File size: 6,815 Bytes
9477b5c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
175
176
"""Identifiability smoke test — pre-registered architectural decision gate.

Take a fixed batch of N=32 GSM8K problems. Freeze the base model. Train ONLY
the latent projector + InfoNCE head with the InfoNCE loss alone, for up to
200 steps. Measure the retrieval accuracy of z↔y.

Pre-registered decision rule (before launching the 24h pilot):
    best retrieval_acc >= 0.70 within 200 steps  →  PASS  → launch pilot
    best retrieval_acc near chance (~1/N=0.031)  →  FAIL → architecture broken

Why this gate matters: a constant-z attractor is mechanically incompatible
with high retrieval accuracy (InfoNCE on a batch of B identical z gives
loss = log B by construction). So PASS proves the architecture is at least
capable of producing identifiable latents — a necessary condition for the
pilot to be worth running.

Cost: ~5-10 minutes on a single GH200. Cheap gate for a 24h commitment.
"""
from __future__ import annotations

import argparse
import json
import os
import time
from pathlib import Path

import torch

from .data import GSM8KDataset, collate_batch
from .losses import InfoNCEHead, infonce_loss
from .model import BLTConfig, LatentProjector, build_base, forward_with_latent


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--config", required=True, help="reuse the pilot config (keys: base_model, K_latents, etc.)")
    ap.add_argument("--n_problems", type=int, default=32)
    ap.add_argument("--n_steps", type=int, default=200)
    ap.add_argument("--lr", type=float, default=1e-3)
    ap.add_argument("--tau", type=float, default=0.1)
    ap.add_argument("--threshold", type=float, default=0.70)
    ap.add_argument("--out", default=None)
    args = ap.parse_args()

    with open(args.config) as f:
        cfg = json.load(f)

    out_dir = Path(args.out or os.path.join(cfg["output_dir"], "smoke"))
    out_dir.mkdir(parents=True, exist_ok=True)
    log_path = out_dir / "smoke_log.txt"
    summary_path = out_dir / "summary.json"

    torch.manual_seed(cfg.get("seed", 42))
    device = "cuda" if torch.cuda.is_available() else "cpu"

    blt_cfg = BLTConfig(
        base_model=cfg["base_model"],
        use_lora=cfg.get("use_lora", False),
        lora_r=cfg.get("lora_r", 16), lora_alpha=cfg.get("lora_alpha", 32),
        lora_dropout=cfg.get("lora_dropout", 0.05),
        lora_target_modules=tuple(cfg.get("lora_target_modules",
                                          ("q_proj", "k_proj", "v_proj", "o_proj"))),
        K_latents=cfg["K_latents"], block_y_to_x=cfg.get("block_y_to_x", True),
        proj_init_scale=cfg.get("proj_init_scale", 0.02),
        dtype=cfg.get("dtype", "bfloat16"),
        attn_impl=cfg.get("attn_impl", "eager"),
    )
    model, tok = build_base(blt_cfg)
    model.to(device)

    # Freeze the base model so the only trainable params are projector + head.
    for p in model.parameters():
        p.requires_grad_(False)
    model.eval()

    inner = model.get_base_model() if hasattr(model, "get_base_model") else model
    d_model = inner.config.hidden_size
    dtype = getattr(torch, blt_cfg.dtype)

    projector = LatentProjector(d_model, init_scale=blt_cfg.proj_init_scale).to(device=device, dtype=dtype)
    head = InfoNCEHead(d_z=d_model, d_y=d_model, d_out=cfg.get("nce_proj_dim", 256)).to(device=device, dtype=dtype)

    # Fixed batch — same problems every step (true identifiability test).
    ds = GSM8KDataset(split="train", max_examples=args.n_problems)
    batch = collate_batch(
        [ds[i] for i in range(args.n_problems)], tok,
        max_prompt_len=cfg.get("max_prompt_len", 256),
        max_answer_len=cfg.get("max_answer_len", 256),
    )
    x_ids  = batch.x_ids.to(device)
    x_attn = batch.x_attn.to(device)
    y_ids  = batch.y_ids.to(device)
    y_attn = batch.y_attn.to(device)

    # Frozen-base y encoding (target for InfoNCE positives).
    with torch.no_grad():
        from .losses import encode_answer_for_infonce
        # We can also just use the y_ids end-of-answer text; here we feed gold "#### N" strings.
        f_y = encode_answer_for_infonce(model, tok, batch.final_strs, device=device, max_len=16)

    opt = torch.optim.AdamW(
        list(projector.parameters()) + list(head.parameters()),
        lr=args.lr, weight_decay=0.0,
    )

    K = blt_cfg.K_latents
    log = open(log_path, "w")
    t0 = time.time()
    best_acc = 0.0
    converged_step = None

    def _log(msg: str):
        line = f"[{time.time() - t0:6.1f}s] {msg}"
        print(line, flush=True)
        log.write(line + "\n"); log.flush()

    _log(f"smoke start: N={args.n_problems} K={K} steps={args.n_steps} thr={args.threshold}")
    _log(f"trainable proj+head params = "
         f"{sum(p.numel() for p in list(projector.parameters()) + list(head.parameters()))}")

    for step in range(args.n_steps):
        _, z, _ = forward_with_latent(
            model, x_ids, x_attn, y_ids, projector, K,
            block_y_to_x=blt_cfg.block_y_to_x,
        )
        z_pool = z.mean(dim=1).float()             # [B, d]
        z_emb, y_emb = head(z_pool, f_y.float())   # both L2-normalized
        loss = infonce_loss(z_emb, y_emb, tau=args.tau)

        # Diagnostic: nearest-neighbor retrieval accuracy.
        with torch.no_grad():
            sims = z_emb @ y_emb.t()
            preds = sims.argmax(dim=-1)
            acc = float((preds == torch.arange(sims.size(0), device=device)).float().mean().item())

        opt.zero_grad(set_to_none=True)
        loss.backward()
        torch.nn.utils.clip_grad_norm_(
            list(projector.parameters()) + list(head.parameters()), 1.0,
        )
        opt.step()

        if acc > best_acc:
            best_acc = acc
        if converged_step is None and acc >= args.threshold:
            converged_step = step

        if step % 10 == 0 or step == args.n_steps - 1:
            _log(f"step={step:4d}  loss={loss.item():.3f}  retr_acc={acc:.3f}  best={best_acc:.3f}")

        # Early stop after sustained convergence to save time.
        if converged_step is not None and step >= converged_step + 20:
            _log("converged + 20 buffer steps reached, stopping early.")
            break

    chance = 1.0 / args.n_problems
    decision = "PASS" if best_acc >= args.threshold else "FAIL"
    summary = {
        "N": args.n_problems, "K": K,
        "steps_run": step + 1,
        "best_retr_acc": best_acc,
        "converged_step": converged_step,
        "threshold": args.threshold,
        "chance": chance,
        "decision": decision,
        "duration_s": time.time() - t0,
    }
    summary_path.write_text(json.dumps(summary, indent=2))
    _log(f"summary: {summary}")
    log.close()
    print(f"[smoke] decision={decision}  best_acc={best_acc:.3f}  chance={chance:.4f}")


if __name__ == "__main__":
    main()