fms-acceleration-peft 0.1.1

FMS Acceleration for PeFT

FMS Acceleration for Accelerated PeFT Techniques

Currently only supports LoRA-related techniques, but more are in the pipeline to be added:

Plugins

Plugin	Description	Depends	Loading	Augmentation	Callbacks
autogptq	Loads 4bit GPTQ-LoRA with quantized GPTQ as base	AutoGPTQ	✅	✅
bnb	Loads 4bit QLoRA with quantized bitsandbytes Linear4	Huggingface bitsandbytes	✅	✅

Key Points

• fix upcasting (resulting in slowdown) issue for bnb plugin, originally discovered by inventors of Unsloth.
• bnb properly configured to work with FSDP following this guide.
• triton_v2 kernels are not yet properly integrated into huggingface optimum.
• triton_v2 kernels are the only 4bit kernels that work for training.

GPTQ-LORA's AutoGPTQ - Current Implementation vs Legacy Implementation

GPTQ-LORA depends on an AutoGPTQ backend to run. There are 2 backend options

1. Current Implementation
   • This is an extracted local subset from ModelCloud's refactored fork.
   • It removes redundant code to simplify build and installation of the plugin
2. Legacy Implementation

   • This requires building the package from the official AutoGPTQ repository

   • To replicate this implementation, follow the installation below

     • The legacy implementation of GPTQ-LORA uses an external AutoGPTQ package, you must ensure the specific commit is installed

       pip install git+https://github.com/AutoGPTQ/AutoGPTQ.git@ea829c7bbe83561c2b1de26795b6592992373ef7
       

     • To construct the plugin, in the configuration object that is passed to the plugin - set use_external_lib: True (otherwise defaults to use the local AutoGPTQ package)

         peft:
         quantization: 
             auto_gptq:
             kernel: triton_v2
             from_quantized: True
             use_external_lib: True
     

Known Issues

• Models with sliding windows (e.g., Mistral, Mixtral) will have memory and throughout issues.
• GPTQ-LORA sometimes observed to have nan grad norms in the begining of training, but training proceeds well otherwise.
• low_cpu_mem_usage temporarily disabled for AutoGPTQ until bug with make_sure_no_tensor_in_meta_device is resolved.