File size: 4,515 Bytes
ed2140a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""Inference script for Diffusion-LM Riddle Solver (Hugging Face model).

Usage:
    python3 inference.py --riddle "i speak without a mouth what am i"
"""

import json
import sys
import argparse
import warnings


def load_model(model_dir: str = "."):
    """Load model weights and config from HF model directory."""
    import torch
    
    with open(f"{model_dir}/config.json") as f:
        config = json.load(f)
    
    with open(f"{model_dir}/vocab.json") as f:
        vocab = json.load(f)
    
    inv_vocab = {int(v): k for k, v in vocab.items()}
    
    # Add riddle_diffusion.py to Python path if running from HF directory
    sys.path.insert(0, model_dir)
    from riddle_diffusion import DiffusionRiddleModel
    from riddle_diffusion import get_schedule
    
    model = DiffusionRiddleModel(
        vocab_size=config["vocab_size"],
        d_model=config["d_model"],
        n_layers=config["n_layers"],
        d_ff=config["d_ff"],
        n_heads=config["n_heads"],
        a_len=config["a_len"],
        q_len=config["q_len"],
        T=config["T"],
    )
    
    state = torch.load(f"{model_dir}/model.safetensors", map_location="cpu",
                       weights_only=True)
    model.load_state_dict(state)
    model.eval()
    
    return model, config, vocab, inv_vocab


def predict(model, config, vocab, inv_vocab, riddle: str, k_samples: int = 10,
            device: str = "cpu"):
    """Run prediction on a single riddle."""
    import torch
    import torch.nn.functional as F
    
    model.to(device)
    
    # Tokenize
    tokens = [vocab.get(w, vocab.get("<UNK>", 1)) for w in riddle.lower().split()]
    if len(tokens) > config["q_len"]:
        tokens = tokens[:config["q_len"]]
    
    q_tokens = torch.tensor([tokens], device=device)
    
    # Diffusion schedule (sqrt power law)
    betas = torch.sqrt(torch.linspace(1e-4, 0.02, config["T"])).to(device)
    alphas = 1.0 - betas
    alpha_bars = torch.cumprod(alphas, dim=0)
    
    # Reverse diffusion
    x_t = torch.randn(k_samples, config["a_len"], config["d_model"], device=device)
    
    for t in reversed(range(config["T"])):
        t_tensor = torch.full((k_samples,), t, device=device, dtype=torch.long)
        pred_x0 = model(x_t, t_tensor, q_tokens)
        
        # Euclidean clamping
        logits = 2.0 * F.linear(pred_x0, model.emb.weight)
        logits = logits - model.emb.weight.square().sum(dim=-1).unsqueeze(0).unsqueeze(0)
        x0_tokens = logits.argmax(dim=-1)
        x0_emb = model.emb(x0_tokens)
        
        if t > 0:
            alpha_bar = alpha_bars[t]
            alpha_bar_prev = alpha_bars[t - 1]
            beta_tilde = betas[t] * (1 - alpha_bar_prev) / (1 - alpha_bar)
            noise = torch.randn_like(x_t)
            coef1 = torch.sqrt(alpha_bar_prev) * betas[t] / (1 - alpha_bar)
            coef2 = torch.sqrt(alpha_bar) * (1 - alpha_bar_prev) / (1 - alpha_bar)
            mu = coef1 * x0_emb + coef2 * x_t
            x_t = mu + torch.sqrt(beta_tilde) * noise
        else:
            x_t = x0_emb
    
    # Decode
    pred_tokens = []
    for b in range(k_samples):
        pred = ""
        for pos in range(config["a_len"]):
            tok_id = x0_tokens[b, pos].item()
            if tok_id == 0:
                break
            pred += inv_vocab.get(tok_id, "?") + " "
        pred_tokens.append(pred.strip())
    
    # Majority vote
    from collections import Counter
    counts = Counter(pred_tokens)
    winner = counts.most_common(1)[0][0]
    
    return winner, pred_tokens


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument("--riddle", required=True, help="Riddle text")
    parser.add_argument("--model-dir", default=".", help="Model directory")
    parser.add_argument("--device", default="cpu", help="Device (cpu, mps, cuda)")
    parser.add_argument("--k-samples", type=int, default=10)
    args = parser.parse_args()
    
    model, config, vocab, inv_vocab = load_model(args.model_dir)
    answer, candidates = predict(model, config, vocab, inv_vocab,
                                 args.riddle, args.k_samples, args.device)
    
    print(f"Riddle:  {args.riddle}")
    print(f"Answer:  {answer}")
    if len(set(candidates)) > 1:
        from collections import Counter
        counts = Counter(candidates)
        print(f"Candidates ({args.k_samples} samples):")
        for text, count in counts.most_common():
            print(f"  {text:<20} ({count} votes)")


if __name__ == "__main__":
    main()