github/ColossalAI
1
0
Fork 0
mirror of https://github.com/hpcaitech/ColossalAI.git synced 2025-09-08 20:40:34 +00:00
Code Issues Packages Projects Releases Wiki Activity

refactor evaluation

Browse Source
This commit is contained in:
YeAnbang
2024-07-22 05:57:39 +00:00
parent c5f582f666
commit 12fe8b5858
22 changed files with 309 additions and 1388 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
applications/ColossalChat/examples/training_scripts/train_dpo.py
View File

@@ -287,9 +287,12 @@ def train(args):
LORA_MANAGER.merge_weights = True
model.eval()
# save model checkpoint after fitting on only rank0
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True)
coordinator.print_on_master(f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}")
if args.save_dir is not None:
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True)
coordinator.print_on_master(
f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}"
)
coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB")
@@ -328,8 +331,8 @@ if __name__ == "__main__":
parser.add_argument(
"--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
)
parser.add_argument("--config_file", type=str, default="config_file", help="Config file")
parser.add_argument("--save_dir", type=str, default="output")
parser.add_argument("--config_file", type=str, default=None, help="Config file")
parser.add_argument("--save_dir", type=str, default=None)
parser.add_argument("--max_length", type=int, default=2048, help="Model max length")
parser.add_argument("--max_epochs", type=int, default=3)
parser.add_argument("--batch_size", type=int, default=4)
@@ -351,7 +354,7 @@ if __name__ == "__main__":
parser.add_argument("--merge_lora_weights", type=bool, default=True)
parser.add_argument("--lr", type=float, default=5e-6)
parser.add_argument("--accumulation_steps", type=int, default=8)
parser.add_argument("--log_dir", default="logs", type=str)
parser.add_argument("--log_dir", default=None, type=str)
parser.add_argument("--use_wandb", default=False, action="store_true")
parser.add_argument("--grad_checkpoint", default=False, action="store_true")
parser.add_argument("--use_flash_attn", default=False, action="store_true")
@@ -362,7 +365,8 @@ if __name__ == "__main__":
args.length_normalization = True
args.gamma = args.gamma if args.gamma > 0 else 1.4
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
if args.config_file is not None:
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
train(args)

3
applications/ColossalChat/examples/training_scripts/train_dpo.sh
View File

@@ -18,6 +18,7 @@ set_n_least_used_CUDA_VISIBLE_DEVICES 4
PROJECT_NAME="DPO"
PARENT_SAVE_DIR="" # Path to a folder to save checkpoints
PARENT_CONFIG_FILE="" # Path to a folder to save training config logs
PARENT_LOG_DIR="" # Path to a folder to save training config logs
PRETRAINED_MODEL_PATH="" # huggingface or local model path
PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path
@@ -38,6 +39,7 @@ TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S)
FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}"
SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}"
CONFIG_FILE="${PARENT_CONFIG_FILE}${FULL_PROJECT_NAME}.json"
LOG_DIR="${PARENT_LOG_DIR}${FULL_PROJECT_NAME}"
colossalai run --nproc_per_node 4 --hostfile hostfile --master_port 31313 train_dpo.py \
--pretrain $PRETRAINED_MODEL_PATH \
@@ -47,6 +49,7 @@ colossalai run --nproc_per_node 4 --hostfile hostfile --master_port 31313 train_
--save_interval 1000 \
--save_dir $SAVE_DIR \
--config_file $CONFIG_FILE \
--log_dir $LOG_DIR \
--max_epochs 1 \
--accumulation_steps 2 \
--batch_size 16 \

