tools/coverage.py

"""Vocabulary coverage and document-length distribution for a calibration file.

    python tools/coverage.py --gguf /workspace/gguf/base/Muse-Glimmer-30B-BF16.gguf \
        builds/muse-glimmer-30b/calib_train.txt

    python tools/coverage.py --backend hf \
        --tokenizer /workspace/models/muse-glimmer-30b/tokenizer.json calib_train.txt

Two backends, because they answer slightly different questions:

* `llama-cpp` runs `llama-tokenize` against the GGUF the imatrix will actually
  be computed from.  This is the authoritative number — it goes through the same
  vocabulary and the same `llama4` pre-tokenizer that `llama-imatrix` will use.
* `hf` uses the model's `tokenizer.json` in-process.  Much faster, and the two
  are expected to agree; `--compare` checks that they do on a sample.

Coverage is reported against embedding rows, since that is the thing an imatrix
either has statistics for or does not: a row no calibration token ever selects
gets no importance data, and the quantiser has nothing to protect it with.
"""

from __future__ import annotations

import argparse
import json
import os
import subprocess
import sys
from collections import Counter

HERE = os.path.dirname(os.path.abspath(__file__))
sys.path.insert(0, HERE)

import poollib as P

DEFAULT_LLAMA_TOKENIZE = "/workspace/src/llama.cpp/build/bin/llama-tokenize"
DEFAULT_GGUF = "/workspace/gguf/base/Muse-Glimmer-30B-BF16.gguf"


def tokenize_llama_cpp(path: str, gguf: str, binary: str) -> list[int]:
    """Token ids from llama-tokenize.

    That tool defaults to parse_special=true, so `<|start|>` and friends in the
    text become their own ids -- the same thing `llama-imatrix --parse-special`
    will do.
    """
    res = subprocess.run(
        [binary, "-m", gguf, "-f", path, "--ids", "--log-disable"],
        capture_output=True, text=True)
    if res.returncode != 0:
        raise SystemExit(f"llama-tokenize failed:\n{res.stderr[-3000:]}")
    out = res.stdout.strip()
    start = out.rfind("[")
    if start < 0:
        raise SystemExit(f"unexpected llama-tokenize output: {out[:300]!r}")
    return json.loads(out[start:])


def doc_lengths(text: str, tok, sep: str = "\n\n",
                manifest: str | None = None) -> list[int]:
    """Token count per document.

    Splitting on the blank-line separator is wrong for anything real: source
    files and prose both contain blank lines, so it reported 65,838 documents
    for a 3,356-document build and a p50 of 31 tokens. When the build manifest
    is available its `chars` column gives the exact boundaries, the same way
    pipeline/build.py wrote them.
    """
    if manifest and os.path.exists(manifest):
        recs = [json.loads(l) for l in open(manifest, encoding="utf-8") if l.strip()]
        # `chars` counts the characters that were written. Reading back with
        # universal newlines silently folds every CRLF into one character, so
        # the offsets drift and the split falls back to blank lines -- which is
        # how a 3,363-document build was reported as 65,838 documents.
        docs, pos = [], 0
        for i, r in enumerate(recs):
            docs.append(text[pos:pos + r["chars"]])
            pos += r["chars"] + (len(sep) if i < len(recs) - 1 else 0)
        if abs(pos - len(text)) <= 2:
            return [len(ids) for ids in tok.encode_batch(docs)]
        print(f"  ! {manifest} does not line up with the file "
              f"({pos} vs {len(text)} characters); falling back to separator split")
    docs = [d for d in text.split(sep) if d.strip()]
    return [len(ids) for ids in tok.encode_batch(docs)]


def percentiles(xs: list[int]) -> dict:
    if not xs:
        return {}
    s = sorted(xs)

    def q(p):
        return s[min(len(s) - 1, int(p * len(s)))]
    return {"p50": q(0.50), "p90": q(0.90), "p95": q(0.95), "p99": q(0.99),
            "min": s[0], "max": s[-1], "mean": round(sum(s) / len(s), 1)}


