scripts/12_extract_mlp_path.py

"""Stage 1e: the per-layer MLP write-vectors at the read position, as the
eleven path descriptors of Eusebi et al. (2026).

Decoders: a forward hook on every decoder layer's MLP captures its output at
the last prompt position (the answer position; left padding), for the same
prompts in the same split order as 11_extract_decoder_states.py, which is
checked by recomputing the prediction. Encoders: a hook on every layer's
feed-forward output projection captures it at the [CLS] position, in the
prediction parquet's row order, as 07_extract_layer_states.py does. Only the
descriptors are written (data/features/<cfg>/mlppath_<split>.npz); the
write-vectors themselves are not kept.

    python extraction/scripts/12_extract_mlp_path.py --config extraction/configs/qwen05_mmlu.yaml
"""
from __future__ import annotations

import argparse
import importlib.util
import sys
import time
from pathlib import Path

import numpy as np
import pandas as pd
import torch

ROOT = Path(__file__).resolve().parents[1]
REPO = ROOT.parent
sys.path.insert(0, str(ROOT)); sys.path.insert(0, str(REPO))
from src.utils import load_config, seed_everything, device_from_cfg  # noqa: E402
from cct_reproduce.grids import is_decoder  # noqa: E402
from cct_reproduce.mlp_path import write_features  # noqa: E402


def _load_11():
    spec = importlib.util.spec_from_file_location("dec11", ROOT / "scripts" / "11_extract_decoder_states.py")
    mod = importlib.util.module_from_spec(spec); spec.loader.exec_module(mod)
    return mod


class _Capture:
    """Forward hooks that keep the module output at one position."""

    def __init__(self, modules, position):
        self.buf, self.position = [None] * len(modules), position
        self.handles = [m.register_forward_hook(self._hook(i)) for i, m in enumerate(modules)]

    def _hook(self, i):
        def f(_m, _in, out):
            o = out[0] if isinstance(out, tuple) else out
            self.buf[i] = o[:, self.position, :].detach().float().cpu()
        return f

    def stack(self):
        return torch.stack(self.buf, dim=1).numpy()                 # (B, L, D)

    def close(self):
        for h in self.handles:
            h.remove()


@torch.no_grad()
def decoder(cfg, name, a):
    d11 = _load_11()
    df = d11.load_questions(cfg["data"])
    df["split"] = d11.partition(df, cfg["data"].get("fractions", [0.5, 0.2, 0.3]),
                                int(cfg["data"].get("split_seed", 0)))
    mcfg = cfg["model"]
    from transformers import AutoModelForCausalLM, AutoTokenizer
    tok = AutoTokenizer.from_pretrained(mcfg["pretrained_path"])
    tok.padding_side = "left"; tok.truncation_side = "left"
    if tok.pad_token is None:
        tok.pad_token = tok.eos_token
    dtype = {"bfloat16": torch.bfloat16, "float16": torch.float16,
             "float32": torch.float32}[mcfg.get("dtype", "bfloat16")]
    model = AutoModelForCausalLM.from_pretrained(mcfg["pretrained_path"], torch_dtype=dtype)
    model = model.to(cfg["inference"].get("device", "cuda")).eval()
    n_cls = int(df["choices"].apply(len).max())
    letter_ids = d11.letter_token_ids(tok, n_cls)
    template = mcfg.get("prompt", "The following is a multiple choice question about "
                        "{subject}.\n\n{question}\n{options}\n\nAnswer with the letter only.")
    df["prompt"] = [d11.build_prompt(tok, r, template, mcfg.get("system", None)) for _, r in df.iterrows()]
    df = df.set_index("idx")
    cap = _Capture([layer.mlp for layer in model.model.layers], position=-1)
    bs, max_length = int(cfg["inference"].get("batch_size", 8)), int(mcfg.get("max_length", 2048))
    for split in ("train", "validation", "test"):
        pdf = pd.read_parquet(REPO / "data" / "features" / name / f"predictions_{split}.parquet")
        sub = df.loc[pdf["qid"].values]
        prompts = sub["prompt"].tolist()
        N = len(prompts)
        lengths = np.array([len(tok.encode(p, add_special_tokens=False)) for p in prompts])
        order = np.argsort(-lengths)
        M = None; pred = np.zeros(N, np.int64); t0 = time.time()
        for b in range(0, N, bs):
            rows = order[b:b + bs]
            enc = tok([prompts[i] for i in rows], return_tensors="pt", padding=True,
                      truncation=True, max_length=max_length)
            enc = {k: v.to(model.device) for k, v in enc.items()}
            out = model(**enc)
            m = cap.stack()
            if M is None:
                M = np.zeros((N, m.shape[1], m.shape[2]), np.float16)
            M[rows] = m.astype(np.float16)
            last = out.logits[:, -1, :].float()
            lt = torch.stack([last[:, ids].max(1).values for ids in letter_ids], 1)
            pred[rows] = lt.argmax(1).cpu().numpy()
            if (b // bs) % 100 == 0:
                print(f"    {b + len(rows):6d}/{N}  {time.time() - t0:.0f}s", flush=True)
        agree = float((pred == pdf["pred"].values).mean())
        print(f"  [{split}] n={N}  M={M.shape}  prediction agreement with the cache {agree:.4f}  "
              f"{time.time() - t0:.0f}s", flush=True)
        if agree < 0.98:
            raise RuntimeError(f"{name} {split}: prompts or order differ from the cached run")
        write_features(name, split, M, pdf["idx"].values, REPO / "data")
    cap.close()


