scripts/full_cloud_eval.sh

#!/bin/bash
set -e
WORKDIR=/workspace
HF_TOKEN="$1"
if [ -z "$HF_TOKEN" ]; then echo "Usage: bash full_cloud_eval.sh <HF_TOKEN>"; exit 1; fi
HF_REPO="Jakubrd4/bielik-quip-e8p12"
LIMIT=200
export HF_DATASETS_TRUST_REMOTE_CODE=1

echo "========================================"
echo "  QuIP# Bielik Eval - FULL AUTO SETUP"
echo "  RTX 4090 / A100 / H100 (NOT Blackwell)"
echo "========================================"
echo "Start: $(date)"
echo "GPU: $(python3 -c 'import torch; print(torch.cuda.get_device_name(0))' 2>/dev/null || echo 'unknown')"
echo ""

# ============================================
# 1. Clone QuIP#
# ============================================
echo "[1/8] Cloning QuIP#..."
cd $WORKDIR
if [ -d quip-sharp ]; then
    echo "  Already exists, skipping clone"
else
    git clone https://github.com/Cornell-RelaxML/quip-sharp.git
fi
cd quip-sharp

# ============================================
# 2. Apply patches
# ============================================
echo "[2/8] Applying patches..."
sed -i 's/from \.lm_eval_adaptor import.*/# disabled for lm-eval 0.4.x/' lib/utils/__init__.py
echo "  __init__.py patched"

python3 << 'PATCHPY'
path = 'lib/utils/unsafe_import.py'
with open(path) as f:
    code = f.read()
if 'from model.mistral' not in code:
    code = code.replace(
        'from model.llama import LlamaForCausalLM',
        'from model.llama import LlamaForCausalLM\nfrom model.mistral import MistralForCausalLM'
    )
if "model_type == 'mistral'" not in code:
    old = "    else:\n        raise Exception"
    new = "    elif model_type == 'mistral':\n        model_str = transformers.MistralConfig.from_pretrained(path)._name_or_path\n        model_cls = MistralForCausalLM\n    else:\n        raise Exception"
    code = code.replace(old, new)

# Also force eager attention (QuIP# fused qkv_proj breaks sdpa)
code = code.replace("attn_implementation='sdpa'", "attn_implementation='eager'")

with open(path, 'w') as f:
    f.write(code)
print('  unsafe_import.py patched for Mistral')
PATCHPY

python3 << 'PATCHPY2'
path = 'model/llama.py'
with open(path) as f:
    code = f.read()
old_line = "                    causal_mask = AttentionMaskConverter._unmask_unattended("
if old_line in code:
    new_block = """                    if hasattr(AttentionMaskConverter, '_unmask_unattended'):
                        causal_mask = AttentionMaskConverter._unmask_unattended(
                            causal_mask, min_dtype
                        )"""
    code = code.replace(
        old_line + "\n                        causal_mask, min_dtype\n                    )",
        new_block
    )
    with open(path, 'w') as f:
        f.write(code)
    print('  llama.py patched (_unmask_unattended)')
else:
    print('  llama.py: patch not needed or already applied')
PATCHPY2

# Patch: add rope_theta default for Mistral config
sed -i 's/self.rope_theta = config.rope_theta/self.rope_theta = getattr(config, "rope_theta", 1000000.0)/' model/mistral.py 2>/dev/null || true
echo "  rope_theta patched"

# ============================================
# 3. Fix Python dependencies
# ============================================
echo "[3/8] Fixing Python dependencies..."
pip install glog primefac protobuf 2>&1 | tail -3
pip install 'transformers==4.38.0' 2>&1 | tail -3
pip install 'datasets==2.20.0' 2>&1 | tail -3
# peft compatible with transformers 4.38
pip install 'peft==0.9.0' 2>&1 | tail -3
echo "  Dependencies fixed"

# ============================================
# 4. Compile QuIP# CUDA kernels
# ============================================
echo "[4/8] Compiling QuIP# CUDA kernels..."
cd $WORKDIR/quip-sharp/quiptools
pip install --no-build-isolation -e . 2>&1 | tail -5
echo "  quiptools installed"
echo "  Installing fast-hadamard-transform..."
pip install --no-build-isolation fast-hadamard-transform 2>&1 | tail -3 || {
    echo "  PyPI install failed, trying from git..."
    pip install --no-build-isolation git+https://github.com/Dao-AILab/fast-hadamard-transform.git 2>&1 | tail -3
}
echo "  fast-hadamard-transform installed"

# ============================================
# 5. Install lm-eval Polish fork
# ============================================
echo "[5/8] Installing lm-evaluation-harness (Polish fork)..."
cd $WORKDIR
if [ -d lm-evaluation-harness ]; then
    echo "  Already exists, skipping clone"
else
    git clone --branch polish4_shuf https://github.com/speakleash/lm-evaluation-harness.git
fi
cd lm-evaluation-harness
pip install -e . 2>&1 | tail -5
echo "  lm-eval installed"

# ============================================
# 6. Download model from HuggingFace
# ============================================
echo "[6/8] Downloading model from HuggingFace..."
python3 << DLPY
from huggingface_hub import snapshot_download
print("  Starting download...")
snapshot_download('${HF_REPO}', local_dir='${WORKDIR}/model', token='${HF_TOKEN}')
print("  Model downloaded!")
DLPY
echo "  Model files:"
ls -lh $WORKDIR/model/

