Habana/falcon

Optimum Habana is the interface between the Hugging Face Transformers and Diffusers libraries and Habana's Gaudi processor (HPU). It provides a set of tools enabling easy and fast model loading, training and inference on single- and multi-HPU settings for different downstream tasks. Learn more about how to take advantage of the power of Habana HPUs to train and deploy Transformers and Diffusers models at hf.co/hardware/habana.

Llama model HPU configuration

This model only contains the GaudiConfig file for running Falcon models on Habana's Gaudi processors (HPU).

This model contains no model weights, only a GaudiConfig.

This enables to specify:

• use_fused_adam: whether to use Habana's custom AdamW implementation
• use_fused_clip_norm: whether to use Habana's fused gradient norm clipping operator
• use_torch_autocast: whether to use PyTorch's autocast mixed precision

Usage

The model is instantiated the same way as in the Transformers library. The only difference is that there are a few new training arguments specific to HPUs.

Here is a causal language modeling example script to pre-train/fine-tune a model. You can run it with Falcon with the following command:

LOWER_LIST=ops_bf16.txt python3 run_lora_clm.py \
    --model_name_or_path tiiuae/falcon-40b \
    --dataset_name timdettmers/openassistant-guanaco \
    --bf16 True \
    --output_dir ./model_lora_falcon \
    --num_train_epochs 3 \
    --per_device_train_batch_size 1 \
    --per_device_eval_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --evaluation_strategy "no" \
    --save_strategy "no" \
    --learning_rate 3e-4 \
    --max_grad_norm  0.3 \
    --warmup_ratio  0.03 \
    --lr_scheduler_type "constant" \
    --logging_steps 1 \
    --do_train \
    --use_habana \
    --use_lazy_mode \
    --pipelining_fwd_bwd \
    --throughput_warmup_steps 3 \
    --lora_rank=64 \
    --lora_alpha=16 \
    --lora_dropout=0.1 \
    --lora_target_modules "query_key_value" "dense" "dense_h_to_4h" "dense_4h_to_h" \
    --dataset_concatenation \
    --max_seq_length 256 \
    --low_cpu_mem_usage True \
    --adam_epsilon 1e-08 \
    --do_eval \
    --validation_split_percentage 5

You will need to install the PEFT library with pip install peft to run the command above.

Check the documentation out for more advanced usage and examples.