quantization/hsaq/profile_sensitivity.py

"""
HSAQ Phase-1 Sensitivity Profiler — Hessian-diag + Activation Magnitude
(Note: HF Jobs UV-runner is at mxguru1/hsaq-sensitivity-profiles/runners/profile_granite_8b.py)
========================================================================

Standalone module per 2026-05-19 plan. Produces a JSON dict of
{layer_name: scores} for granite-3.3-8b (or any HF causal-LM) using a single
calibration pass.

Two signals combined per channel, then top-decile-mean'd to a layer score:
  - Hessian-diag proxy:    sum_i (x_i^2)     # input variance per channel
  - Activation magnitude:  max_i |x_i|       # AWQ-style outlier presence

Diagonal-only Hessian: state per layer is [in_features], not [in_features, in_features].
For ~280 linear layers on granite-8B at hidden=4096, peak hook state is ~16 MB,
not ~18 GB.

This is NOT wired into HSAQ assignment.py. Phase-1 deliverable is the JSON only —
integration shape (sensitivity floor / quantizer routing / weighted greedy) is
deferred until we eyeball the scores.

Run:
    python profile_sensitivity.py \\
        --model ibm-granite/granite-3.3-8b-instruct \\
        --calib-dataset mxguru1/master-chief-benign-calibration-v1 \\
        --n-samples 32 --seq-len 512 \\
        --out sensitivity_scores_granite_8b.json
"""

from __future__ import annotations

import argparse
import json
import logging
import re
import time
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Iterable

import torch
import torch.nn as nn

logger = logging.getLogger("hsaq.profile_sensitivity")
PIPELINE_VERSION = "profile_sensitivity.v0.1.0"

# Match projection layers inside transformer blocks. Excludes embed_tokens,
# lm_head, norms, and auxiliary Linear modules outside the block stack.
# Matches granite/llama/qwen/mistral families.
LAYER_NAME_RE = re.compile(r"\.layers\.\d+\.(self_attn|mlp)\.[a-z_]+_proj$")

# Below this in_features, top-decile aggregation is noisy (top-10% of 128
# channels = 12 channels). Fall back to max-pool for tiny layers.
MIN_IN_FEATURES_FOR_TOPK = 256


@dataclass
class LayerScore:
    layer_name: str
    layer_idx: int
    layer_type: str           # "q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"
    in_features: int
    hess_diag_score: float    # mean of normalized hessian-diag
    act_max_score: float      # mean of normalized act-max
    combined_score: float     # top-decile-mean of 0.6*hess + 0.4*act (or max for tiny)
    aggregator: str           # "topk_mean" or "max"
    finite: bool              # False if Inf/NaN encountered (clamped before scoring)
    n_token_positions: int    # how many token positions contributed to stats


def _parse_layer_idx(name: str) -> int:
    m = re.search(r"\.layers\.(\d+)\.", name)
    return int(m.group(1)) if m else -1


def _layer_type(name: str) -> str:
    return name.rsplit(".", 1)[-1]


def _select_target_layers(model: nn.Module) -> dict[str, nn.Linear]:
    targets: dict[str, nn.Linear] = {}
    for name, mod in model.named_modules():
        if not isinstance(mod, nn.Linear):
            continue
        if not LAYER_NAME_RE.search(name):
            continue
        targets[name] = mod
    return targets


def _safe_max(t: torch.Tensor) -> float:
    finite = t[torch.isfinite(t)]
    return float(finite.max().item()) if finite.numel() > 0 else 0.0


def _combine_to_layer_score(
    hess_diag: torch.Tensor,
    act_max: torch.Tensor,
    in_features: int,
    hess_weight: float = 0.6,
    act_weight: float = 0.4,
) -> tuple[float, float, float, str]:
    hess_norm = hess_diag / (hess_diag.max() + 1e-8)
    act_norm  = act_max  / (act_max.max()  + 1e-8)
    combined  = hess_weight * hess_norm + act_weight * act_norm

    hess_s = float(hess_norm.mean().item())
    act_s  = float(act_norm.mean().item())

    if in_features < MIN_IN_FEATURES_FOR_TOPK:
        score = float(combined.max().item())
        aggregator = "max"
    else:
        k = max(1, combined.numel() // 10)
        score = float(combined.topk(k).values.mean().item())
        aggregator = "topk_mean"

    return hess_s, act_s, score, aggregator


def profile_sensitivity_unified(
    model: nn.Module,
    calib_iter: Iterable[dict],
    n_samples: int,
    device: str = "cuda",
) -> list[LayerScore]:
    """Single calibration pass — all target layers profiled simultaneously."""
    targets = _select_target_layers(model)
    if not targets:
        raise RuntimeError(
            "No target layers matched the filter. "
            "Check LAYER_NAME_RE against this model's module names."
        )

