How to fine tune vicuna-7b with A40

[yqh984638220]

How to fine tune vicuna-7b with A40

[gabinguo]

FlashAttention backward for head dim > 64 requires A100 or H100
GPUs as the implementation needs a large amount of shared memory.

This might be related.. got this error while using train_mem.py with A40s..

[jhu10]

same problem. running train_mem.py using following args: torchrun --nproc_per_node=2 --master_port=20001 /data/ljn/Vicuna-13B/model/FastChat/fastchat/train/train_mem.py \ --model_name_or_path /data/ljn/Vicuna-13B/model/FastChat/data/hf \ --data_path /data/ljn/Vicuna-13B/model/FastChat/playground/data/leecode_new.json \ --bf16 True \ --output_dir /data/ljn/Vicuna-13B/model/FastChat/output \ --num_train_epochs 3 \ --per_device_train_batch_size 2 \ --per_device_eval_batch_size 2 \ --gradient_accumulation_steps 16 \ --evaluation_strategy "no" \ --save_strategy "steps" \ --save_steps 1200 \ --save_total_limit 10 \ --learning_rate 2e-5 \ --weight_decay 0. \ --warmup_ratio 0.03 \ --lr_scheduler_type "cosine" \ --logging_steps 1 \ --fsdp "full_shard auto_wrap" \ --fsdp_transformer_layer_cls_to_wrap 'LlamaDecoderLayer' \ --tf32 True \ --model_max_length 2048 \ --gradient_checkpointing True \ --lazy_preprocess True

Error: Loading checkpoint shards: 100%|████████████████████████████████████████████████████████████████████| 3/3 [03:19<00:00, 66.41s/it] wandb: (1) Create a W&B account wandb: (2) Use an existing W&B account wandb: (3) Don't visualize my results wandb: Enter your choice: 3 wandb: You chose "Don't visualize my results" wandb: Tracking run with wandb version 0.15.3 wandb: W&B syncing is set to offline in this directory. wandb: Run wandb online or set WANDB_MODE=online to enable cloud syncing. 0%| | 0/537 [00:00<?, ?it/s]Traceback (most recent call last): File "/data/ljn/Vicuna-13B/model/FastChat/fastchat/train/train_mem.py", line 13, in train() File "/data/ljn/Vicuna-13B/model/FastChat/fastchat/train/train.py", line 263, in train trainer.train() File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/transformers/trainer.py", line 1662, in train return inner_training_loop( File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/transformers/trainer.py", line 1927, in _inner_training_loop tr_loss_step = self.training_step(model, inputs) File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/transformers/trainer.py", line 2717, in training_step loss.backward() File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/torch/_tensor.py", line 487, in backward torch.autograd.backward( File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/torch/autograd/init.py", line 200, in backward Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/torch/autograd/function.py", line 274, in apply return user_fn(self, args) File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/torch/utils/checkpoint.py", line 157, in backward torch.autograd.backward(outputs_with_grad, args_with_grad) File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/torch/autograd/init.py", line 200, in backward Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/torch/autograd/function.py", line 274, in apply return user_fn(self, args) File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/flash_attn/flash_attn_interface.py", line 75, in backward _flash_attn_backward( File "/data/ljn/Vicuna-13B/lib/python3.9/site-packages/flash_attn/flash_attn_interface.py", line 42, in _flash_attnbackward , , , softmax_d = flash_attn_cuda.bwd( RuntimeError: FlashAttention backward for head dim > 64 requires A100 or H100 GPUs as the implementation needs a large amount of shared memory.

I used 2 * A40 on my machine, is there any suggestions to solve this?

[Ted8000]

maybe you can remove flash_attn() ,

[prateeky2806]

I am facing this similar issues, anyone figured out how to solve this?

[Hzzhang-nlp]

我似乎也遇到了同样的问题

[Hzzhang-nlp]

How to solve it

[aduan]

@Hzzhang-nlp if cuda is 12.x, can install pytorch 12.1 from nightly and install flash-attention from source

pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu121

git clone https://github.com/HazyResearch/flash-attention.git
python setup.py install

[Len-Li]

It works for me. But still OOM for a single A6000 finetuning.

[surak]

We can't do much about this, I fear. If the gpu has not enough memory, it has not enough memory...