fix minor bug

finetune.py

@@ -36,7 +36,7 @@ LEARNING_RATE = 2e-4
 CUTOFF_LEN = 256
 LORA_R = 8
 LORA_ALPHA = 16
-LORA_DROPOUT = 0.05
+LORA_DROPOUT = 0.05 # should be 0 if gradient checkpointing is on
 VAL_SET_SIZE = 0
 TARGET_MODULES = [
     "q_proj",
@@ -49,6 +49,10 @@ save_steps = 50
 save_total_limit = 3
 logging_steps = 10
 
+if LORA_DROPOUT > 0 and GRADIENT_CHECKPOINTING:
+    LORA_DROPOUT = 0
+    print('Disable Dropout.')
+
 # Load Basic Model
 model, tokenizer = load_llama_model_4bit_low_ram(config_path, model_path)
 

install.bat

@@ -15,8 +15,9 @@ REM replace ./repository/GPTQ-for-LLaMa/quant_cuda.cpp and quant_cuda_kernel.cu
 copy .\GPTQ-for-LLaMa\quant_cuda.cpp .\repository\GPTQ-for-LLaMa\quant_cuda.cpp /Y
 copy .\GPTQ-for-LLaMa\quant_cuda_kernel.cu .\repository\GPTQ-for-LLaMa\quant_cuda_kernel.cu /Y
 
-REM copy autograd_4bit.py into ./repository/GPTQ-for-LLaMa/autograd_4bit.py
+REM copy files into ./repository/GPTQ-for-LLaMa/
 copy .\GPTQ-for-LLaMa\autograd_4bit.py .\repository\GPTQ-for-LLaMa\autograd_4bit.py /Y
+copy .\GPTQ-for-LLaMa\gradient_checkpointing.py .\repository\GPTQ-for-LLaMa\gradient_checkpointing.py /Y
 
 REM install quant_cuda
 cd .\repository\GPTQ-for-LLaMa

install.sh

@@ -19,8 +19,9 @@ cp ./peft/tuners/lora.py ./repository/peft/src/peft/tuners/lora.py
 cp ./GPTQ-for-LLaMa/quant_cuda.cpp ./repository/GPTQ-for-LLaMa/quant_cuda.cpp
 cp ./GPTQ-for-LLaMa/quant_cuda_kernel.cu ./repository/GPTQ-for-LLaMa/quant_cuda_kernel.cu
 
-# Copy autograd_4bit.py into ./repository/GPTQ-for-LLaMa/autograd_4bit.py
+# Copy files into ./repository/GPTQ-for-LLaMa/
 cp ./GPTQ-for-LLaMa/autograd_4bit.py ./repository/GPTQ-for-LLaMa/autograd_4bit.py
+cp ./GPTQ-for-LLaMa/gradient_checkpointing.py ./repository/GPTQ-for-LLaMa/gradient_checkpointing.py
 
 # Install quant_cuda and cd into ./repository/GPTQ-for-LLaMa
 cd ./repository/GPTQ-for-LLaMa