Upload full_cloud_eval.sh with huggingface_hub

full_cloud_eval.sh

@@ -5,14 +5,20 @@ 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 ""
 
-echo "[1/7] Cloning QuIP#..."
+# ============================================
+# 1. Clone QuIP#
+# ============================================
+echo "[1/8] Cloning QuIP#..."
 cd $WORKDIR
 if [ -d quip-sharp ]; then
     echo "  Already exists, skipping clone"
@@ -21,7 +27,10 @@ else
 fi
 cd quip-sharp
 
-echo "[2/7] Applying patches..."
+# ============================================
+# 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"
 
@@ -38,6 +47,10 @@ 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')
@@ -64,15 +77,39 @@ else:
     print('  llama.py: patch not needed or already applied')
 PATCHPY2
 
-echo "[3/7] Compiling QuIP# CUDA kernels..."
+# 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
+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"
 
-echo "[4/7] Installing lm-evaluation-harness (Polish fork)..."
+# ============================================
+# 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"
@@ -83,7 +120,10 @@ cd lm-evaluation-harness
 pip install -e . 2>&1 | tail -5
 echo "  lm-eval installed"
 
-echo "[5/7] Downloading model from HuggingFace..."
+# ============================================
+# 6. Download model from HuggingFace
+# ============================================
+echo "[6/8] Downloading model from HuggingFace..."
 python3 << DLPY
 from huggingface_hub import snapshot_download
 print("  Starting download...")
@@ -93,7 +133,29 @@ DLPY
 echo "  Model files:"
 ls -lh $WORKDIR/model/
 
-echo "[6/7] Creating eval script..."
+# ============================================
+# 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")
@@ -197,7 +259,7 @@ print("Saved to %s" % fn)
 PYEOF
 echo "  Eval script created"
 
-echo "[7/7] Running evaluation with limit=$LIMIT..."
+echo "Running evaluation with limit=$LIMIT..."
 echo "========================================"
 cd $WORKDIR/quip-sharp
 python3 $WORKDIR/run_eval.py --limit $LIMIT