41
applications/ColossalChat/examples/training_scripts/train_kto.py
View File

@@ -174,14 +174,16 @@ def train(args):
# Check if the user specified weights fit into the theoratical lower and upper bounds from Eq. (8) of https://arxiv.org/abs/2402.01306
actual_ratio = (args.desirable_weight * num_desirable) / (args.undesirable_weight * num_undesirable)
if actual_ratio <= 1:
raise AssertionError(
f"Desirable weight and undesirable weight are not within the theoratical bounds, [1, 4/3]. Actual ratio: {actual_ratio}, please increase desirable weight or decrease undesirable weight."
)
elif actual_ratio > 4 / 3:
raise AssertionError(
f"Desirable weight and undesirable weight are not within the theoratical bounds, [1, 4/3]. Actual ratio: {actual_ratio}, please decrease desirable weight or increase undesirable weight."
)
if actual_ratio < 1 or actual_ratio > 4 / 3:
if not args.auto_weight:
raise AssertionError(
f"Desirable weight and undesirable weight are not within the theoratical bounds, [1, 4/3]. Actual ratio: {actual_ratio}, please increase/decrease desirable weight or decrease/increase undesirable weight."
)
else:
args.desirable_weight = args.desirable_weight / actual_ratio
coordinator.print_on_master(
f"Desirable weight and undesirable weight are not within the theoratical bounds, [1, 4/3]. Actual ratio: {actual_ratio}, auto weight is enabled, set desirable weight to {args.desirable_weight} and undesirable weight to {args.undesirable_weight}"
)
data_collator = DataCollatorForKTODataset(tokenizer=tokenizer, max_length=args.max_length)
@@ -304,9 +306,12 @@ def train(args):
LORA_MANAGER.merge_weights = True
model.eval()
# save model checkpoint after fitting on only rank0
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True)
coordinator.print_on_master(f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}")
if args.save_dir is not None:
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True)
coordinator.print_on_master(
f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}"
)
coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB")
@@ -343,8 +348,8 @@ if __name__ == "__main__":
parser.add_argument(
"--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
)
parser.add_argument("--config_file", type=str, default="config_file", help="Config file")
parser.add_argument("--save_dir", type=str, default="output")
parser.add_argument("--config_file", type=str, default=None, help="Config file")
parser.add_argument("--save_dir", type=str, default=None)
parser.add_argument("--max_length", type=int, default=2048, help="Model max length")
parser.add_argument("--max_epochs", type=int, default=3)
parser.add_argument("--batch_size", type=int, default=4)
@@ -359,14 +364,16 @@ if __name__ == "__main__":
)
parser.add_argument("--save_interval", type=int, default=1000, help="number of step between two checkpoints")
parser.add_argument("--merge_lora_weights", type=bool, default=True)
parser.add_argument("--auto_weight", default=False, action="store_true")
parser.add_argument("--lr", type=float, default=5e-6)
parser.add_argument("--accumulation_steps", type=int, default=8)
parser.add_argument("--log_dir", default="logs", type=str)
parser.add_argument("--log_dir", default=None, type=str)
parser.add_argument("--use_wandb", default=False, action="store_true")
parser.add_argument("--grad_checkpoint", default=False, action="store_true")
parser.add_argument("--use_flash_attn", default=False, action="store_true")
args = parser.parse_args()
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
if args.config_file is not None:
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
train(args)

4
applications/ColossalChat/examples/training_scripts/train_kto.sh
View File

@@ -19,6 +19,7 @@ PROJECT_NAME="kto"
PARENT_SAVE_DIR="" # Path to a folder to save checkpoints
PARENT_TENSORBOARD_DIR="" # Path to a folder to save logs
PARENT_CONFIG_FILE="" # Path to a folder to save training config logs
PARENT_LOG_DIR="" # Path to a folder to save training config logs
PRETRAINED_MODEL_PATH="" # huggingface or local model path
PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path
@@ -39,6 +40,7 @@ TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S)
FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}"
SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}"
CONFIG_FILE="${PARENT_CONFIG_FILE}-${FULL_PROJECT_NAME}.json"
LOG_DIR="${PARENT_LOG_DIR}${FULL_PROJECT_NAME}"
colossalai run --nproc_per_node 4 --master_port 31313 train_kto.py \
--pretrain $PRETRAINED_MODEL_PATH \
@@ -48,9 +50,11 @@ colossalai run --nproc_per_node 4 --master_port 31313 train_kto.py \
--save_interval 1000 \
--save_dir $SAVE_DIR \
--config_file $CONFIG_FILE \
--log_dir $LOG_DIR \
--max_epochs 1 \
--accumulation_steps 1 \
--batch_size 8 \
--auto_weight \
--lr 1e-5 \
--beta 0.1 \
--mixed_precision "bf16" \

