scripts/eval_polish_quip.py

#!/usr/bin/env python3
"""
Polish LLM Leaderboard evaluation for QuIP# Bielik-Q2-Sharp Variant A.

Custom wrapper that loads QuIP# model via quip-sharp and runs eval
through speakleash/lm-evaluation-harness (polish3 branch).

Task groups:
  - polish_generate_few  (5-shot generative: polemo2, 8tags, cbd, ppc, psc)
  - polish_mc            (5-shot multiple choice variants)

Usage:
  python eval_polish_quip.py \
    --model_path /dev/shm/eval/model \
    --tokenizer speakleash/Bielik-11B-v2.3-Instruct \
    --output_dir /dev/shm/eval/results_a \
    --num_fewshot 5
"""
import sys
import os
import json
import time
import argparse

# Add quip-sharp to path BEFORE other imports
QUIP_DIR = os.environ.get('QUIP_DIR', '/dev/shm/eval/quip-sharp')
sys.path.insert(0, QUIP_DIR)

import torch

# PyTorch 2.10+ changed torch.load default to weights_only=True
_orig_load = torch.load
def _compat_load(*a, **kw):
    kw.setdefault('weights_only', False)
    return _orig_load(*a, **kw)
torch.load = _compat_load
torch.set_grad_enabled(False)

import numpy as np
from transformers import AutoTokenizer

# quip-sharp model loading
from lib.utils.unsafe_import import model_from_hf_path

# lm-eval imports — detect API version
import lm_eval
from lm_eval import evaluator

# Try new API (v0.4.x) first, fall back to old (v0.3.x)
try:
    from lm_eval.api.model import LM as BaseLMClass
    API_VERSION = "new"
except ImportError:
    from lm_eval.base import BaseLM as BaseLMClass
    API_VERSION = "old"


def log(msg):
    print(f"[{time.strftime('%H:%M:%S')}] {msg}", flush=True)


class QuIPSharpLM(BaseLMClass):
    """
    lm-eval compatible wrapper for QuIP# quantized models.

    Supports both old (BaseLM) and new (LM) lm-eval APIs.
    Old API: implements _model_call / _model_generate (batching handled by BaseLM).
    New API: implements loglikelihood / loglikelihood_rolling / generate_until directly.
    """

    def __init__(self, model_path, tokenizer_path, batch_size=1, max_length=2048):
        super().__init__()
        log(f"Loading QuIP# model from {model_path}...")
        t0 = time.time()
        self._model, _ = model_from_hf_path(model_path, use_cuda_graph=False)
        self._model.eval()
        log(f"Model loaded in {time.time()-t0:.1f}s")

        self._tokenizer = AutoTokenizer.from_pretrained(tokenizer_path)
        if self._tokenizer.pad_token is None:
            self._tokenizer.pad_token = self._tokenizer.eos_token
        log(f"Tokenizer: {tokenizer_path} (vocab={self._tokenizer.vocab_size})")

        self._batch_size = batch_size
        self._max_length = max_length
        self._device = torch.device("cuda")

    # ─── Properties (both APIs) ─────────────────────────────────
    @property
    def eot_token_id(self):
        return self._tokenizer.eos_token_id

    @property
    def max_length(self):
        return self._max_length

    @property
    def max_gen_toks(self):
        return 64

    @property
    def batch_size(self):
        return self._batch_size

    @property
    def device(self):
        return self._device

    @property
    def rank(self):
        return 0

    @property
    def world_size(self):
        return 1

    @property
    def tokenizer_name(self):
        return self._tokenizer.name_or_path

    def tok_encode(self, string, **kwargs):
        return self._tokenizer.encode(string, add_special_tokens=False)

    def tok_decode(self, tokens, **kwargs):
        return self._tokenizer.decode(tokens)

    # ─── Old API (BaseLM) ──────────────────────────────────────
    def _model_call(self, inps):
        """Forward pass — used by BaseLM for loglikelihood."""
        with torch.no_grad():
            return self._model(inps.to(self._device)).logits

    def _model_generate(self, context, max_length, eos_token_id):
        """Generate — used by BaseLM for generate_until."""
        with torch.no_grad():
            return self._model.generate(
                context.to(self._device),
                max_length=max_length,
                eos_token_id=eos_token_id,
                do_sample=False,
            )

