blob: 7dcad6549a445b97d165b30426300cf1a73c6439 (
plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
|
#!/bin/bash
#SBATCH --job-name=exp_rag
#SBATCH --account=bfqt-delta-gpu
#SBATCH --partition=gpuH200x8
#SBATCH --gres=gpu:4
#SBATCH --nodes=1
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=32
#SBATCH --mem=256G
#SBATCH --time=24:00:00
#SBATCH --output=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/exp_rag-%j.out
#SBATCH --error=/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/exp_rag-%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"
PROFILE_PATH="/projects/bfqt/users/yurenh2/ml-projects/personalization-user-model/collaborativeagents/data/complex_profiles_v2/profiles_200.jsonl"
echo "=== rag (PersonalizedLLMAdapter with local transformers) ==="
date
nvidia-smi --query-gpu=index,name,memory.total --format=csv
pkill -f "vllm.entrypoints" 2>/dev/null || true
sleep 2
# GPU 0,1: vLLM server for user simulator (port 8004)
# PersonalizedLLMAdapter uses local transformers for embedding/reranker/chat/extractor
echo "Starting user simulator vLLM server on GPU 0,1 (port 8004)..."
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 &
echo "Waiting for vLLM server..."
for i in $(seq 1 200); do
if curl -s http://localhost:8004/health > /dev/null 2>&1; then
echo "Server ready after $((i*3))s"; break
fi
sleep 3
done
# GPU 2,3: PersonalizedLLMAdapter's transformers models
# (embedding ~8B, reranker ~8B, chat ~1.5B, extractor ~0.6B)
CUDA_VISIBLE_DEVICES=2,3 python scripts/run_experiments.py \
--methods rag \
--datasets math-hard,math-500,bigcodebench \
--n-profiles 200 --n-sessions 30 --max-turns 15 \
--use-vllm --parallel-profiles 10 \
--output-dir ../results/full_h200 \
--profile-path "$PROFILE_PATH"
pkill -f "vllm.entrypoints" 2>/dev/null || true
echo "Done: $(date)"
|