def report(name: str, ids: list[int], n_vocab: int, lengths: list[int] | None,
           label: str) -> dict:
    c = Counter(ids)
    ge1 = len(c)
    ge10 = sum(1 for v in c.values() if v >= 10)
    ge100 = sum(1 for v in c.values() if v >= 100)
    out = {
        "file": name,
        "backend": label,
        "tokens": len(ids),
        "vocab_rows": n_vocab,
        "coverage": {
            "seen_ge_1": {"ids": ge1, "percent": round(100.0 * ge1 / n_vocab, 3)},
            "seen_ge_10": {"ids": ge10, "percent": round(100.0 * ge10 / n_vocab, 3)},
            "seen_ge_100": {"ids": ge100, "percent": round(100.0 * ge100 / n_vocab, 3)},
            "unseen": {"ids": n_vocab - ge1,
                       "percent": round(100.0 * (n_vocab - ge1) / n_vocab, 3)},
        },
    }
    print(f"\n=== {name}  [{label}] ===")
    print(f"tokens: {len(ids):,}")
    print(f"vocabulary coverage (denominator = {n_vocab:,} embedding rows):")
    print(f"  seen >=1   : {ge1:>9,}  ({100.0*ge1/n_vocab:6.2f}%)")
    print(f"  seen >=10  : {ge10:>9,}  ({100.0*ge10/n_vocab:6.2f}%)")
    print(f"  seen >=100 : {ge100:>9,}  ({100.0*ge100/n_vocab:6.2f}%)")
    print(f"  unseen     : {n_vocab-ge1:>9,}  ({100.0*(n_vocab-ge1)/n_vocab:6.2f}%)")
    if lengths:
        p = percentiles(lengths)
        out["documents"] = len(lengths)
        out["document_tokens"] = p
        big = sum(1 for x in lengths if x >= 8192)
        big_tok = sum(x for x in lengths if x >= 8192)
        out["documents_ge_8k"] = {"documents": big,
                                  "percent_of_tokens": round(100.0 * big_tok / max(1, sum(lengths)), 2)}
        print(f"documents: {len(lengths):,}")
        print(f"  document tokens: p50={p['p50']:,}  p90={p['p90']:,}  "
              f"p95={p['p95']:,}  p99={p['p99']:,}  max={p['max']:,}")
        print(f"  docs >= 8k tokens: {big:,}  ({out['documents_ge_8k']['percent_of_tokens']}% of tokens)")
    return out


def main() -> int:
    ap = argparse.ArgumentParser()
    ap.add_argument("files", nargs="+")
    ap.add_argument("--backend", choices=("llama-cpp", "hf"), default="llama-cpp")
    ap.add_argument("--gguf", default=DEFAULT_GGUF)
    ap.add_argument("--llama-tokenize", default=DEFAULT_LLAMA_TOKENIZE)
    ap.add_argument("--tokenizer", default=None, help="tokenizer.json for the hf backend")
    ap.add_argument("--vocab-size", type=int, default=None)
    ap.add_argument("--sep", default="\n\n")
    ap.add_argument("--json-out", default=None)
    ap.add_argument("--manifest", default=None,
                    help="build manifest giving exact document boundaries; "
                         "defaults to <file>.manifest.jsonl beside the input")
    ap.add_argument("--compare", action="store_true",
                    help="tokenize with both backends and report disagreement")
    args = ap.parse_args()

    hf = P.TargetTokenizer(args.tokenizer, n_vocab=args.vocab_size) if args.tokenizer else None
    n_vocab = args.vocab_size or (hf.n_vocab if hf else 202048)

    results = []
    for path in args.files:
        text = open(path, encoding="utf-8", newline="").read()
        if args.backend == "hf":
            if hf is None:
                raise SystemExit("--tokenizer is required for --backend hf")
            ids = hf.encode(text)
            label = f"hf:{os.path.basename(args.tokenizer)}"
        else:
            ids = tokenize_llama_cpp(path, args.gguf, args.llama_tokenize)
            label = f"llama-tokenize:{os.path.basename(args.gguf)}"
        mf = args.manifest or os.path.splitext(path)[0] + ".manifest.jsonl"
        lengths = doc_lengths(text, hf, args.sep, mf) if hf else None
        results.append(report(path, ids, n_vocab, lengths, label))

        if args.compare and hf is not None and args.backend != "hf":
            hf_ids = hf.encode(text)
            same = hf_ids == ids
            print(f"  backend agreement: {'identical' if same else 'DIFFER'} "
                  f"({len(ids):,} vs {len(hf_ids):,} tokens)")
            results[-1]["backend_agreement"] = {
                "identical": same, "llama_cpp_tokens": len(ids),
                "hf_tokens": len(hf_ids),
                # llama-tokenize prepends BOS; one extra token is expected
                "difference": len(ids) - len(hf_ids)}

    if args.json_out:
        with open(args.json_out, "w", encoding="utf-8") as f:
            json.dump(results, f, indent=2)
            f.write("\n")
        print(f"\nwrote {args.json_out}")
    return 0


if __name__ == "__main__":
    sys.exit(main())