@torch.no_grad()
def encoder(cfg, name, a):
    from src.load_data import load_splits
    from src.load_model import load_classification_model, move
    from src.extract_attention import _build_dataloader
    device = device_from_cfg(cfg)
    splits_data = load_splits(cfg)
    model, tokenizer = load_classification_model(
        cfg["model"]["pretrained_path"], num_labels=cfg["model"].get("num_labels"),
        base_tokenizer=cfg["model"].get("base_tokenizer"),
        do_lower_case=cfg["model"].get("do_lower_case"),
        is_peft=cfg["model"].get("is_peft", False), base_model=cfg["model"].get("base_model"))
    model = move(model, device).eval()
    layers = model.base_model.encoder.layer
    cap = _Capture([layer.output.dense for layer in layers], position=0)
    for split in ("train", "validation", "test"):
        p = REPO / "data" / "features" / name / f"predictions_{split}.parquet"
        pdf = pd.read_parquet(p)
        idxs = pdf["idx"].tolist()
        raw = splits_data[split].reset_index(drop=True)
        raw_m = raw.loc[raw["idx"].isin(idxs)].set_index("idx").loc[idxs].reset_index()
        loader = _build_dataloader(raw_m, tokenizer, cfg["data"]["text_col"],
                                   cfg["data"]["label_col"], cfg["model"]["max_length"],
                                   cfg["inference"]["batch_size"])
        Ms, preds = [], []
        t0 = time.time()
        for batch in loader:
            o = model(input_ids=batch["input_ids"].to(device),
                      attention_mask=batch["attention_mask"].to(device))
            Ms.append(cap.stack().astype(np.float16)); preds.append(o.logits.argmax(-1).cpu().numpy())
        M = np.concatenate(Ms); pred = np.concatenate(preds)
        agree = float((pred == pdf["pred"].values).mean())
        print(f"  [{split}] n={len(M)}  M={M.shape}  prediction agreement {agree:.4f}  "
              f"{time.time() - t0:.0f}s", flush=True)
        if agree < 0.98:
            raise RuntimeError(f"{name} {split}: row order differs from the cached predictions")
        write_features(name, split, M, np.asarray(idxs), REPO / "data")
    cap.close()


def main():
    ap = argparse.ArgumentParser(description=__doc__.split("\n")[0])
    ap.add_argument("--config", required=True)
    ap.add_argument("--force", action="store_true")
    a = ap.parse_args()
    cfg = load_config(a.config)
    name = cfg["run_name"]
    if (REPO / "data" / "features" / name / "mlppath_test.npz").exists() and not a.force:
        print(f"[{name}] mlppath features already present"); return
    seed_everything(int(cfg.get("seed", 0)))
    (decoder if is_decoder(name) else encoder)(cfg, name, a)
    print(f"[{name}] DONE")


if __name__ == "__main__":
    main()