22
applications/ColossalChat/examples/training_scripts/train_orpo.py
View File

@@ -269,9 +269,12 @@ def train(args):
LORA_MANAGER.merge_weights = True
model.eval()
# save model checkpoint after fitting on only rank0
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True)
coordinator.print_on_master(f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}")
if args.save_dir is not None:
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True)
coordinator.print_on_master(
f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}"
)
coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB")
@@ -307,8 +310,8 @@ if __name__ == "__main__":
parser.add_argument(
"--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
)
parser.add_argument("--config_file", type=str, default="config_file", help="Config file")
parser.add_argument("--save_dir", type=str, default="output")
parser.add_argument("--config_file", type=str, default=None, help="Config file")
parser.add_argument("--save_dir", type=str, default=None)
parser.add_argument("--max_length", type=int, default=2048, help="Model max length")
parser.add_argument("--max_epochs", type=int, default=3)
parser.add_argument("--batch_size", type=int, default=4)
@@ -330,12 +333,13 @@ if __name__ == "__main__":
parser.add_argument("--merge_lora_weights", type=bool, default=True)
parser.add_argument("--lr", type=float, default=5e-6)
parser.add_argument("--accumulation_steps", type=int, default=8)
parser.add_argument("--log_dir", default="logs", type=str)
parser.add_argument("--log_dir", default=None, type=str)
parser.add_argument("--use_wandb", default=False, action="store_true")
parser.add_argument("--grad_checkpoint", default=False, action="store_true")
parser.add_argument("--use_flash_attn", default=False, action="store_true")
args = parser.parse_args()
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
if args.config_file is not None:
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
train(args)

3
applications/ColossalChat/examples/training_scripts/train_orpo.sh
View File

@@ -18,6 +18,7 @@ set_n_least_used_CUDA_VISIBLE_DEVICES 2
PROJECT_NAME="ORPO"
PARENT_SAVE_DIR="" # Path to a folder to save checkpoints
PARENT_CONFIG_FILE="" # Path to a folder to save training config logs
PARENT_LOG_DIR="" # Path to a folder to save training config logs
PRETRAINED_MODEL_PATH="" # huggingface or local model path
PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path
@@ -38,6 +39,7 @@ TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S)
FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}"
SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}"
CONFIG_FILE="${PARENT_CONFIG_FILE}${FULL_PROJECT_NAME}.json"
LOG_DIR="${PARENT_LOG_DIR}${FULL_PROJECT_NAME}"
colossalai run --nproc_per_node 8 --hostfile hostfile --master_port 31313 train_orpo.py \
--pretrain $PRETRAINED_MODEL_PATH \
@@ -47,6 +49,7 @@ colossalai run --nproc_per_node 8 --hostfile hostfile --master_port 31313 train_
--save_interval 1000 \
--save_dir $SAVE_DIR \
--config_file $CONFIG_FILE \
--log_dir $LOG_DIR \
--max_epochs 3 \
--accumulation_steps 1 \
--batch_size 16 \

22
applications/ColossalChat/examples/training_scripts/train_rm.py
View File

@@ -284,9 +284,12 @@ def train(args):
LORA_MANAGER.merge_weights = True
model.eval()
# save model checkpoint after fitting on only rank0
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True)
coordinator.print_on_master(f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}")
if args.save_dir is not None:
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_dir, "modeling"), shard=True)
coordinator.print_on_master(
f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_dir}"
)
coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB")
@@ -320,8 +323,8 @@ if __name__ == "__main__":
parser.add_argument(
"--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
)
parser.add_argument("--config_file", type=str, default="config_file", help="Config file")
parser.add_argument("--save_dir", type=str, default="output")
parser.add_argument("--config_file", type=str, default=None, help="Config file")
parser.add_argument("--save_dir", type=str, default=None)
parser.add_argument("--max_length", type=int, default=2048, help="Model max length")
parser.add_argument("--max_epochs", type=int, default=3)
parser.add_argument("--batch_size", type=int, default=4)
@@ -338,12 +341,13 @@ if __name__ == "__main__":
parser.add_argument("--merge_lora_weights", type=bool, default=True)
parser.add_argument("--lr", type=float, default=5e-6)
parser.add_argument("--accumulation_steps", type=int, default=8)
parser.add_argument("--log_dir", default="logs", type=str)
parser.add_argument("--log_dir", default=None, type=str)
parser.add_argument("--use_wandb", default=False, action="store_true")
parser.add_argument("--grad_checkpoint", default=False, action="store_true")
parser.add_argument("--use_flash_attn", default=False, action="store_true")
args = parser.parse_args()
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
if args.config_file is not None:
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
train(args)

3
applications/ColossalChat/examples/training_scripts/train_rm.sh
View File

@@ -18,6 +18,7 @@ set_n_least_used_CUDA_VISIBLE_DEVICES 8
PROJECT_NAME="RM"
PARENT_SAVE_DIR="" # Path to a folder to save checkpoints
PARENT_CONFIG_FILE="" # Path to a folder to save training config logs
PARENT_LOG_DIR="" # Path to a folder to save training config logs
PRETRAINED_MODEL_PATH="" # huggingface or local model path
PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path
@@ -38,6 +39,7 @@ TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S)
FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}"
SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}"
CONFIG_FILE="${PARENT_CONFIG_FILE}${FULL_PROJECT_NAME}.json"
LOG_DIR="${PARENT_LOG_DIR}${FULL_PROJECT_NAME}"
colossalai run --nproc_per_node 8 --hostfile hostfile --master_port 31312 train_rm.py \
--pretrain $PRETRAINED_MODEL_PATH \
@@ -47,6 +49,7 @@ colossalai run --nproc_per_node 8 --hostfile hostfile --master_port 31312 train_
--save_interval 1000 \
--save_dir $SAVE_DIR \
--config_file $CONFIG_FILE \
--log_dir $LOG_DIR \
--max_epochs 3 \
--accumulation_steps 1 \
--batch_size 8 \