    # ─── New API (LM v0.4.x) — batched ─────────────────────────
    def _encode_pair(self, ctx, cont):
        """Encode context+continuation, return (full_tokens, cont_length)."""
        ctx_enc = self._tokenizer.encode(ctx, add_special_tokens=False)
        cont_enc = self._tokenizer.encode(cont, add_special_tokens=False)
        full = ctx_enc + cont_enc
        if len(full) > self._max_length:
            full = full[-self._max_length:]
            cont_len = min(len(cont_enc), len(full))
        else:
            cont_len = len(cont_enc)
        return full, cont_len

    def loglikelihood(self, requests):
        """Compute log-likelihood with length-sorted batching for speed."""
        if API_VERSION == "old":
            return super().loglikelihood(requests)

        # Prepare all encodings
        encoded = []
        for req in requests:
            ctx, cont = req.args if hasattr(req, 'args') else req
            full, cont_len = self._encode_pair(ctx, cont)
            encoded.append((full, cont_len))

        total = len(encoded)
        results = [None] * total
        bs = max(self._batch_size, 8)  # Use at least 8 for length-sorted batching
        pad_id = self._tokenizer.pad_token_id or 0

        # Sort by sequence length for efficient batching (less padding waste)
        sorted_indices = sorted(range(total), key=lambda i: len(encoded[i][0]))

        log(f"  loglikelihood: {total} requests, batch_size={bs} (length-sorted)")
        lens = [len(encoded[i][0]) for i in sorted_indices]
        log(f"  sequence lengths: min={lens[0]}, max={lens[-1]}, "
            f"median={lens[len(lens)//2]}")
        t0 = time.time()
        processed = 0

        for batch_start in range(0, total, bs):
            batch_end = min(batch_start + bs, total)
            batch_indices = sorted_indices[batch_start:batch_end]
            batch = [encoded[i] for i in batch_indices]

            # Pad to same length within batch (minimal waste due to sorting)
            max_len = len(batch[-1][0])  # Last item is longest (sorted)

            input_ids = torch.full(
                (len(batch), max_len), pad_id,
                dtype=torch.long, device=self._device
            )
            attention_mask = torch.zeros(
                (len(batch), max_len),
                dtype=torch.long, device=self._device
            )

            for i, (tokens, _) in enumerate(batch):
                # Right-align (pad on left)
                offset = max_len - len(tokens)
                input_ids[i, offset:] = torch.tensor(tokens, dtype=torch.long)
                attention_mask[i, offset:] = 1

            with torch.no_grad():
                logits = self._model(
                    input_ids, attention_mask=attention_mask
                ).logits

            # Extract log probs for each item (vectorized)
            for i, (tokens, cont_len) in enumerate(batch):
                offset = max_len - len(tokens)
                seq_logits = logits[i, offset:]  # unpadded logits
                seq_ids = input_ids[i, offset:]

                shift_logits = seq_logits[:-1]
                shift_labels = seq_ids[1:]
                log_probs = torch.nn.functional.log_softmax(shift_logits, dim=-1)

                cont_start = len(tokens) - cont_len - 1
                if cont_start < 0:
                    cont_start = 0

                # Vectorized log prob computation
                cont_labels = shift_labels[cont_start:]
                cont_lps = log_probs[cont_start:]
                cont_log_prob = cont_lps[
                    torch.arange(len(cont_labels), device=self._device),
                    cont_labels
                ].sum().item()
                is_greedy = (
                    shift_logits[cont_start:].argmax(dim=-1) == cont_labels
                ).all().item()

                results[batch_indices[i]] = (cont_log_prob, is_greedy)

            processed += len(batch)
            if processed % (bs * 50) < bs:
                elapsed = time.time() - t0
                speed = processed / elapsed
                eta = (total - processed) / speed if speed > 0 else 0
                log(f"  loglikelihood: {processed}/{total} "
                    f"({speed:.1f} req/s, ETA {eta/60:.1f}min)")

        elapsed = time.time() - t0
        log(f"  loglikelihood done: {total} in {elapsed:.0f}s "
            f"({total/elapsed:.1f} req/s)")
        return results

