)
# Sample from the distribution
next_token = torch.multinomial(probs, num_samples=1)
tokens = torch.cat((tokens, next_token), dim=1)
# Decode back into text using our custom decode function
output_tokens = tokens[0].tolist()
output_text = decode(output_tokens)
return output_text
# print model parameters
print(
f"Model has {sum(p.numel() for p in model.parameters()) / 1000000:.6f} million parameters."
)
prompt = "this new company does [/INST]"
print(
generate_text(
model, prompt, max_new_tokens=500, block_size=block_size, device=device
)
)
|