23
applications/ColossalChat/examples/training_scripts/train_sft.py
View File

@@ -284,10 +284,12 @@ def train(args):
LORA_MANAGER.merge_weights = True
model.eval()
# save model checkpoint after fitting on only rank0
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_path, "modeling"), shard=True)
coordinator.print_on_master(f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_path}")
if args.save_path is not None:
coordinator.print_on_master("Start saving final model checkpoint")
booster.save_model(model, os.path.join(args.save_path, "modeling"), shard=True)
coordinator.print_on_master(
f"Saved final model checkpoint at epoch {args.max_epochs} at folder {args.save_path}"
)
coordinator.print_on_master(f"Max CUDA memory usage: {torch.cuda.max_memory_allocated()/1024**2:.2f} MB")
@@ -321,7 +323,7 @@ if __name__ == "__main__":
parser.add_argument(
"--checkpoint_path", type=str, default=None, help="Checkpoint path if need to resume training form a checkpoint"
)
parser.add_argument("--save_path", type=str, default="output")
parser.add_argument("--save_path", type=str, default=None)
parser.add_argument("--max_epochs", type=int, default=3)
parser.add_argument("--batch_size", type=int, default=4)
parser.add_argument("--max_len", type=int, default=512)
@@ -336,14 +338,15 @@ if __name__ == "__main__":
parser.add_argument("--save_interval", type=int, default=1000, help="number of step between two checkpoints")
parser.add_argument("--merge_lora_weights", type=bool, default=True)
parser.add_argument("--lr", type=float, default=5e-6)
parser.add_argument("--config_file", type=str, default="config_file", help="Config file")
parser.add_argument("--config_file", type=str, default=None, help="Config file")
parser.add_argument("--accumulation_steps", type=int, default=8)
parser.add_argument("--log_dir", default="logs", type=str)
parser.add_argument("--log_dir", default=None, type=str)
parser.add_argument("--use_wandb", default=False, action="store_true")
parser.add_argument("--grad_checkpoint", default=False, action="store_true")
parser.add_argument("--use_flash_attn", default=False, action="store_true")
args = parser.parse_args()
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
if args.config_file is not None:
os.makedirs(os.path.dirname(args.config_file), exist_ok=True)
with open(args.config_file, "w") as f:
json.dump(args.__dict__, f, indent=4)
train(args)

3
applications/ColossalChat/examples/training_scripts/train_sft.sh
View File

@@ -17,6 +17,7 @@ set_n_least_used_CUDA_VISIBLE_DEVICES 4
PROJECT_NAME="SFT"
PARENT_SAVE_DIR="" # Path to a folder to save checkpoints
PARENT_CONFIG_FILE="" # Path to a folder to save training config logs
PARENT_LOG_DIR="" # Path to a folder to save training config logs
PRETRAINED_MODEL_PATH="" # huggingface or local model path
PRETRAINED_TOKENIZER_PATH="" # huggingface or local tokenizer path
declare -a dataset=(
@@ -36,6 +37,7 @@ TIMESTAMP=$(date +%Y-%m-%d-%H-%M-%S)
FULL_PROJECT_NAME="${PROJECT_NAME}-${TIMESTAMP}"
SAVE_DIR="${PARENT_SAVE_DIR}${FULL_PROJECT_NAME}"
CONFIG_FILE="${PARENT_CONFIG_FILE}${FULL_PROJECT_NAME}.json"
LOG_DIR="${PARENT_LOG_DIR}${FULL_PROJECT_NAME}"
echo $(which colossalai)
echo $(which python)
@@ -47,6 +49,7 @@ colossalai run --nproc_per_node 4 --master_port 31312 --hostfile ./hostfile trai
--dataset ${dataset[@]} \
--save_path $SAVE_DIR \
--config_file $CONFIG_FILE \
--log_dir $LOG_DIR \
--lora_rank 0 \
--plugin zero2 \
--batch_size 8 \