Accelerate config file missing for LoRA

[psyhtest]

The LoRA reference implementation has a broken link to an Accelerate config file:

where the Accelerate config file is this one.

[psyhtest]

Or is it the same file stored under configs/default_config.yaml?

[ShriyaPalsamudram]

@itayhubara could you please confirm if the configs/default_config.yaml is the Accelerate config file?

If yes, we should also update the link to avoid confusion in the future.

[regisss]

The link is broken because it points to a private repo and you cannot access it (it's the repo I created at the beginning of the project).
The Accelerate config in my repo is:

compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
  gradient_accumulation_steps: 1
  offload_optimizer_device: none
  offload_param_device: none
  zero3_init_flag: true
  zero3_save_16bit_model: true
  zero_stage: 3
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

While https://github.com/mlcommons/training/blob/master/llama2_70b_lora/configs/default_config.yaml is:

compute_environment: LOCAL_MACHINE
debug: false
deepspeed_config:
  gradient_clipping: 0.3
  gradient_accumulation_steps: 1
  offload_optimizer_device: none
  offload_param_device: none
  zero3_init_flag: true
  zero3_save_16bit_model: true
  zero_stage: 3
distributed_type: DEEPSPEED
downcast_bf16: 'no'
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 8
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

https://github.com/mlcommons/training/blob/master/llama2_70b_lora/configs/default_config.yaml is the right config to use as gradient clipping was added later but I didn't update the config in my repo.

[ShriyaPalsamudram]

@regisss thank you very much for clarifying the difference. Could you please update the link to the accelerate config file in the README by creating a PR to the repo so we can avoid similar issues in the future?

I can help merge that PR for you if you can tag me in it.