Blame for scripts/eval_quantized.py

#!/usr/bin/env python3
"""Teacher-forced evaluation: BF16 vs quantized model on a fixed corpus.

Computes, over fixed token chunks:
  * Perplexity (independent chunks; identical chunking for both models)
  * NLL delta per token
  * Top-1 / Top-5 agreement
  * KL(BF16 || candidate)
  * Cosine similarity of logits
  * Top-1 logit margin change

No sampling, no generation: the same token sequence is fed to both models.

Usage:
  JAX_PLATFORM_NAME=cpu python scripts/eval_quantized.py \
    --bf16-dir /tmp/qwen35_0.8b_bf16_lalamo_lalamo_origcfg \
    --candidate-dir /tmp/qwen35_0.8b_rht_int4_lalamo_lalamo_origcfg \
    --corpus experiments/qwen35-0.8b/data/wikitext2_sample.txt \
    --max-sequences 100 --max-tokens 128 \
    --json-out experiments/qwen35-0.8b/rtn-int4-g32/eval_wikitext2.json
"""

from __future__ import annotations

import argparse
import hashlib
import json
import math
import os
import sys
import time
from pathlib import Path

os.environ.setdefault("JAX_PLATFORM_NAME", "cpu")


def parse_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser()
    parser.add_argument("--bf16-dir", type=Path, required=True)
    parser.add_argument("--candidate-dir", type=Path, required=True)
    parser.add_argument("--corpus", type=Path, required=True, help="UTF-8 text file")
    parser.add_argument("--lalamo-path", default=os.environ.get("LALAMO_PATH", "/tmp/lalamo-direct"))
    parser.add_argument("--max-sequences", type=int, default=100)
    parser.add_argument("--max-tokens", type=int, default=128)
    parser.add_argument(
        "--dtype",
        choices=("bfloat16", "float32"),
        default="bfloat16",
        help="model dtype; T4 does not support BF16 matmul, use float32 there",
    )
    parser.add_argument("--min-tokens", type=int, default=16)
    parser.add_argument("--json-out", type=Path, default=None)
    parser.add_argument(
        "--progress-jsonl",
        type=Path,
        default=None,
        help="default: <json-out>.progress.jsonl or <candidate-dir>/eval_progress.jsonl",
    )
    return parser.parse_args()


def chunks_from_text(tokenizer, text: str, max_tokens: int, min_tokens: int, limit: int) -> list[list[int]]:
    ids = tokenizer.encode(text, add_special_tokens=True).ids
    chunks = [ids[start : start + max_tokens] for start in range(0, len(ids), max_tokens)]
    chunks = [chunk for chunk in chunks if len(chunk) >= min_tokens]
    return chunks[:limit]


def scored_chars(tokenizer, chunks: list[list[int]]) -> int:
    """Characters of the scored stream.

    Every chunk contributes len(chunk) - 1 predictions, so the context-only first
    token of each chunk is removed. Decoding the concatenation once keeps
    multi-byte characters intact.
    """
    flat = [token for chunk in chunks for token in chunk]
    full = tokenizer.decode(flat)
    skipped = sum(len(tokenizer.decode([chunk[0]])) for chunk in chunks if chunk)
    return len(full) - skipped


def bpc(ppl: float, tokens: int, chars: int) -> float | None:
    if tokens <= 0 or chars <= 0 or ppl <= 0:
        return None
    return math.log2(ppl) * tokens / chars


def main() -> None:
    args = parse_args()
    from lalamo_forward import bf16_block_reason, gpu_capabilities

    reason = bf16_block_reason(args.dtype, gpu_capabilities())
    if reason:
        raise SystemExit(reason)
    if args.progress_jsonl:
        progress_path = args.progress_jsonl
    elif args.json_out:
        progress_path = args.json_out.with_suffix(".progress.jsonl")
    else:
        progress_path = args.candidate_dir / "eval_progress.jsonl"
    progress_path.parent.mkdir(parents=True, exist_ok=True)
    run_id = time.strftime("%Y%m%d-%H%M%S")
    progress_path.write_text("", encoding="utf-8")
    eval_config = {
        "corpus": str(args.corpus),
        "corpus_sha256": hashlib.sha256(args.corpus.read_bytes()).hexdigest()[:16],
        "max_sequences": args.max_sequences,
        "max_tokens": args.max_tokens,
        "min_tokens": args.min_tokens,
    }
    config_hash = hashlib.sha256(json.dumps(eval_config, sort_keys=True).encode("utf-8")).hexdigest()[:16]

