File size: 5,395 Bytes
3217baa
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Fourth GPT model definition and inference using PyTorch (CPU)."""

import torch
import torch.nn as nn
import torch.nn.functional as F
import math
import json
import os
import re


class RMSNorm(nn.Module):
    def __init__(self, dim, eps=1e-6):
        super().__init__()
        self.weight = nn.Parameter(torch.ones(dim))
        self.eps = eps

    def forward(self, x):
        norm = torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
        return x * norm * self.weight


class TransformerBlock(nn.Module):
    def __init__(self, n_embd, n_head):
        super().__init__()
        self.n_head = n_head
        self.head_dim = n_embd // n_head
        self.norm1 = RMSNorm(n_embd)
        self.wq = nn.Linear(n_embd, n_embd, bias=False)
        self.wk = nn.Linear(n_embd, n_embd, bias=False)
        self.wv = nn.Linear(n_embd, n_embd, bias=False)
        self.wo = nn.Linear(n_embd, n_embd, bias=False)
        self.norm2 = RMSNorm(n_embd)
        self.mlp_fc1 = nn.Linear(n_embd, 4 * n_embd, bias=False)
        self.mlp_fc2 = nn.Linear(4 * n_embd, n_embd, bias=False)

    def forward(self, x, mask):
        B, T, _ = x.shape
        xn = self.norm1(x)
        q = self.wq(xn).reshape(B, T, self.n_head, self.head_dim).transpose(1, 2)
        k = self.wk(xn).reshape(B, T, self.n_head, self.head_dim).transpose(1, 2)
        v = self.wv(xn).reshape(B, T, self.n_head, self.head_dim).transpose(1, 2)
        att = (q @ k.transpose(-2, -1)) / math.sqrt(self.head_dim)
        att = att + mask
        att = F.softmax(att, dim=-1)
        out = (att @ v).transpose(1, 2).reshape(B, T, -1)
        x = x + self.wo(out)
        xn2 = self.norm2(x)
        h = F.relu(self.mlp_fc1(xn2))
        x = x + self.mlp_fc2(h)
        return x


class GPT(nn.Module):
    def __init__(self, vocab_size, n_layer, n_embd, block_size, n_head):
        super().__init__()
        self.block_size = block_size
        self.wte = nn.Embedding(vocab_size, n_embd)
        self.wpe = nn.Embedding(block_size, n_embd)
        self.ln_pre = RMSNorm(n_embd)
        self.layers = nn.ModuleList([TransformerBlock(n_embd, n_head) for _ in range(n_layer)])
        self.ln_post = RMSNorm(n_embd)
        self.lm_head = nn.Linear(n_embd, vocab_size, bias=False)

    def forward(self, tokens):
        B, T = tokens.shape
        x = self.wte(tokens) + self.wpe(torch.arange(T, device=tokens.device))
        x = self.ln_pre(x)
        mask = torch.triu(torch.full((T, T), -1e9, device=tokens.device), diagonal=1)
        for layer in self.layers:
            x = layer(x, mask)
        x = self.ln_post(x)
        return self.lm_head(x)


class FourthModel:
    """Wraps the GPT model with tokenizer and generation logic."""

    def __init__(self, checkpoint_dir=None):
        if checkpoint_dir is None:
            checkpoint_dir = os.path.join(os.path.dirname(__file__) or ".", "model_weights")
        self.checkpoint_dir = checkpoint_dir
        self.model = None
        self.stoi = None
        self.itos = None
        self.bos = None
        self.config = None

    def load(self):
        config_path = os.path.join(self.checkpoint_dir, "config.json")
        with open(config_path) as f:
            self.config = json.load(f)

        self.stoi = self.config["stoi"]
        self.bos = self.config["bos"]
        self.itos = {int(i): c for c, i in self.stoi.items()}
        self.itos[self.bos] = ""

        self.model = GPT(
            vocab_size=self.config["vocab_size"],
            n_layer=self.config["n_layer"],
            n_embd=self.config["n_embd"],
            block_size=self.config["block_size"],
            n_head=self.config["n_head"],
        )

        # Load weights — try PyTorch format first, fall back to npz
        pt_path = os.path.join(self.checkpoint_dir, "weights.pt")
        npz_path = os.path.join(self.checkpoint_dir, "weights.npz")

        if os.path.exists(pt_path):
            state_dict = torch.load(pt_path, map_location="cpu", weights_only=True)
        else:
            import numpy as np
            npz = np.load(npz_path)
            state_dict = {k: torch.tensor(npz[k]) for k in npz.files}

        self.model.load_state_dict(state_dict)
        self.model.eval()

        nparams = sum(p.numel() for p in self.model.parameters())
        print(f"Loaded model: {nparams} params, vocab={self.config['vocab_size']}")

    @torch.no_grad()
    def generate(self, prompt: str, max_tokens: int = 128, temperature: float = 0.7) -> str:
        """Generate a response to a prompt."""
        clean = re.sub(r'[^a-z |]', '', prompt.lower().strip())
        clean = re.sub(r'  +', ' ', clean).strip()

        if not clean.endswith("|"):
            clean += "|"

        block_size = self.config["block_size"]
        tokens = [self.bos] + [self.stoi.get(c, self.bos) for c in clean]

        for _ in range(min(max_tokens, block_size - len(tokens))):
            x = torch.tensor([tokens[-block_size:]], dtype=torch.long)
            logits = self.model(x)
            logits = logits[0, -1] / temperature
            probs = F.softmax(logits, dim=-1)
            tok = torch.multinomial(probs, 1).item()
            if tok == self.bos:
                break
            tokens.append(tok)

        full = "".join(self.itos.get(t, "?") for t in tokens[1:])
        parts = full.split("|", 1)
        return parts[1] if len(parts) > 1 else full