Mistral: Sliding Window Attention with Flash Attention and Sample Packing

[casper-hansen]

Main benefits of this PR:

• Loss is reduced 5x on long-context datasets due to passing window_size to Flash Attention.
• Memory usage is reduced by 3GB due to the usage of a sliding window mask.

Memory usage

Memory usage with SWA. The conclusion is that you save 3GB when using a sliding window mask.

• _prepare_decoder_attention_mask with window_size=(4096, 4096) parameter to flash attention.
  • PR, with sliding window mask: 4.851GB (+24.606GB cache, +0.781GB misc)
  • PR, without sliding window mask: 4.851GB (+27.778GB cache, +0.781GB misc)
• _prepare_decoder_attention_mask with window_size=(-1, -1) parameter to flash attention.
  • PR, with sliding window mask: 4.851GB (+24.606GB cache, +0.781GB misc)
  • Main, without sliding window mask: 4.851GB (+27.778GB cache, +0.781GB misc)

Long context (casperhansen/longalpaca_1k_test)

I test with a long context dataset, minimum 16k tokens and maximum 32k tokens. Minimum 48GB VRAM needed to run this.

Results after a few steps:

• PR: Loss starts at 1.9, goes down to 1.54 after 4 steps, 1.396 after 19 steps
• Main: Loss starts at 9.98, goes down to 9.11 after 4 steps

base_model: mistralai/Mistral-7B-v0.1
base_model_config: mistralai/Mistral-7B-v0.1
model_type: MistralForCausalLM
tokenizer_type: LlamaTokenizer
is_mistral_derived_model: true

load_in_8bit: false
load_in_4bit: true
strict: false

datasets:
  - path: casperhansen/longalpaca_1k_test
    type: alpaca
dataset_prepared_path: last_run_prepared
val_set_size: 0.01
output_dir: ./qlora-out

adapter: qlora
lora_model_dir:

sequence_len: 32768
sample_packing: true
pad_to_sequence_len: true

lora_r: 32
lora_alpha: 16
lora_dropout: 0.05
lora_target_linear: true
lora_fan_in_fan_out:
lora_target_modules:
  - gate_proj
  - down_proj
  - up_proj
  - q_proj
  - v_proj
  - k_proj
  - o_proj

wandb_mode: 
wandb_project: 
wandb_entity: 
wandb_watch: 
wandb_run_id: 
wandb_log_model:

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_bnb_8bit
lr_scheduler: cosine
learning_rate: 0.0002

train_on_inputs: false
group_by_length: false
bf16: true
fp16: false
tf32: false

gradient_checkpointing: true
early_stopping_patience:
resume_from_checkpoint:
local_rank:
logging_steps: 1
xformers_attention:
flash_attention: true

warmup_steps: 10
eval_steps: 20
eval_table_size: 5
eval_table_max_new_tokens: 128
save_steps:
debug:
deepspeed:
weight_decay: 0.0
fsdp:
fsdp_config:
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"

Short context (mhenrichsen/alpaca_2k_test)

Loss on short-context datasets is tested to be the same.

Used default config in examples/mistral/qlora.yml.

Other notes:

• attention_mask and sliding_window_mask are not broadcastable in the first iteration after eval loss. However, this is only the case when wandb is enabled. This error is handled by attention_mask.shape[0] != 1 so that it does not trigger.
• I tried to use _expand_mask and it did not work with Flash Attention. I tried other methods too, but same problem.

[TJ-Solergibert]

Hi! I'm trying to fine-tune Mistral 7b with a chat dataset, packing, train on competitions only and FA. I got spammed with the warning skipping sliding window mask, not broadcastable with attention mask. I don't understand why this is happening in the first iteration after eval loss and only when wandb is activated. Should I worry for the performance of the model?

Thanks,

Toni