    def progress(record: dict) -> None:
        record = {"run_id": run_id, "config_hash": config_hash, **record}
        with open(progress_path, "a", encoding="utf-8") as file:
            file.write(json.dumps(record) + "\n")

    corpus_lines = [
        line.strip()
        for line in args.corpus.read_text(encoding="utf-8").splitlines()
        if line.strip()
    ]
    duplicate_lines = len(corpus_lines) - len(set(corpus_lines))
    if duplicate_lines:
        raise SystemExit(
            f"eval corpus has {duplicate_lines} duplicate non-empty lines; "
            "use scripts/dedup_lines.py first"
        )
    text = "\n".join(corpus_lines)

    sys.path.insert(0, args.lalamo_path)

    import jax
    import jax.numpy as jnp
    import numpy as np

    from lalamo.model import Model
    from lalamo.module import Keychain
    from lalamo.modules import DecoderForwardPassConfig, ForwardPassMode
    from lalamo_forward import apply_attention_env, describe, enter_mesh, force_float32_activations
    from lalamo.utils.sharding import ShardingConfig
    from tokenizers import Tokenizer

    for label, directory in (("bf16", args.bf16_dir), ("candidate", args.candidate_dir)):
        config_path = directory / "config.json"
        if not config_path.exists():
            raise SystemExit(f"{label}: missing {config_path}")
        lalamo_config = json.loads(config_path.read_text())
        token_codec = lalamo_config.get("token_codec_config", {})
        if "response_format" not in token_codec:
            raise SystemExit(
                f"{label}: {directory} is uzu-normalized; create a lalamo eval copy "
                "(scripts/make_lalamo_eval_copy.py)"
            )

    def tokenizer_hash(directory: Path) -> str:
        return hashlib.sha256((directory / "tokenizer.json").read_bytes()).hexdigest()

    if tokenizer_hash(args.bf16_dir) != tokenizer_hash(args.candidate_dir):
        raise SystemExit("bf16/candidate tokenizer.json mismatch")

    sharding = ShardingConfig.replicated()
    _mesh = enter_mesh(sharding)
    keychain = Keychain.init(seed=0, sharding_config=sharding)
    forward_pass_config = DecoderForwardPassConfig.for_inference(mode=ForwardPassMode.MULTI_TOKEN)
    forward_pass_config = force_float32_activations(apply_attention_env(forward_pass_config))
    print(f"attention: {describe()}", flush=True)

    print("loading BF16 model...")
    model_dtype = jnp.dtype(args.dtype)
    print(f"model dtype: {args.dtype}", flush=True)
    bf16 = Model.load(args.bf16_dir, sharding, dtype=model_dtype)
    print("loading candidate model...")
    candidate = Model.load(args.candidate_dir, sharding, dtype=model_dtype)
    tokenizer = Tokenizer.from_file(str(args.bf16_dir / "tokenizer.json"))
    print("loaded")

    text = "\n".join(corpus_lines)
    chunks = chunks_from_text(tokenizer, text, args.max_tokens, args.min_tokens, args.max_sequences)
    if not chunks:
        raise SystemExit("no chunks: corpus too small or min-tokens too high")
    print(f"chunks: {len(chunks)}  mean tokens: {sum(len(c) for c in chunks) / len(chunks):.1f}")

    def logits_for(model, ids: list[int]) -> np.ndarray:
        token_ids = jnp.asarray([ids], dtype=jnp.int32)
        positions = jnp.arange(len(ids), dtype=jnp.int32)[None, :]
        result = model.decoder(
            token_ids=token_ids,
            token_positions=positions,
            state=None,
            return_updated_state=False,
            lengths_without_padding=None,
            forward_pass_config=forward_pass_config,
            keychain=keychain.split(1)[0],
        )
        return np.asarray(result.logits[0], dtype=np.float32)

    total_nll_bf16 = 0.0
    total_nll_candidate = 0.0
    total_positions = 0
    top1_matches = 0
    top5_overlap_sum = 0.0
    kl_sum = 0.0
    cosine_sum = 0.0
    margin_delta_sum = 0.0
    started_at = time.time()

    for index, ids in enumerate(chunks):
        bf_logits = logits_for(bf16, ids)
        qa_logits = logits_for(candidate, ids)
        # Predict token i+1 from position i.
        bf_pred = bf_logits[:-1]
        qa_pred = qa_logits[:-1]
        targets = np.asarray(ids[1:], dtype=np.int64)

        bf_log_prob = bf_pred - bf_pred.max(axis=-1, keepdims=True)
        bf_log_prob = bf_log_prob - np.log(np.exp(bf_log_prob).sum(axis=-1, keepdims=True))
        qa_log_prob = qa_pred - qa_pred.max(axis=-1, keepdims=True)
        qa_log_prob = qa_log_prob - np.log(np.exp(qa_log_prob).sum(axis=-1, keepdims=True))

