1 files changed, 94 insertions, 0 deletions
diff --git a/collaborativeagents/slurm/fullscale/test_local_user.sh b/collaborativeagents/slurm/fullscale/test_local_user.sh
new file mode 100644
index 0000000..8374a93
--- /dev/null
+++ b/collaborativeagents/slurm/fullscale/test_local_user.sh
@@ -0,0 +1,94 @@
+#!/bin/bash
+#SBATCH --job-name=test_local_user
+#SBATCH --account=bfqt-delta-gpu
+#SBATCH --partition=gpuA100x4
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=32
+#SBATCH --gres=gpu:nvidia_a100:4
+#SBATCH --mem=200G
+#SBATCH --time=1:00:00
+#SBATCH --output=test_local_user_%j.out
+#SBATCH --error=test_local_user_%j.err
+
+# Test with LOCAL vLLM user simulator (70B AWQ) instead of OpenAI
+# Expected: ~2000+ sessions/hr (vs ~60 with OpenAI API latency)
+#
+# GPU Layout:
+#   GPU 0-1: 70B user simulator (AWQ INT4, TP=2)
+#   GPU 2: 8B agent (~24GB)
+#   GPU 3: Available for embedding/reranker if needed
+
+cd /projects/bfqt/users/yurenh2/ml-projects/personalization-user-model
+source /u/yurenh2/miniforge3/etc/profile.d/conda.sh
+conda activate eval
+
+export HF_HOME=/projects/bfqt/users/yurenh2/hf_cache/huggingface
+export PYTHONPATH="${PWD}/src:${PWD}/collaborativeagents:${PYTHONPATH}"
+export NCCL_P2P_DISABLE=1
+
+set -a
+source .env
+set +a
+
+# Models
+MODEL_70B="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"
+
+# Kill any existing vLLM servers
+pkill -f "vllm.entrypoints" 2>/dev/null || true
+sleep 3
+
+echo "=== Starting 70B User Simulator (GPU 0-1, TP=2) ==="
+CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
+    --model $MODEL_70B \
+    --port 8004 \
+    --tensor-parallel-size 2 \
+    --gpu-memory-utilization 0.90 \
+    --max-model-len 4096 \
+    --quantization awq \
+    --dtype float16 \
+    --disable-log-requests \
+    --guided-decoding-backend outlines &
+
+echo "=== Starting 8B Agent (GPU 2) ==="
+CUDA_VISIBLE_DEVICES=2 python -m vllm.entrypoints.openai.api_server \
+    --model $MODEL_8B \
+    --port 8003 \
+    --tensor-parallel-size 1 \
+    --gpu-memory-utilization 0.50 \
+    --max-model-len 8192 \
+    --dtype bfloat16 \
+    --disable-log-requests &
+
+# Wait for both servers
+echo "Waiting for vLLM servers..."
+for port in 8004 8003; do
+    for i in $(seq 1 120); do
+        curl -s http://localhost:$port/health > /dev/null 2>&1 && break
+        sleep 2
+    done
+    echo "  Port $port ready."
+done
+
+cd collaborativeagents/scripts
+
+echo ""
+echo "=== Running Test: 10 profiles × 2 sessions with LOCAL user simulator ==="
+python run_experiments.py \
+    --methods vanilla \
+    --datasets math-hard \
+    --n-profiles 10 \
+    --n-sessions 2 \
+    --max-turns 8 \
+    --use-vllm \
+    --vllm-user-url http://localhost:8004/v1 \
+    --vllm-agent-url http://localhost:8003/v1 \
+    --reward-mode llm \
+    --parallel-profiles 10 \
+    --profile-path ../data/complex_profiles_v2/profiles_200.jsonl \
+    --output-dir ../results/test_local_user_$(date +%Y%m%d_%H%M%S)
+
+echo ""
+echo "=== Test Complete ==="
+pkill -f "vllm.entrypoints" 2>/dev/null || true