    def loglikelihood_rolling(self, requests):
        """Compute full-string log-likelihood (for perplexity)."""
        if API_VERSION == "old":
            return super().loglikelihood_rolling(requests)

        results = []
        for req in requests:
            text = req.args[0] if hasattr(req, 'args') else req[0]
            enc = self._tokenizer.encode(text, add_special_tokens=False)
            if len(enc) > self._max_length:
                enc = enc[-self._max_length:]

            inp = torch.tensor([enc], device=self._device)
            with torch.no_grad():
                logits = self._model(inp).logits

            shift_logits = logits[0, :-1]
            shift_labels = inp[0, 1:]
            log_probs = torch.nn.functional.log_softmax(shift_logits, dim=-1)
            total_lp = sum(
                log_probs[i, shift_labels[i]].item()
                for i in range(len(shift_labels))
            )
            results.append(total_lp)
        return results

    def generate_until(self, requests):
        """Generate text with batched inference for speed."""
        if API_VERSION == "old":
            return super().generate_until(requests)

        total = len(requests)
        results = [None] * total
        bs = max(self._batch_size, 8)
        pad_id = self._tokenizer.pad_token_id or 0

        # Parse all requests
        parsed = []
        for idx, req in enumerate(requests):
            if hasattr(req, 'args'):
                ctx, gen_kwargs = req.args
            else:
                ctx, gen_kwargs = req
            until = gen_kwargs.get('until', [self._tokenizer.eos_token])
            if '\n' not in until:
                until = until + ['\n']
            max_gen = gen_kwargs.get('max_gen_toks', self.max_gen_toks)
            enc = self._tokenizer.encode(ctx, add_special_tokens=False)
            if len(enc) > self._max_length - max_gen:
                enc = enc[-(self._max_length - max_gen):]
            parsed.append((enc, until, max_gen))

        # Sort by length for efficient batching
        sorted_indices = sorted(range(total), key=lambda i: len(parsed[i][0]))

        lens = [len(parsed[i][0]) for i in sorted_indices]
        t0 = time.time()
        log(f"  generate_until: {total} requests, batch_size={bs} (length-sorted)")
        log(f"  context lengths: min={lens[0]}, max={lens[-1]}, "
            f"median={lens[len(lens)//2]}, max_gen_toks={self.max_gen_toks}")
        processed = 0

        for batch_start in range(0, total, bs):
            batch_end = min(batch_start + bs, total)
            batch_indices = sorted_indices[batch_start:batch_end]
            batch = [parsed[i] for i in batch_indices]

            # Use the max_gen from the first item (should be same for all)
            max_gen = batch[0][2]

            # Pad contexts to same length (left-pad)
            max_ctx_len = max(len(enc) for enc, _, _ in batch)
            input_ids = torch.full(
                (len(batch), max_ctx_len), pad_id,
                dtype=torch.long, device=self._device
            )
            attention_mask = torch.zeros(
                (len(batch), max_ctx_len),
                dtype=torch.long, device=self._device
            )
            ctx_lengths = []
            for i, (enc, _, _) in enumerate(batch):
                offset = max_ctx_len - len(enc)
                input_ids[i, offset:] = torch.tensor(enc, dtype=torch.long)
                attention_mask[i, offset:] = 1
                ctx_lengths.append(len(enc))

            # Batched generate
            with torch.no_grad():
                out = self._model.generate(
                    input_ids,
                    attention_mask=attention_mask,
                    max_new_tokens=max_gen,
                    do_sample=False,
                    eos_token_id=self._tokenizer.eos_token_id,
                )

            # Extract generated text per item
            for i, (enc, until, _) in enumerate(batch):
                offset = max_ctx_len - len(enc)
                gen_start = max_ctx_len  # generated tokens start after context
                gen_tokens = out[i, gen_start:]
                text = self._tokenizer.decode(gen_tokens, skip_special_tokens=True)
                for stop in until:
                    if stop in text:
                        text = text[:text.index(stop)]
                results[batch_indices[i]] = text

            processed += len(batch)
            if processed % (bs * 10) < bs:
                elapsed = time.time() - t0
                speed = processed / elapsed * 60
                eta = (total - processed) / (processed / elapsed) if processed > 0 else 0
                log(f"  generate_until: {processed}/{total} "
                    f"({speed:.1f} req/min, ETA {eta/60:.1f}min)")

