MSAGPT / scripts /cli_sat.sh
Bo1015's picture
Upload 27 files
0dce0bd verified
raw
history blame contribute delete
No virus
1.35 kB
#!/bin/bash
script_path=$(realpath $0)
script_dir=$(dirname $script_path)
main_dir=$(dirname $script_dir)
MP_SIZE=1
# MODEL_NAME="MSAGPT-"
# MODEL_NAME="MSAGPT-dpo"
SEED=12345
MAX_GEN_LENGTH=128
MIN_GEN_LENGTH=0
# BeamSearchStrategy args
NUM_BEAMS=4
LENGTH_PENALTY=1.0
NO_REPEAT_NGRAM=0
# BaseStrategy args
TEMP=0.8
TOPK=0
TOPP=0.9
PORT=19865
MODEL_ARGS="--bf16 \
--skip-init \
--mode finetune \
--rotary-embedding-2d"
# --mode inference \ TODO: sat ds_config bug?
GENERATION_ARGS="--seed $SEED \
--sampling-strategy BaseStrategy \
--max-gen-length $MAX_GEN_LENGTH \
--min-gen-length $MIN_GEN_LENGTH \
--num-beams $NUM_BEAMS \
--length-penalty $LENGTH_PENALTY \
--no-repeat-ngram-size $NO_REPEAT_NGRAM \
--multiline_stream \
--temperature $TEMP \
--top_k $TOPK \
--top_p $TOPP
"
# --sampling-strategy BeamSearchStrategy \
# --no-gap
OPTIONS_NCCL="NCCL_DEBUG=VERSION NCCL_IB_DISABLE=0 NCCL_NET_GDR_LEVEL=2 CUDA_LAUNCH_BLOCKING=0"
ARGS="${main_dir}/cli_sat.py \
$MODEL_ARGS \
$GENERATION_ARGS \
$*"
run_cmd="${OPTIONS_NCCL} torchrun --nproc_per_node $MP_SIZE --master_port=$PORT ${ARGS}"
echo ${run_cmd}
eval ${run_cmd}
set +x