    logger.info("Hooking %d target layers", len(targets))

    state: dict[str, dict] = {}
    for name, mod in targets.items():
        in_dim = mod.in_features
        state[name] = {
            "H_diag": torch.zeros(in_dim, device=device, dtype=torch.float32),
            "max":    torch.zeros(in_dim, device=device, dtype=torch.float32),
            "n_seen": 0,
            "in_features": in_dim,
        }

    handles = []

    def make_hook(layer_name: str):
        def hook(module, input_tuple, output):
            x = input_tuple[0].detach().to(torch.float32)
            x = x.reshape(-1, x.shape[-1])
            s = state[layer_name]
            s["H_diag"].add_((x * x).sum(dim=0))
            torch.maximum(s["max"], x.abs().amax(dim=0), out=s["max"])
            s["n_seen"] += x.shape[0]
        return hook

    # Gotcha (4.7): register hooks AFTER model.eval(); if accelerate.dispatch_model
    # was used, hooks must also be registered after dispatch. Caller's responsibility
    # — but for single-GPU bf16 (granite-8B on A100), no dispatch is involved.
    for name, mod in targets.items():
        handles.append(mod.register_forward_hook(make_hook(name)))

    try:
        model.eval()
        with torch.no_grad():
            consumed = 0
            for i, batch in enumerate(calib_iter):
                if i >= n_samples:
                    break
                input_ids = batch["input_ids"].to(device)
                model(input_ids=input_ids)
                consumed = i + 1
                if consumed % 8 == 0:
                    logger.info("Calibration batch %d/%d", consumed, n_samples)
            logger.info("Calibration loop consumed %d/%d batches", consumed, n_samples)
    finally:
        for h in handles:
            h.remove()

    scores: list[LayerScore] = []
    for name, s in state.items():
        n_seen = max(s["n_seen"], 1)
        H_diag = s["H_diag"] / n_seen
        act_max = s["max"]

        # Gotcha (4.7): NaN/Inf check. Outlier activations squared can overflow
        # even in fp32 accumulator. Clamp before scoring; flag in output.
        finite_h = torch.isfinite(H_diag).all().item()
        finite_a = torch.isfinite(act_max).all().item()
        finite = bool(finite_h and finite_a)
        if not finite:
            logger.warning("Non-finite values in %s — clamping (h=%s a=%s)", name, finite_h, finite_a)
            H_diag = torch.nan_to_num(H_diag, nan=0.0, posinf=_safe_max(H_diag), neginf=0.0)
            act_max = torch.nan_to_num(act_max, nan=0.0, posinf=_safe_max(act_max), neginf=0.0)

        hess_s, act_s, combined_s, aggregator = _combine_to_layer_score(
            H_diag.cpu(), act_max.cpu(), s["in_features"]
        )
        scores.append(LayerScore(
            layer_name=name,
            layer_idx=_parse_layer_idx(name),
            layer_type=_layer_type(name),
            in_features=s["in_features"],
            hess_diag_score=hess_s,
            act_max_score=act_s,
            combined_score=combined_s,
            aggregator=aggregator,
            finite=finite,
            n_token_positions=s["n_seen"],
        ))

    scores.sort(key=lambda r: (r.layer_idx, r.layer_type))
    return scores


def _build_calib_iter(dataset, tokenizer, seq_len: int, batch_size: int):
    """Concatenate dataset[text] into a token stream, chunk into seq_len blocks,
    yield as batches of input_ids."""
    eos = tokenizer.eos_token_id if tokenizer.eos_token_id is not None else 0
    buf: list[int] = []
    for row in dataset:
        text = row.get("text") or row.get("prompt") or ""
        if not text:
            continue
        ids = tokenizer(text, add_special_tokens=False)["input_ids"]
        buf.extend(ids)
        buf.append(eos)

    n_full = (len(buf) // seq_len) * seq_len
    blocks = [buf[i:i + seq_len] for i in range(0, n_full, seq_len)]
    logger.info("Built %d blocks of %d tokens from %d total tokens", len(blocks), seq_len, len(buf))

    for i in range(0, (len(blocks) // batch_size) * batch_size, batch_size):
        chunk = blocks[i:i + batch_size]
        yield {"input_ids": torch.tensor(chunk, dtype=torch.long)}


def main():
    p = argparse.ArgumentParser()
    p.add_argument("--model", default="ibm-granite/granite-3.3-8b-instruct")
    p.add_argument("--calib-dataset", default="mxguru1/master-chief-benign-calibration-v1")
    p.add_argument("--calib-split", default="train")
    p.add_argument("--n-samples", type=int, default=32)
    p.add_argument("--seq-len", type=int, default=512)
    p.add_argument("--batch-size", type=int, default=1)
    p.add_argument("--out", default="sensitivity_scores.json")
    p.add_argument("--device", default="cuda")
    p.add_argument("--dtype", default="bfloat16", choices=["bfloat16", "float16", "float32"])
    args = p.parse_args()

