Add collaborativeagents module and update gitignore

- Add collaborativeagents subproject with adapters, agents, and evaluation modules
- Update .gitignore to exclude large binary files (.whl, .tar), wandb logs, and results

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

1 files changed, 205 insertions, 0 deletions

diff --git a/collaborativeagents/scripts/test_parallel_vllm.sh b/collaborativeagents/scripts/test_parallel_vllm.sh
new file mode 100755
index 0000000..0cd0f1f
--- /dev/null
+++ b/collaborativeagents/scripts/test_parallel_vllm.sh
@@ -0,0 +1,205 @@
+#!/bin/bash
+# Test parallel vLLM processing on H200x8-interactive
+# Usage: Run this on an interactive H200 node
+#
+# srun --account=bfqt-delta-gpu --partition=gpuH200x8-interactive \
+#      --nodes=1 --gpus-per-node=4 --time=02:00:00 --mem=200G --pty bash
+# cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents
+# bash scripts/test_parallel_vllm.sh
+
+set -e
+
+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="${PWD}:${PWD}/../src:${PYTHONPATH}"
+
+# Configuration
+MODEL_70B_AWQ="hugging-quants/Meta-Llama-3.1-70B-Instruct-AWQ-INT4"
+MODEL_8B="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
+PORT_70B=8004
+PORT_8B=8003
+
+echo "============================================"
+echo "Parallel vLLM Experiment Test"
+echo "============================================"
+echo "Date: $(date)"
+echo "Node: $(hostname)"
+echo ""
+
+echo "=== GPU Info ==="
+nvidia-smi --query-gpu=index,name,memory.total,memory.free --format=csv
+echo ""
+
+# Kill any existing vLLM servers
+echo "Cleaning up any existing vLLM servers..."
+pkill -f "vllm.entrypoints" 2>/dev/null || true
+sleep 2
+
+echo "============================================"
+echo "Starting vLLM Servers"
+echo "============================================"
+
+# Start 70B AWQ server on GPU 0-1 (needs 2 GPUs for tensor parallelism)
+echo ""
+echo "Starting 70B AWQ vLLM Server (GPU 0-1, TP=2)..."
+CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
+    --model $MODEL_70B_AWQ \
+    --port $PORT_70B \
+    --tensor-parallel-size 2 \
+    --gpu-memory-utilization 0.90 \
+    --max-model-len 4096 \
+    --disable-log-requests \
+    --quantization awq \
+    --dtype float16 &
+SERVER_70B_PID=$!
+echo "70B Server PID: $SERVER_70B_PID"
+
+# Start 8B server on GPU 2
+echo ""
+echo "Starting 8B vLLM Server (GPU 2)..."
+CUDA_VISIBLE_DEVICES=2 python -m vllm.entrypoints.openai.api_server \
+    --model $MODEL_8B \
+    --port $PORT_8B \
+    --gpu-memory-utilization 0.90 \
+    --max-model-len 4096 \
+    --disable-log-requests \
+    --dtype bfloat16 &
+SERVER_8B_PID=$!
+echo "8B Server PID: $SERVER_8B_PID"
+
+echo ""
+echo "Waiting for servers to start..."
+
+# Wait for 70B (may take 3-5 minutes)
+for i in $(seq 1 120); do
+    if curl -s http://localhost:$PORT_70B/health > /dev/null 2>&1; then
+        echo "70B Server ready after $((i*3)) seconds"
+        break
+    fi
+    if [ $((i % 20)) -eq 0 ]; then
+        echo "  Waiting for 70B... ($((i*3)) seconds)"
+    fi
+    sleep 3
+done
+
+# Wait for 8B
+for i in $(seq 1 60); do
+    if curl -s http://localhost:$PORT_8B/health > /dev/null 2>&1; then
+        echo "8B Server ready after $((i*2)) seconds"
+        break
+    fi
+    sleep 2
+done
+
+# Check both servers
+echo ""
+if ! curl -s http://localhost:$PORT_70B/health > /dev/null 2>&1; then
+    echo "ERROR: 70B server failed to start"
+    kill $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null
+    exit 1
+fi
+echo "✓ 70B server healthy"
+
+if ! curl -s http://localhost:$PORT_8B/health > /dev/null 2>&1; then
+    echo "ERROR: 8B server failed to start"
+    kill $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null
+    exit 1
+fi
+echo "✓ 8B server healthy"
+
+echo ""
+echo "=== vLLM Server Info ==="
+echo "70B model:"
+curl -s http://localhost:$PORT_70B/v1/models | python -m json.tool 2>/dev/null | head -10
+echo ""
+echo "8B model:"
+curl -s http://localhost:$PORT_8B/v1/models | python -m json.tool 2>/dev/null | head -10
+
+echo ""
+echo "============================================"
+echo "Test 1: Sequential Processing (1 profile)"
+echo "============================================"
+
+cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents/scripts
+
+time python run_experiments.py \
+    --methods vanilla \
+    --datasets mmlu \
+    --n-profiles 1 \
+    --n-sessions 3 \
+    --use-vllm \
+    --vllm-user-url http://localhost:$PORT_70B/v1 \
+    --vllm-agent-url http://localhost:$PORT_8B/v1 \
+    --parallel-profiles 1 \
+    --output-dir ../results/parallel_test_seq \
+    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl
+
+echo ""
+echo "============================================"
+echo "Test 2: Parallel Processing (4 profiles)"
+echo "============================================"
+
+time python run_experiments.py \
+    --methods vanilla \
+    --datasets mmlu \
+    --n-profiles 4 \
+    --n-sessions 3 \
+    --use-vllm \
+    --vllm-user-url http://localhost:$PORT_70B/v1 \
+    --vllm-agent-url http://localhost:$PORT_8B/v1 \
+    --parallel-profiles 4 \
+    --output-dir ../results/parallel_test_4 \
+    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl
+
+echo ""
+echo "============================================"
+echo "Test 3: Parallel Processing (8 profiles)"
+echo "============================================"
+
+time python run_experiments.py \
+    --methods vanilla \
+    --datasets mmlu \
+    --n-profiles 8 \
+    --n-sessions 3 \
+    --use-vllm \
+    --vllm-user-url http://localhost:$PORT_70B/v1 \
+    --vllm-agent-url http://localhost:$PORT_8B/v1 \
+    --parallel-profiles 8 \
+    --output-dir ../results/parallel_test_8 \
+    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl
+
+echo ""
+echo "============================================"
+echo "Test 4: Parallel Processing (16 profiles)"
+echo "============================================"
+
+time python run_experiments.py \
+    --methods vanilla \
+    --datasets mmlu \
+    --n-profiles 16 \
+    --n-sessions 3 \
+    --use-vllm \
+    --vllm-user-url http://localhost:$PORT_70B/v1 \
+    --vllm-agent-url http://localhost:$PORT_8B/v1 \
+    --parallel-profiles 16 \
+    --output-dir ../results/parallel_test_16 \
+    --profile-path ../data/complex_profiles_v2/profiles_100.jsonl
+
+# Cleanup
+echo ""
+echo "Cleaning up..."
+kill $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null
+wait $SERVER_70B_PID $SERVER_8B_PID 2>/dev/null
+
+echo ""
+echo "============================================"
+echo "TEST COMPLETE!"
+echo "============================================"
+echo ""
+echo "Compare the timing results above to estimate optimal parallelism."
+echo "Expected scaling: Higher parallelism → Higher throughput (until bottleneck)"
+echo ""
+date