#!/bin/bash
#SBATCH --job-name=vllm-serve
#SBATCH --nodes=1
#SBATCH --gpus=4
#SBATCH --time=4:00:00
#SBATCH --exclusive
#SBATCH --output=out/%x.%j.out

echo SERVING ON $HOSTNAME

module reset
module load brics/nccl
module list

source .venv/bin/activate

YAML_CONFIG="/projects/public/brics/distributed_vllm/GPT-OSS_Hopper.yaml"
HF_HOME=/projects/public/brics/hf
MODEL_PATH=$HF_HOME/hub/models--openai--gpt-oss-120b/snapshots/b5c939de8f754692c1647ca79fbf85e8c1e70f8a/
MODEL_NAME="openai/gpt-oss-120b"

export TIKTOKEN_ENCODINGS_BASE="/projects/public/brics/distributed_vllm/etc/encodings"

srun \
    --nodes=$SLURM_NNODES \
    --gpus=$SLURM_GPUS \
    --cpus-per-task 72 \
    --ntasks-per-node 1 \
    vllm serve $MODEL_PATH \
    --served-model-name $MODEL_NAME \
    --config $YAML_CONFIG \
    --host 0.0.0.0 \
    --port 8000 \
    --max-num-seqs 512 \
    --tensor_parallel_size=4