# ============================================
# 7. Add rope_theta to model config if missing
# ============================================
echo "[7/8] Checking model config..."
python3 << 'CFGPY'
import json
p = '/workspace/model/config.json'
c = json.load(open(p))
changed = False
if 'rope_theta' not in c:
    c['rope_theta'] = 1000000.0
    changed = True
if changed:
    json.dump(c, open(p, 'w'), indent=2)
    print("  Added rope_theta to config")
else:
    print("  Config OK")
CFGPY

# ============================================
# 8. Create eval script and run
# ============================================
echo "[8/8] Creating eval script and running..."
cat > $WORKDIR/run_eval.py << 'PYEOF'
import sys, os, json, time, torch, argparse
sys.path.insert(0, "/workspace/quip-sharp")
torch.set_grad_enabled(False)
from transformers import AutoTokenizer
from lm_eval import evaluator
from lm_eval.models.huggingface import HFLM
from lib.utils.unsafe_import import model_from_hf_path

MC_TASKS = [
    "polemo2_in_multiple_choice", "polemo2_out_multiple_choice",
    "polish_8tags_multiple_choice", "polish_belebele_mc",
    "polish_dyk_multiple_choice", "polish_ppc_multiple_choice",
    "polish_psc_multiple_choice", "polish_cbd_multiple_choice",
    "polish_klej_ner_multiple_choice", "polish_polqa_reranking_multiple_choice",
]
PPL_TASKS = ["polish_poleval2018_task3_test_10k"]
BASELINES = {
    "polemo2_in_multiple_choice": 0.416, "polemo2_out_multiple_choice": 0.368,
    "polish_8tags_multiple_choice": 0.143, "polish_belebele_mc": 0.279,
    "polish_dyk_multiple_choice": 0.289, "polish_ppc_multiple_choice": 0.419,
    "polish_psc_multiple_choice": 0.466, "polish_cbd_multiple_choice": 0.149,
    "polish_klej_ner_multiple_choice": 0.343, "polish_polqa_reranking_multiple_choice": 0.534,
}

parser = argparse.ArgumentParser()
parser.add_argument("--limit", type=int, default=None)
parser.add_argument("--batch_size", type=int, default=1)
parser.add_argument("--model_path", type=str, default="/workspace/model")
args = parser.parse_args()

ALL_TASKS = MC_TASKS + PPL_TASKS
start = time.time()
lstr = str(args.limit) if args.limit else "FULL"
print("=" * 70)
print("Open PL LLM Leaderboard - QuIP# E8P12 2-bit Instruct")
print("Batch: %d | Limit: %s" % (args.batch_size, lstr))
print("GPU: %s" % torch.cuda.get_device_name(0))
print("=" * 70)

print("Loading model...")
model, model_str = model_from_hf_path(args.model_path, use_cuda_graph=False, use_flash_attn=False)
tokenizer = AutoTokenizer.from_pretrained(model_str)
tokenizer.pad_token = tokenizer.eos_token
lm = HFLM(pretrained=model, tokenizer=tokenizer, backend="causal", batch_size=args.batch_size, max_length=4096, trust_remote_code=True)

ekw = dict(model=lm, tasks=ALL_TASKS, num_fewshot=5, batch_size=args.batch_size, log_samples=False)
if args.limit:
    ekw["limit"] = args.limit

print("Running eval...")
results = evaluator.simple_evaluate(**ekw)

elapsed = time.time() - start
print("\n" + "=" * 70)
print("RESULTS (5-shot, limit=%s)" % lstr)
print("=" * 70)
scores = {}
nscores = {}
for t in ALL_TASKS:
    if t not in results.get("results", {}):
        print("  %-45s MISSING" % t)
        continue
    tr = results["results"][t]
    score = None
    metric = "?"
    for mk in ["acc,none", "f1,none", "word_perplexity,none"]:
        if mk in tr:
            score = tr[mk]
            metric = mk.split(",")[0]
            break
    if score is None:
        continue
    bl = BASELINES.get(t, 0)
    is_ppl = t in PPL_TASKS
    if is_ppl:
        norm = None
    elif 0 < bl < 1.0:
        norm = max(0, (score - bl) / (1.0 - bl))
    else:
        norm = max(0, score)
    scores[t] = score
    if norm is not None:
        nscores[t] = norm
    ns = "norm=%.4f" % norm if norm is not None else ""
    print("  %-45s %s=%.4f  %s" % (t, metric, score, ns))

print("-" * 70)
avg = sum(nscores.values()) / len(nscores) if nscores else 0
print("  %-45s %.4f  (%.2f%%)" % ("Avg MC (normalized)", avg, avg * 100))
print("=" * 70)
print("Time: %.1f min" % (elapsed / 60))
print("\nComparison:")
print("  SpeakLeash IQ2_XXS  = 61.34%%")
print("  FP16 baseline       = 65.71%%")
print("  QuIP# E8P12 2-bit   = %.2f%%" % (avg * 100))
os.makedirs("/workspace/eval_results", exist_ok=True)
fn = "/workspace/eval_results/results_limit%s.json" % (str(args.limit) if args.limit else "full")
json.dump({"avg_mc": float(avg), "scores": {k: float(v) for k,v in scores.items()}, "normalized": {k: float(v) for k,v in nscores.items()}, "full": results.get("results", {})}, open(fn, "w"), indent=2, default=str)
print("Saved to %s" % fn)
PYEOF
echo "  Eval script created"

echo "Running evaluation with limit=$LIMIT..."
echo "========================================"
cd $WORKDIR/quip-sharp
python3 $WORKDIR/run_eval.py --limit $LIMIT

echo ""
echo "========================================"
echo "  ALL DONE! $(date)"
echo "========================================"