Changes to make phi-3 tuning work

AI-Commandos · May 7, 2024 · 1426045 · 1426045
1 parent eacaf86
commit 1426045
Show file tree

Hide file tree

Showing 4 changed files with 13 additions and 10 deletions.
diff --git a/finetune.py b/finetune.py
@@ -111,7 +111,7 @@ def main():
 
     if noquant:
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, device_map={"": 0}, trust_remote_code=True)
     elif quant8:
         quant_config = BitsAndBytesConfig(
             load_in_8bit=True,
@@ -120,7 +120,7 @@ def main():
             bnb_8bit_use_double_quant=False
         )
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0}, trust_remote_code=True)
     else:
         # Set up quantization config
         quant_config = BitsAndBytesConfig(
@@ -130,7 +130,7 @@ def main():
             bnb_4bit_use_double_quant=True,
         )
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0}, trust_remote_code=True)
 
     model.config.use_cache = False
     model.config.pretraining_tp = 1
@@ -150,6 +150,7 @@ def main():
         r=64,
         bias="none",
         task_type="CAUSAL_LM",
+        target_modules = 'all-linear',
     )
 
     # Pass quant and lora to trainer

diff --git a/finetune_dpo.py b/finetune_dpo.py
@@ -113,7 +113,7 @@ def main():
 
     if noquant:
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, device_map={"": 0}, trust_remote_code=True)
     elif quant8:
         quant_config = BitsAndBytesConfig(
             load_in_8bit=True,
@@ -122,7 +122,7 @@ def main():
             bnb_8bit_use_double_quant=False
         )
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0}, trust_remote_code=True)
     else:
         # Set up quantization config
         quant_config = BitsAndBytesConfig(
@@ -132,7 +132,7 @@ def main():
             bnb_4bit_use_double_quant=True,
         )
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0}, trust_remote_code=True)
 
     model.config.use_cache = False
     model.config.pretraining_tp = 1
@@ -154,6 +154,7 @@ def main():
         r=64,
         bias="none",
         task_type="CAUSAL_LM",
+        target_modules='all-linear',
     )
 
     # Pass quant and lora to trainer

diff --git a/finetune_orpo.py b/finetune_orpo.py
@@ -111,7 +111,7 @@ def main():
 
     if noquant:
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, device_map={"": 0}, trust_remote_code=True)
     elif quant8:
         quant_config = BitsAndBytesConfig(
             load_in_8bit=True,
@@ -120,7 +120,7 @@ def main():
             bnb_8bit_use_double_quant=False
         )
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0}, trust_remote_code=True)
     else:
         # Set up quantization config
         quant_config = BitsAndBytesConfig(
@@ -130,7 +130,7 @@ def main():
             bnb_4bit_use_double_quant=True,
         )
         # Load base model
-        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0})
+        model = AutoModelForCausalLM.from_pretrained(base_model, quantization_config=quant_config, device_map={"": 0}, trust_remote_code=True)
 
     model.config.use_cache = False
     model.config.pretraining_tp = 1
@@ -162,6 +162,7 @@ def main():
         report_to=None,
         remove_unused_columns=False,
         beta=0.1, # the lambda/alpha hyperparameter in the paper/code
+        target_modules='all-linear',
     )
 
     trainer = ORPOTrainer(

diff --git a/requirements.txt b/requirements.txt
@@ -1,5 +1,5 @@
 accelerate==0.25.0
-peft==0.6.2
+peft==0.10.0
 bitsandbytes==0.41.2.post2
 transformers==4.36.2
 trl==0.8.2