dreamerdeo
commited on
Update run_train_test.sh
Browse files- run_train_test.sh +6 -1
run_train_test.sh
CHANGED
@@ -5,7 +5,10 @@ TRAIN_ARGS="--train_iters 50000 --lr_decay_style cosine
|
|
5 |
--lr_warmup_iters 50 --lr 1e-5 --min_lr 1e-6 --use_flash_attn
|
6 |
--attention_dropout 0.0 --adam_beta1 0.9 --adam_beta2 0.95 --adam_eps 1e-5"
|
7 |
|
|
|
8 |
DISTRIBUTED_ARGS="--nproc_per_node 8 --nnodes 1 --node_rank 0 --master_addr localhost --master_port 8000"
|
|
|
|
|
9 |
# DISTRIBUTED_ARGS="--nproc_per_node 8 --nnodes 2 --node_rank $RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
|
10 |
|
11 |
LLAMA2_ARGS=" --no_bias_gelu_fusion --no_bias_dropout_fusion
|
@@ -25,9 +28,11 @@ torchrun $DISTRIBUTED_ARGS finetune.py \
|
|
25 |
--vocab_file megatron_test/llama2_tokenizer.model \
|
26 |
--bf16 \
|
27 |
--micro_batch_size 4 \
|
28 |
-
--global_batch_size
|
29 |
--sequence_parallel \
|
30 |
--recompute_granularity selective \
|
31 |
--use_checkpoint_args \
|
32 |
$COMMON_ARGS $LOG_ARGS $TRAIN_ARGS $LLAMA2_ARGS
|
33 |
|
|
|
|
|
|
5 |
--lr_warmup_iters 50 --lr 1e-5 --min_lr 1e-6 --use_flash_attn
|
6 |
--attention_dropout 0.0 --adam_beta1 0.9 --adam_beta2 0.95 --adam_eps 1e-5"
|
7 |
|
8 |
+
### for one node, 8GPUs
|
9 |
DISTRIBUTED_ARGS="--nproc_per_node 8 --nnodes 1 --node_rank 0 --master_addr localhost --master_port 8000"
|
10 |
+
|
11 |
+
# ### for multi nodes, 8GPUs
|
12 |
# DISTRIBUTED_ARGS="--nproc_per_node 8 --nnodes 2 --node_rank $RANK --master_addr $MASTER_ADDR --master_port $MASTER_PORT"
|
13 |
|
14 |
LLAMA2_ARGS=" --no_bias_gelu_fusion --no_bias_dropout_fusion
|
|
|
28 |
--vocab_file megatron_test/llama2_tokenizer.model \
|
29 |
--bf16 \
|
30 |
--micro_batch_size 4 \
|
31 |
+
--global_batch_size 128 \
|
32 |
--sequence_parallel \
|
33 |
--recompute_granularity selective \
|
34 |
--use_checkpoint_args \
|
35 |
$COMMON_ARGS $LOG_ARGS $TRAIN_ARGS $LLAMA2_ARGS
|
36 |
|
37 |
+
### Increase the micro_batch_size, if you have 80G GPU
|
38 |
+
### Decrease the micro_batch_size, if you need to train larger model
|