sft have bug while lora run successfully

[TimeFlysLeo]

Reminder

• I have read the README and searched the existing issues.

System Info

• llamafactory version: 0.9.2.dev0
• Platform: Linux-6.8.0-49-generic-x86_64-with-glibc2.35
• Python version: 3.10.16
• PyTorch version: 2.5.1+cu124 (GPU)
• Transformers version: 4.46.1
• Datasets version: 3.1.0
• Accelerate version: 1.0.1
• PEFT version: 0.12.0
• TRL version: 0.9.6
• GPU type: NVIDIA GeForce RTX 4090 D
• DeepSpeed version: 0.15.4
• Bitsandbytes version: 0.45.0

Reproduction

0/1000 [00:00<?, ?it/s]/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/utils/checkpoint.py:87: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
warnings.warn(
/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/utils/checkpoint.py:87: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
warnings.warn(
/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/utils/checkpoint.py:87: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
warnings.warn(
/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/utils/checkpoint.py:87: UserWarning: None of the inputs have requires_grad=True. Gradients will be None
warnings.warn(
[WARNING|logging.py:168] 2024-12-20 18:40:09,732 >> use_cache=True is incompatible with gradient checkpointing. Setting use_cache=False.
[rank1]: Traceback (most recent call last):
[rank1]: File "/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py", line 23, in
[rank1]: launch()
[rank1]: File "/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py", line 19, in launch
[rank1]: run_exp()
[rank1]: File "/home/lzx/LLaMA-Factory/src/llamafactory/train/tuner.py", line 50, in run_exp
[rank1]: run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
[rank1]: File "/home/lzx/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line 163, in run_sft
[rank1]: train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 2122, in train
[rank1]: return inner_training_loop(
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 2474, in _inner_training_loop
[rank1]: tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 3606, in training_step
[rank1]: self.accelerator.backward(loss, **kwargs)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/accelerate/accelerator.py", line 2238, in backward
[rank1]: self.deepspeed_engine_wrapped.backward(loss, **kwargs)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/accelerate/utils/deepspeed.py", line 186, in backward
[rank1]: self.engine.backward(loss, **kwargs)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank1]: ret_val = func(*args, **kwargs)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2020, in backward
[rank1]: self.optimizer.backward(loss, retain_graph=retain_graph)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank1]: ret_val = func(*args, **kwargs)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/zero/stage3.py", line 2259, in backward
[rank1]: self.loss_scaler.backward(loss.float(), retain_graph=retain_graph)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/fp16/loss_scaler.py", line 63, in backward
[rank1]: scaled_loss.backward(retain_graph=retain_graph)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/_tensor.py", line 581, in backward
[rank1]: torch.autograd.backward(
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/init.py", line 347, in backward
[rank1]: _engine_run_backward(
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/graph.py", line 825, in _engine_run_backward
[rank1]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/function.py", line 307, in apply
[rank1]: return user_fn(self, *args)
[rank1]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/utils/checkpoint.py", line 317, in backward
[rank1]: raise RuntimeError(
[rank1]: RuntimeError: none of output has requires_grad=True, this checkpoint() is not necessary
[rank2]: Traceback (most recent call last):
[rank2]: File "/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py", line 23, in
[rank2]: launch()
[rank2]: File "/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py", line 19, in launch
[rank2]: run_exp()
[rank2]: File "/home/lzx/LLaMA-Factory/src/llamafactory/train/tuner.py", line 50, in run_exp
[rank2]: run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
[rank2]: File "/home/lzx/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line 163, in run_sft
[rank2]: train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 2122, in train
[rank2]: return inner_training_loop(
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 2474, in _inner_training_loop
[rank2]: tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 3606, in training_step
[rank2]: self.accelerator.backward(loss, **kwargs)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/accelerate/accelerator.py", line 2238, in backward
[rank2]: self.deepspeed_engine_wrapped.backward(loss, **kwargs)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/accelerate/utils/deepspeed.py", line 186, in backward
[rank2]: self.engine.backward(loss, **kwargs)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank2]: ret_val = func(*args, **kwargs)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2020, in backward
[rank2]: self.optimizer.backward(loss, retain_graph=retain_graph)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank2]: ret_val = func(*args, **kwargs)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/zero/stage3.py", line 2259, in backward
[rank2]: self.loss_scaler.backward(loss.float(), retain_graph=retain_graph)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/fp16/loss_scaler.py", line 63, in backward
[rank2]: scaled_loss.backward(retain_graph=retain_graph)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/_tensor.py", line 581, in backward
[rank2]: torch.autograd.backward(
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/init.py", line 347, in backward
[rank2]: _engine_run_backward(
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/graph.py", line 825, in _engine_run_backward
[rank2]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/function.py", line 307, in apply
[rank2]: return user_fn(self, *args)
[rank2]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/utils/checkpoint.py", line 317, in backward
[rank2]: raise RuntimeError(
[rank2]: RuntimeError: none of output has requires_grad=True, this checkpoint() is not necessary
[rank3]: Traceback (most recent call last):
[rank3]: File "/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py", line 23, in
[rank3]: launch()
[rank3]: File "/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py", line 19, in launch
