msj19/gated_deltaproduct
05
1# export PDSH_RCMD_TYPE=ssh2# export NCCL_SOCKET_IFNAME=bond13# export NCCL_IB_DISABLE=0 # 明确启用IB4bash train.sh \5 type=emgla \6 lr=3e-4 \7 scheduler=cosine_with_min_lr \8 batch=16 \9 update=2 \10 warmup=512 \11 steps=30720 \12 context=2048 \13 gpus=8 \14 nodes=1 \15 path=/mnt/jfzn/msj/train_exp/emgla_340M \16 project=fla \17 model=configs/emgla_340M.json \18 data=cerebras/SlimPajama-627B \19 name=SlimPajama \20 cache=/mnt/jfzn/data/SlimPajama-627B/pre_slimp_chunk1/slimp/train \21 tasks=run \22 23 24 25 26 