Zero loss and nan grad_norm when Flash Attention is enabled

[fgdfgfthgr-fox]

Please check that this issue hasn't been reported before.

• I searched previous Bug Reports didn't find any similar reports.

Expected Behavior

I expect similar loss and grad_norm when training a model with the same setting regardless whether flash attention is enabled or not.

Current behaviour

Currently, during training steps (right from the start), I can see messages of
{'loss': 0.0, 'grad_norm': nan, 'learning_rate': 6.545084971874738e-06, 'epoch': 0.4}
for few steps, before a

  File "/home/huada524/ondemand/data/sys/myjobs/projects/default/1/huada524-prune-env/lib/python3.10/site-packages/flash_attn/bert_padding.py", line 110, in unpad_input
    indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
RuntimeError: CUDA error: an illegal memory access was encountered

error appear and the training stops.

However, if flash attention is disabled with flash_attention: false, then the network trains normally.
{'loss': 3.0972, 'grad_norm': 0.76171875, 'learning_rate': 3.4549150281252635e-06, 'epoch': 0.6}

Steps to reproduce

1. I Installed my axolotl on a remote cluster with 3x L40 graphic cards with slurm, using the following script:

module load python
module load cuda

echo "Setting up python venv..."
python -m venv venv
source venv/bin/activate
python -m pip install --upgrade pip
pip install -U wheel
pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu124 -I
pip install ninja
export TORCH_CUDA_ARCH_LIST="8.6;8.9"
export CUDA_VISIBLE_DEVICES=2
export LD_LIBRARY_PATH=/home/huada524/ondemand/data/sys/myjobs/projects/default/1/venv/lib64/python3.10/site-packages/nvidia/nvjitlink/lib:$LD_LIBRARY_PATH
pip install -v -U "git+https://github.com/facebookresearch/xformers.git@main#egg=xformers"

cd axolotl
git pull
pip install packaging
pip install -e '.[flash-attn,deepspeed]'
# I manually disabled xformers installation from axolotl/requirements.txt so it won't attempt to override the one I just compiled with.
# I also have to apply this patch https://github.com/microsoft/DeepSpeed/issues/5603 to make sure axolotl would launch
cd ..

2. I started the training using the script below:

module load python
module load cuda

source venv/bin/activate

export CUDA_VISIBLE_DEVICES=2

export WANDB_API_KEY=xxxxxxx
export LD_LIBRARY_PATH=/home/huada524/ondemand/data/sys/myjobs/projects/default/1/venv/lib64/python3.10/site-packages/nvidia/nvjitlink/lib:$LD_LIBRARY_PATH

accelerate launch -m axolotl.cli.train config_llama3_40B_dora.yaml

Note the model I am training the lora with is meta's llama-3-70B model with some of its layers removed.
GPU are running on a CUDA version of 12.5, while the loaded module is 12.3.

---

**** Axolotl Dependency Versions *****
accelerate: 0.30.1
peft: 0.11.1
transformers: 4.41.1
trl: 0.8.7.dev0
torch: 2.4.0.dev20240610+cu124
bitsandbytes: 0.43.1

---

Config yaml

base_model: /home/huada524/ondemand/data/sys/myjobs/projects/default/1/PruneMe/slice_with_mergekit/merged
model_type: LlamaForCausalLM
tokenizer_type: AutoTokenizer

load_in_8bit: false
load_in_4bit: true
strict: false

pretraining_dataset:
  - path: HuggingFaceFW/fineweb-edu
    name: default
    type: completion
val_set_size: 0.05
max_steps: 10
output_dir: ./outputs/out

adapter: qlora
lora_r: 8
lora_alpha: 4
lora_dropout: 0.0
lora_target_linear: true
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj
#  - lm_head
peft_use_dora: false
lora_model_dir:

sequence_len: 1024
sample_packing: true
pad_to_sequence_len: true

wandb_mode: online
wandb_project: Creating Llama-3-40B
wandb_entity:
wandb_watch:
wandb_name: Experimental_Runs

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch
lr_scheduler: cosine
learning_rate: 1e-5
#loraplus_lr_embedding: 1e-6

train_on_inputs: false
group_by_length: false
bf16: auto
fp16:
tf32: false

gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
early_stopping_patience:
resume_from_checkpoint:
logging_steps: 1
xformers_attention:
flash_attention: false

warmup_steps: 0
evals_per_epoch: 1
eval_table_size:
saves_per_epoch: 1
debug:
deepspeed:
weight_decay: 0.0
fsdp:
#  - full_shard
#  - auto_wrap
fsdp_config:
#  fsdp_limit_all_gathers: true
#  fsdp_sync_module_states: true
#  fsdp_offload_params: true
#  fsdp_use_orig_params: false
#  fsdp_cpu_ram_efficient_loading: true
#  fsdp_auto_wrap_policy: TRANSFORMER_BASED_WRAP
#  fsdp_transformer_layer_cls_to_wrap: LlamaDecoderLayer
#  fsdp_state_dict_type: FULL_STATE_DICT
#  fsdp_sharding_strategy: FULL_SHARD
special_tokens:
  pad_token: <|end_of_text|>

Possible solution

No response

Which Operating Systems are you using?

• Linux
• macOS
• Windows

Python Version

3.10

axolotl branch-commit

5783839

Acknowledgements

• My issue title is concise, descriptive, and in title casing.
• I have searched the existing issues to make sure this bug has not been reported yet.
• I am using the latest version of axolotl.
• I have provided enough information for the maintainers to reproduce and diagnose the issue.