[rank3]: run_exp()
[rank3]: File "/home/lzx/LLaMA-Factory/src/llamafactory/train/tuner.py", line 50, in run_exp
[rank3]: run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
[rank3]: File "/home/lzx/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line 163, in run_sft
[rank3]: train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 2122, in train
[rank3]: return inner_training_loop(
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 2474, in _inner_training_loop
[rank3]: tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 3606, in training_step
[rank3]: self.accelerator.backward(loss, **kwargs)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/accelerate/accelerator.py", line 2238, in backward
[rank3]: self.deepspeed_engine_wrapped.backward(loss, **kwargs)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/accelerate/utils/deepspeed.py", line 186, in backward
[rank3]: self.engine.backward(loss, **kwargs)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank3]: ret_val = func(*args, **kwargs)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2020, in backward
[rank3]: self.optimizer.backward(loss, retain_graph=retain_graph)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank3]: ret_val = func(*args, **kwargs)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/zero/stage3.py", line 2259, in backward
[rank3]: self.loss_scaler.backward(loss.float(), retain_graph=retain_graph)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/fp16/loss_scaler.py", line 63, in backward
[rank3]: scaled_loss.backward(retain_graph=retain_graph)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/_tensor.py", line 581, in backward
[rank3]: torch.autograd.backward(
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/init.py", line 347, in backward
[rank3]: _engine_run_backward(
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/graph.py", line 825, in _engine_run_backward
[rank3]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/function.py", line 307, in apply
[rank3]: return user_fn(self, *args)
[rank3]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/utils/checkpoint.py", line 317, in backward
[rank3]: raise RuntimeError(
[rank3]: RuntimeError: none of output has requires_grad=True, this checkpoint() is not necessary
[rank0]: Traceback (most recent call last):
[rank0]: File "/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py", line 23, in
[rank0]: launch()
[rank0]: File "/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py", line 19, in launch
[rank0]: run_exp()
[rank0]: File "/home/lzx/LLaMA-Factory/src/llamafactory/train/tuner.py", line 50, in run_exp
[rank0]: run_sft(model_args, data_args, training_args, finetuning_args, generating_args, callbacks)
[rank0]: File "/home/lzx/LLaMA-Factory/src/llamafactory/train/sft/workflow.py", line 163, in run_sft
[rank0]: train_result = trainer.train(resume_from_checkpoint=training_args.resume_from_checkpoint)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 2122, in train
[rank0]: return inner_training_loop(
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 2474, in _inner_training_loop
[rank0]: tr_loss_step = self.training_step(model, inputs, num_items_in_batch)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/transformers/trainer.py", line 3606, in training_step
[rank0]: self.accelerator.backward(loss, **kwargs)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/accelerate/accelerator.py", line 2238, in backward
[rank0]: self.deepspeed_engine_wrapped.backward(loss, **kwargs)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/accelerate/utils/deepspeed.py", line 186, in backward
[rank0]: self.engine.backward(loss, **kwargs)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank0]: ret_val = func(*args, **kwargs)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/engine.py", line 2020, in backward
[rank0]: self.optimizer.backward(loss, retain_graph=retain_graph)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/utils/nvtx.py", line 18, in wrapped_fn
[rank0]: ret_val = func(*args, **kwargs)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/zero/stage3.py", line 2259, in backward
[rank0]: self.loss_scaler.backward(loss.float(), retain_graph=retain_graph)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/deepspeed/runtime/fp16/loss_scaler.py", line 63, in backward
[rank0]: scaled_loss.backward(retain_graph=retain_graph)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/_tensor.py", line 581, in backward
[rank0]: torch.autograd.backward(
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/init.py", line 347, in backward
[rank0]: _engine_run_backward(
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/graph.py", line 825, in _engine_run_backward
[rank0]: return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/autograd/function.py", line 307, in apply
[rank0]: return user_fn(self, *args)
[rank0]: File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/utils/checkpoint.py", line 317, in backward
[rank0]: raise RuntimeError(
[rank0]: RuntimeError: none of output has requires_grad=True, this checkpoint() is not necessary
0%| | 0/1000 [00:11<?, ?it/s]
W1220 18:40:21.147000 2883502 site-packages/torch/distributed/elastic/multiprocessing/api.py:897] Sending process 2883573 closing signal SIGTERM
W1220 18:40:21.147000 2883502 site-packages/torch/distributed/elastic/multiprocessing/api.py:897] Sending process 2883574 closing signal SIGTERM
W1220 18:40:21.148000 2883502 site-packages/torch/distributed/elastic/multiprocessing/api.py:897] Sending process 2883576 closing signal SIGTERM
E1220 18:40:21.376000 2883502 site-packages/torch/distributed/elastic/multiprocessing/api.py:869] failed (exitcode: 1) local_rank: 2 (pid: 2883575) of binary: /usr/local/miniconda3/envs/llama_factory/bin/python
Traceback (most recent call last):
File "/usr/local/miniconda3/envs/llama_factory/bin/torchrun", line 8, in
sys.exit(main())
File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/init.py", line 355, in wrapper
return f(*args, **kwargs)
File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/distributed/run.py", line 919, in main
run(args)
File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/distributed/run.py", line 910, in run
elastic_launch(
File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 138, in call
return launch_agent(self._config, self._entrypoint, list(args))
File "/usr/local/miniconda3/envs/llama_factory/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 269, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

