#!/bin/bash
# Ablation: Direct Soft Matching (no prototypes)
# Uses train_direct_match.py instead of train.py

rm -rf ms-swift/ms_swift.egg-info

CLUSTER_CONFIG='./cluster.json'
cat <<EOT > ${CLUSTER_CONFIG}
{
    "worker": {
        "gpu": 100,
        "cpu": 1200,
        "memory": 50000
    }
}
EOT

WORKER_COUNT=8

args="--config_file=./args/gsm_codi_direct_match_train.yaml"

nebulactl run mdl --engine=xdl \
                  --refresh \
                  --ignore=data/*,tools/pycocoevalcap/*,tools/pycocotools/*,.git/*,analysis/results/*,plot/* \
                  --queue=pailitao_algo_high_2 \
                  --worker_count=$WORKER_COUNT \
                  --odps_project=pailitao_algo_high_2 \
                  --entry="./train_direct_match.py" \
                  --oss_access_id=LTAI5tMZb5HHH6HHgAExAyiL \
                  --oss_access_key=6MVP8JnTE01t4TLUHhPoFV0swZNTrb \
                  --oss_bucket=pailitao-ai \
                  --oss_endpoint=ossproxy.alibaba-inc.com \
                  --algo_name=pytorch260 \
                  --user_params="$args" \
                  --job_name=gpt2-gsm8k_ablation_direct_match_mt0.1_fac0.5_l6_lr3e3_btz256_pos2_ep30-train \
                  --file.cluster_file=$CLUSTER_CONFIG \
                  --force \
                  --max_failover_times=2 \
                  --requirements_file_name="./requirements.txt"

rm ${CLUSTER_CONFIG}