        bf_target_log_prob = bf_log_prob[np.arange(len(targets)), targets]
        qa_target_log_prob = qa_log_prob[np.arange(len(targets)), targets]
        total_nll_bf16 -= float(bf_target_log_prob.sum())
        total_nll_candidate -= float(qa_target_log_prob.sum())
        total_positions += len(targets)

        bf_top1 = bf_pred.argmax(axis=-1)
        qa_top1 = qa_pred.argmax(axis=-1)
        top1_matches += int((bf_top1 == qa_top1).sum())

        bf_sorted = np.argsort(-bf_pred, axis=-1)[:, :5]
        qa_sorted = np.argsort(-qa_pred, axis=-1)[:, :5]
        for bf_row, qa_row in zip(bf_sorted, qa_sorted):
            top5_overlap_sum += len(set(bf_row.tolist()) & set(qa_row.tolist())) / 5.0

        bf_prob = np.exp(bf_log_prob)
        qa_prob = np.exp(qa_log_prob)
        kl_sum += float(
            np.sum(bf_prob * (np.log(bf_prob + 1e-45) - np.log(qa_prob + 1e-45)), axis=-1).sum()
        )
        cosine_sum += float(
            (
                np.sum(bf_pred * qa_pred, axis=-1)
                / (np.linalg.norm(bf_pred, axis=-1) * np.linalg.norm(qa_pred, axis=-1) + 1e-45)
            ).sum()
        )

        bf_margin = np.sort(bf_pred, axis=-1)[:, -1] - np.sort(bf_pred, axis=-1)[:, -2]
        qa_margin = np.sort(qa_pred, axis=-1)[:, -1] - np.sort(qa_pred, axis=-1)[:, -2]
        margin_delta_sum += float(np.abs(bf_margin - qa_margin).sum())

        if total_positions:
            progress(
                {
                    "chunk": index + 1,
                    "chunks": len(chunks),
                    "tokens": total_positions,
                    "bf16_ppl": math.exp(total_nll_bf16 / total_positions),
                    "candidate_ppl": math.exp(total_nll_candidate / total_positions),
                    "delta_nll_per_token": (total_nll_candidate - total_nll_bf16) / total_positions,
                    "top1_agreement": top1_matches / total_positions,
                    "kl_bf16_candidate": kl_sum / total_positions,
                    "elapsed_s": time.time() - started_at,
                }
            )
        if (index + 1) % 10 == 0 or index + 1 == len(chunks):
            print(f"  {index + 1}/{len(chunks)} chunks")

    char_total = scored_chars(tokenizer, chunks)
    mean_nll_bf16 = total_nll_bf16 / total_positions
    mean_nll_candidate = total_nll_candidate / total_positions
    results = {
        "bf16_dir": str(args.bf16_dir),
        "candidate_dir": str(args.candidate_dir),
        "corpus": str(args.corpus),
        "config": eval_config,
        "config_hash": config_hash,
        # dtype changes absolute PPL (0.8B runs float32 on CPU/T4, 4B/9B run
        # bfloat16 on A100), so it must travel with the number.
        "eval_dtype": args.dtype,
        # CPU float32 and GPU float32 agree to 0.02%, but recording the device
        # makes every number self-describing.
        "device": jax.devices()[0].device_kind if jax.devices() else "unknown",
        "chunks": len(chunks),
        "tokens": total_positions,
        "chars": char_total,
        "tokens_per_char": (total_positions / char_total) if char_total else None,
        "bf16_ppl": math.exp(mean_nll_bf16),
        "candidate_ppl": math.exp(mean_nll_candidate),
        "bpc_bf16": bpc(math.exp(mean_nll_bf16), total_positions, char_total),
        "bpc_candidate": bpc(math.exp(mean_nll_candidate), total_positions, char_total),
        "delta_nll_per_token": mean_nll_candidate - mean_nll_bf16,
        "top1_agreement": top1_matches / total_positions,
        "top5_overlap": top5_overlap_sum / total_positions,
        "kl_bf16_candidate": kl_sum / total_positions,
        "cosine_logits": cosine_sum / total_positions,
        "mean_abs_top1_margin_delta": margin_delta_sum / total_positions,
    }
    print(json.dumps(results, indent=2))
    if args.json_out:
        args.json_out.parent.mkdir(parents=True, exist_ok=True)
        args.json_out.write_text(json.dumps(results, indent=2) + "\n")
        print(f"saved {args.json_out}")


if __name__ == "__main__":
    main()