1 files changed, 130 insertions, 0 deletions
diff --git a/collaborativeagents/scripts/test_vllm_speed.sbatch b/collaborativeagents/scripts/test_vllm_speed.sbatch
new file mode 100644
index 0000000..070df5d
--- /dev/null
+++ b/collaborativeagents/scripts/test_vllm_speed.sbatch
@@ -0,0 +1,130 @@
+#!/bin/bash
+#SBATCH --job-name=test_vllm_speed
+#SBATCH --account=bfqt-delta-gpu
+#SBATCH --partition=gpuH200x8
+#SBATCH --gres=gpu:4
+#SBATCH --nodes=1
+#SBATCH --ntasks=1
+#SBATCH --cpus-per-task=16
+#SBATCH --mem=128G
+#SBATCH --time=00:30:00
+#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/test_vllm_speed-%j.out
+#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/test_vllm_speed-%j.err
+
+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}/scripts:${PWD}/../src:${PYTHONPATH}"
+
+MODEL_8B="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/models/llama-3.1-8b-instruct"
+
+echo "=== vLLM Speed Test ==="
+date
+nvidia-smi --query-gpu=index,name,memory.used,memory.total --format=csv
+
+pkill -f "vllm.entrypoints" 2>/dev/null || true
+sleep 2
+
+echo ""
+echo "=== Test 1: ContextualAdapter with vLLM (2 servers) ==="
+echo "Starting vLLM servers on GPU 0,1 (user) and GPU 2,3 (agent)..."
+
+CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
+    --model $MODEL_8B --port 8004 --tensor-parallel-size 2 \
+    --gpu-memory-utilization 0.90 --max-model-len 8192 \
+    --disable-log-requests --dtype bfloat16 &
+
+CUDA_VISIBLE_DEVICES=2,3 python -m vllm.entrypoints.openai.api_server \
+    --model $MODEL_8B --port 8003 --tensor-parallel-size 2 \
+    --gpu-memory-utilization 0.90 --max-model-len 8192 \
+    --disable-log-requests --dtype bfloat16 &
+
+for i in $(seq 1 120); do
+    u=$(curl -s http://localhost:8004/health > /dev/null 2>&1 && echo 1 || echo 0)
+    a=$(curl -s http://localhost:8003/health > /dev/null 2>&1 && echo 1 || echo 0)
+    if [ "$u" = "1" ] && [ "$a" = "1" ]; then
+        echo "Both servers ready after $((i*2))s"; break
+    fi
+    sleep 2
+done
+
+python -c "
+import time
+import sys
+sys.path.insert(0, '.')
+from adapters.contextual_adapter import ContextualAdapter
+
+print('Testing ContextualAdapter with vLLM...')
+adapter = ContextualAdapter(vllm_url='http://localhost:8003/v1')
+adapter.initialize()
+adapter.start_session('test_user')
+
+# Warm up
+adapter.generate_response('Hello')
+
+# Benchmark
+n_requests = 20
+start = time.time()
+for i in range(n_requests):
+    resp = adapter.generate_response(f'Solve: What is {i*7} + {i*3}? Give a brief answer.')
+elapsed = time.time() - start
+
+print(f'ContextualAdapter (vLLM): {n_requests} requests in {elapsed:.2f}s')
+print(f'Throughput: {n_requests/elapsed:.2f} req/s = {n_requests/elapsed*3600:.0f} requests/hr')
+print(f'Estimated sessions/hr (assuming 5 turns/session): {n_requests/elapsed*3600/5:.0f}')
+"
+
+pkill -f "vllm.entrypoints" 2>/dev/null || true
+sleep 5
+
+echo ""
+echo "=== Test 2: PersonalizedLLMAdapter (vLLM user + transformers adapter) ==="
+echo "Starting vLLM on GPU 0,1 for user simulation..."
+
+CUDA_VISIBLE_DEVICES=0,1 python -m vllm.entrypoints.openai.api_server \
+    --model $MODEL_8B --port 8004 --tensor-parallel-size 2 \
+    --gpu-memory-utilization 0.90 --max-model-len 8192 \
+    --disable-log-requests --dtype bfloat16 &
+
+for i in $(seq 1 120); do
+    if curl -s http://localhost:8004/health > /dev/null 2>&1; then
+        echo "Server ready after $((i*2))s"; break
+    fi
+    sleep 2
+done
+
+echo "Loading PersonalizedLLMAdapter on GPU 2,3..."
+CUDA_VISIBLE_DEVICES=2,3 python -c "
+import time
+import sys
+sys.path.insert(0, '.')
+from adapters.personalized_llm_adapter import create_baseline_adapter
+
+print('Testing PersonalizedLLMAdapter (all_memory mode)...')
+adapter = create_baseline_adapter('all_memory')
+adapter.initialize()
+adapter.start_session('test_user')
+
+# Warm up
+adapter.generate_response('Hello')
+
+# Benchmark
+n_requests = 10
+start = time.time()
+for i in range(n_requests):
+    resp = adapter.generate_response(f'Solve: What is {i*7} + {i*3}? Give a brief answer.')
+elapsed = time.time() - start
+
+print(f'PersonalizedLLMAdapter (transformers): {n_requests} requests in {elapsed:.2f}s')
+print(f'Throughput: {n_requests/elapsed:.2f} req/s = {n_requests/elapsed*3600:.0f} requests/hr')
+print(f'Estimated sessions/hr (assuming 5 turns/session): {n_requests/elapsed*3600/5:.0f}')
+"
+
+pkill -f "vllm.entrypoints" 2>/dev/null || true
+
+echo ""
+echo "=== Test Complete ==="
+date