        elapsed = time.time() - t0
        log(f"  generate_until done: {total} in {elapsed:.0f}s "
            f"({total/elapsed*60:.1f} req/min)")
        return results


def main():
    parser = argparse.ArgumentParser(
        description="Polish LLM Leaderboard eval for QuIP# models"
    )
    parser.add_argument('--model_path', default='/dev/shm/eval/model',
                        help='Path to QuIP# model directory')
    parser.add_argument('--tokenizer', default='speakleash/Bielik-11B-v2.3-Instruct',
                        help='Tokenizer name or path')
    parser.add_argument('--output_dir', default='/dev/shm/eval/results_a',
                        help='Output directory for results')
    parser.add_argument('--batch_size', type=int, default=1,
                        help='Batch size for eval')
    parser.add_argument('--num_fewshot', type=int, default=5,
                        help='Number of few-shot examples')
    parser.add_argument('--tasks', nargs='+',
                        default=['polish_generate_few', 'polish_mc'],
                        help='Task groups to evaluate')
    args = parser.parse_args()

    os.makedirs(args.output_dir, exist_ok=True)

    log("=" * 60)
    log("  Polish LLM Leaderboard Eval")
    log("  Model: QuIP# Bielik-Q2-Sharp Variant A")
    log(f"  lm-eval API: {API_VERSION}")
    log(f"  Tasks: {args.tasks}")
    log(f"  Few-shot: {args.num_fewshot}")
    log("=" * 60)

    # Load model once
    model = QuIPSharpLM(
        model_path=args.model_path,
        tokenizer_path=args.tokenizer,
        batch_size=args.batch_size,
    )

    # Run all tasks in a single evaluate call
    log(f"\nRunning {len(args.tasks)} tasks...")
    t0 = time.time()

    try:
        results = evaluator.simple_evaluate(
            model=model,
            tasks=args.tasks,
            num_fewshot=args.num_fewshot,
            log_samples=True,
            batch_size=args.batch_size,
        )
    except TypeError as e:
        log(f"simple_evaluate TypeError ({e}), trying older signature...")
        results = evaluator.simple_evaluate(
            model=model,
            tasks=args.tasks,
            num_fewshot=args.num_fewshot,
            no_cache=True,
        )

    elapsed = time.time() - t0
    log(f"\nAll tasks completed in {elapsed:.0f}s")

    # Save full results
    out_file = os.path.join(args.output_dir, 'full_results.json')
    with open(out_file, 'w') as f:
        json.dump(results, f, indent=2, default=str)
    log(f"Saved: {out_file}")

    # Print per-task summary
    all_results = {}
    if 'results' in results:
        for task_name, metrics in results['results'].items():
            log(f"\n  {task_name}:")
            for k, v in metrics.items():
                if isinstance(v, (int, float)):
                    log(f"    {k}: {v:.4f}" if isinstance(v, float) else f"    {k}: {v}")
            all_results[task_name] = metrics

    # Print final summary
    log("\n" + "=" * 60)
    log("  FINAL RESULTS SUMMARY")
    log("=" * 60)
    scores = []
    for group, tasks_res in all_results.items():
        for task_name, metrics in tasks_res.items():
            # Find the main accuracy metric
            for key in ['acc_norm', 'acc', 'f1', 'exact_match']:
                if key in metrics:
                    val = metrics[key]
                    if isinstance(val, (int, float)):
                        scores.append((task_name, key, val))
                        log(f"  {task_name}: {key}={val:.4f}")
                        break

    if scores:
        avg = np.mean([s[2] for s in scores])
        log(f"\n  Average score: {avg:.4f} ({avg*100:.2f}%)")
        log(f"  Baseline (IQ2_XXS): 61.34%")
        log(f"  FP16 Instruct:      65.71%")
        if avg * 100 > 61.34:
            log(f"  >>> BEATS BASELINE by {avg*100 - 61.34:.2f}pp <<<")
        else:
            log(f"  >>> Below baseline by {61.34 - avg*100:.2f}pp <<<")

    log("=" * 60)
    log("  EVALUATION COMPLETE")
    log("=" * 60)


if __name__ == '__main__':
    main()