/home/lzx/LLaMA-Factory/src/llamafactory/launcher.py FAILED

Failures:
<NO_OTHER_FAILURES>

Root Cause (first observed failure):
[0]:
time : 2024-12-20_18:40:21
host : gp-SYS-4029GP-TRT2-EC028B
rank : 2 (local_rank: 2)
exitcode : 1 (pid: 2883575)
error_file: <N/A>
traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html

Expected behavior

我通过这里的代码在workflow里面freeze的大部分参数

for lay_name, param in model.named_parameters():
     if lay_name in llama3_vl_train_layers:
         param.requires_grad = True
     else:
         param.requires_grad = False

我的代码在一块3090上训练可以，在6卡4090d的服务器上使用lora微调可以，但是冻结参数直接sft会有这个报错。

我已经将batchsize=1，cutoff_len=256，显存还是爆炸了，因此您发的那个FAQ是没法解决我这个问题。

我想知道这是什么原因，如何修改。是不是我冻结的参数在使用deepspeed的时候没有均匀分配到每张卡上？非常感谢！！！

Others

我的yaml参数如下

model

model_name_or_path: /home/Llama-3.2-11B-Vision-Instruct

trust_remote_code: true

method

stage: sft
do_train: true
finetuning_type: full
deepspeed: examples/deepspeed/ds_z3_config.json # choices: [ds_z0_config.json, ds_z2_config.json, ds_z3_config.json]

dataset

dataset: tenk

dataset: identity,alpaca_en_demo

template: mllama
cutoff_len: 256
max_samples: 800
overwrite_cache: true
preprocessing_num_workers: 16

output

output_dir: saves/llama3-11B-Vision-Instruct/tenk/DCT_4_on_2k_5/vision_32/sft
logging_steps: 10
save_steps: 500
plot_loss: true
overwrite_output_dir: true

train

per_device_train_batch_size: 1
gradient_accumulation_steps: 1
learning_rate: 1.0e-5
num_train_epochs: 5
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
ddp_timeout: 180000000
optim: paged_adamw_8bit

eval

val_size: 0.001
per_device_eval_batch_size: 1
eval_strategy: steps
eval_steps: 500

[moonlightian]

Hi, could you please share your codes for lora?

[TimeFlysLeo]

Hi, could you please share your codes for lora?
i just use the code from the LLaMA-Factory