    logging.basicConfig(
        level=logging.INFO,
        format="%(asctime)s %(levelname)s %(name)s: %(message)s",
    )

    from datasets import load_dataset
    from transformers import AutoModelForCausalLM, AutoTokenizer

    dtype = {"bfloat16": torch.bfloat16, "float16": torch.float16, "float32": torch.float32}[args.dtype]

    logger.info("Loading tokenizer: %s", args.model)
    tok = AutoTokenizer.from_pretrained(args.model)
    if tok.pad_token_id is None:
        tok.pad_token = tok.eos_token

    logger.info("Loading model: %s (dtype=%s, device=%s)", args.model, args.dtype, args.device)
    model = AutoModelForCausalLM.from_pretrained(args.model, torch_dtype=dtype)
    model.to(args.device)
    model.eval()

    logger.info("Loading calibration set: %s [%s]", args.calib_dataset, args.calib_split)
    ds = load_dataset(args.calib_dataset, split=args.calib_split)
    logger.info("Calibration set: %d rows", len(ds))

    calib_iter = _build_calib_iter(ds, tok, seq_len=args.seq_len, batch_size=args.batch_size)

    t0 = time.time()
    scores = profile_sensitivity_unified(
        model=model,
        calib_iter=calib_iter,
        n_samples=args.n_samples,
        device=args.device,
    )
    elapsed = time.time() - t0
    logger.info("Profiling complete in %.1fs (%d layers)", elapsed, len(scores))

    payload = {
        "pipeline_version": PIPELINE_VERSION,
        "model": args.model,
        "calibration_dataset": args.calib_dataset,
        "calibration_split": args.calib_split,
        "n_samples_requested": args.n_samples,
        "seq_len": args.seq_len,
        "batch_size": args.batch_size,
        "dtype": args.dtype,
        "elapsed_seconds": elapsed,
        "n_layers": len(scores),
        "layers": [asdict(s) for s in scores],
    }
    Path(args.out).write_text(json.dumps(payload, indent=2))
    logger.info("Wrote scores to %s", args.out)

    by_type: dict[str, list[float]] = {}
    for s in scores:
        by_type.setdefault(s.layer_type, []).append(s.combined_score)
    print("\n=== Mean combined_score by layer_type ===")
    for lt, vals in sorted(by_type.items(), key=lambda kv: -sum(kv[1]) / max(1, len(kv[1]))):
        mean = sum(vals) / len(vals)
        print(f"  {lt:12s}  mean={mean:.4f}  n={len(vals)}")

    top10 = sorted(scores, key=lambda r: -r.combined_score)[:10]
    bot10 = sorted(scores, key=lambda r: r.combined_score)[:10]
    print("\n=== Top 10 layers by combined_score ===")
    for s in top10:
        print(f"  {s.combined_score:.4f}  L{s.layer_idx:02d}  {s.layer_type}")
    print("\n=== Bottom 10 layers by combined_score ===")
    for s in bot10:
        print(f"  {s.combined_score:.4f}  L{s.layer_idx:02d}  {s.layer_type}")


if __name__ == "__main__":
    main()


# ── Phase-3a: floor builder ───────────────────────────────────────────────

def build_floor_from_scores_json(
    scores_path: str | Path,
    top_n: int | None = None,
    score_threshold: float | None = None,
    layer_type_filter: str | tuple[str, ...] | None = None,
    min_bits: int = 4,
) -> dict[str, int]:
    """Build a min_bits_floor map for assign_bit_widths from a sensitivity JSON.

    Parameters
    ----------
    scores_path : path to a sensitivity_scores_*.json produced by main()
    top_n : if set, pick top-N layers by combined_score (after layer_type_filter)
    score_threshold : if set, pick all layers with combined_score >= threshold
    layer_type_filter : keep only these layer types (e.g. ("o_proj",) or "o_proj")
    min_bits : the floor value to apply to picked layers

    Returns
    -------
    {component_name: min_bits} — keys match LayerCandidate.component (full module name)
    """
    if top_n is None and score_threshold is None:
        raise ValueError("Specify top_n or score_threshold")

    payload = json.loads(Path(scores_path).read_text())
    layers = payload["layers"]

    if layer_type_filter is not None:
        if isinstance(layer_type_filter, str):
            layer_type_filter = (layer_type_filter,)
        layers = [l for l in layers if l["layer_type"] in layer_type_filter]

    if score_threshold is not None:
        layers = [l for l in layers if l["combined_score"] >= score_threshold]

    if top_n is not None:
        layers = sorted(layers, key=lambda l: -l["combined_score"])[:top_n]

    return {l["layer_name"]: min_bits for l in layers}