1 files changed, 123 insertions, 0 deletions
diff --git a/collaborativeagents/scripts/test_rag_fix.sbatch b/collaborativeagents/scripts/test_rag_fix.sbatch
new file mode 100644
index 0000000..b07d286
--- /dev/null
+++ b/collaborativeagents/scripts/test_rag_fix.sbatch
@@ -0,0 +1,123 @@
+#!/bin/bash
+#SBATCH --job-name=test_rag_fix
+#SBATCH --account=bfqt-delta-gpu
+#SBATCH --partition=gpuH200x8
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=32
+#SBATCH --gres=gpu:4
+#SBATCH --mem=200G
+#SBATCH --time=02:00:00
+#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/test_rag_fix-%j.out
+#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/test_rag_fix-%j.err
+
+# Small-scale test: 5 profiles, 10 sessions each
+# Tests RAG fixes: extract_session accumulation, nopersonal mode
+# Compare: vanilla, rag, rag_vector
+
+cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents
+source /u/yurenh2/miniforge3/etc/profile.d/conda.sh
+conda activate eval
+export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
+export PYTHONPATH="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/src:$PYTHONPATH"
+
+PROFILE_PATH="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents/data/complex_profiles_v2/profiles_200.jsonl"
+AGENT_MODEL="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
+USER_MODEL="meta-llama/Llama-3.1-70B-Instruct"
+
+echo "=== RAG Fix Verification Test ==="
+echo "Testing fixes:"
+echo "  1. extract_session (accumulate all turns)"
+echo "  2. RAG mode=nopersonal (pure dense+rerank)"
+echo "  3. Explicit normalize=True"
+echo ""
+echo "Settings: 5 profiles, 10 sessions each, max 15 turns"
+echo "User simulator: $USER_MODEL (70B)"
+echo "Agent: $AGENT_MODEL (8B)"
+date
+nvidia-smi --query-gpu=index,name,memory.total --format=csv
+
+# Start vLLM servers
+# User simulator on GPUs 0,1 (70B, TP=2)
+echo ""
+echo "Starting vLLM servers..."
+CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
+    --model $USER_MODEL \
+    --port 8004 --tensor-parallel-size 2 --gpu-memory-utilization 0.90 \
+    --max-model-len 16384 --dtype bfloat16 --download-dir $HF_HOME \
+    --disable-log-requests &
+USER_PID=$!
+
+# Agent on GPUs 2,3 (8B, TP=2, lower memory for embedding/reranker)
+CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
+    --model $AGENT_MODEL \
+    --port 8003 --tensor-parallel-size 2 --gpu-memory-utilization 0.45 \
+    --max-model-len 16384 --dtype bfloat16 \
+    --disable-log-requests &
+AGENT_PID=$!
+
+# Wait for servers
+echo "Waiting for vLLM servers (may take 5-10 min)..."
+for i in {1..200}; do
+    if curl -s http://localhost:8004/health > /dev/null 2>&1; then
+        echo "User simulator (8004) ready after $((i*5)) seconds"
+        break
+    fi
+    sleep 5
+done
+for i in {1..60}; do
+    if curl -s http://localhost:8003/health > /dev/null 2>&1; then
+        echo "Agent (8003) ready after $((i*5)) seconds"
+        break
+    fi
+    sleep 5
+done
+
+if ! curl -s http://localhost:8004/health > /dev/null 2>&1; then
+    echo "ERROR: User server not healthy"
+    kill $USER_PID $AGENT_PID 2>/dev/null
+    exit 1
+fi
+if ! curl -s http://localhost:8003/health > /dev/null 2>&1; then
+    echo "ERROR: Agent server not healthy"
+    kill $USER_PID $AGENT_PID 2>/dev/null
+    exit 1
+fi
+echo "Both vLLM servers ready"
+sleep 5
+
+# Test methods: vanilla (baseline), rag (fixed), rag_vector (fixed)
+OUTPUT_DIR="../results/rag_fix_test_$(date +%Y%m%d_%H%M%S)"
+
+for METHOD in vanilla rag rag_vector; do
+    echo ""
+    echo "============================================"
+    echo "Testing method: $METHOD"
+    echo "============================================"
+    date
+    START=$(date +%s)
+
+    python scripts/run_experiments.py --methods $METHOD \
+        --datasets math-hard --n-profiles 5 --n-sessions 10 --max-turns 15 \
+        --use-vllm --no-batch-processing --parallel-profiles 5 \
+        --output-dir $OUTPUT_DIR --profile-path $PROFILE_PATH
+
+    END=$(date +%s)
+    ELAPSED=$((END-START))
+
+    if [ $? -eq 0 ]; then
+        echo "Method $METHOD: SUCCESS (${ELAPSED}s)"
+    else
+        echo "Method $METHOD: FAILED after ${ELAPSED}s"
+    fi
+done
+
+echo ""
+echo "============================================"
+echo "RAG Fix Test Complete"
+echo "============================================"
+echo "Results saved to: $OUTPUT_DIR"
+date
+
+# Cleanup
+pkill -f "vllm.entrypoints" 2>/dev/null || true