accelerated-peft-autogptq-sample-configuration.yaml

# FMS Acceleration Plugin Configuration. 
#
# Each stanza incorporates various configurations for 
# different fine-tuning / training tasks.
plugins:
  # PEFT-related acceleration
  peft:

    # quantization-releated acceleration
    # e.g., kernels for quantized base weights
    quantization:

      # AutoGPTQ quantized base weights.
      auto_gptq:

        # Kernel to be used for GPTQ linear laeyer
        # NOTE: Not all kernels are suitable for PEFT training; need to use 
        # kernels that support autograd forward / backward. The best 
        # recommendation at the moment is "triton_v2".
        kernel: triton_v2

        # If true, then will already expect quantized checkpoint 
        # passed into TrainingArguments.model_name_or_path
        from_quantized: true