Peft fix.

usnistgov · Oct 4, 2024 · f171615 · f171615
1 parent be50121
commit f171615
Showing 1 changed file with 27 additions and 25 deletions.
diff --git a/atomgpt/inverse_models/inverse_models.py b/atomgpt/inverse_models/inverse_models.py
@@ -19,6 +19,8 @@
 from pydantic_settings import BaseSettings
 import sys
 import argparse
+from peft import PeftModelForCausalLM
+
 
 parser = argparse.ArgumentParser(
     description="Atomistic Generative Pre-trained Transformer."
@@ -38,7 +40,7 @@ class TrainingPropConfig(BaseSettings):
     prefix: str = "atomgpt_run"
     model_name: str = "unsloth/mistral-7b-bnb-4bit"
     batch_size: int = 2
-    num_epochs: int = 5
+    num_epochs: int = 2
     seed_val: int = 42
     num_train: Optional[int] = 2
     num_val: Optional[int] = 2
@@ -164,7 +166,7 @@ def text2atoms(response):
     return atoms
 
 
-def gen_atoms(prompt="", max_new_tokens=512, model="", tokenizer=""):
+def gen_atoms(prompt="", max_new_tokens=2048, model="", tokenizer=""):
     inputs = tokenizer(
         [
             alpaca_prompt.format(
@@ -179,9 +181,7 @@ def gen_atoms(prompt="", max_new_tokens=512, model="", tokenizer=""):
     outputs = model.generate(
         **inputs, max_new_tokens=max_new_tokens, use_cache=True
     )
-    response = tokenizer.batch_decode(outputs)
-    print("response", response)
-    response = response[0].split("# Output:")[1]
+    response = tokenizer.batch_decode(outputs)[0].split("# Output:")[1]
     atoms = None
     try:
         atoms = text2atoms(response)
@@ -263,26 +263,28 @@ def run_atomgpt_inverse(config_file="config.json"):
         # token = "hf_...", # use one if using gated models like meta-llama/Llama-2-7b-hf
     )
 
-    model = FastLanguageModel.get_peft_model(
-        model,
-        r=16,  # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
-        target_modules=[
-            "q_proj",
-            "k_proj",
-            "v_proj",
-            "o_proj",
-            "gate_proj",
-            "up_proj",
-            "down_proj",
-        ],
-        lora_alpha=16,
-        lora_dropout=0,  # Supports any, but = 0 is optimized
-        bias="none",  # Supports any, but = "none" is optimized
-        use_gradient_checkpointing=True,
-        random_state=3407,
-        use_rslora=False,  # We support rank stabilized LoRA
-        loftq_config=None,  # And LoftQ
-    )
+    if not isinstance(model, PeftModelForCausalLM):
+
+        model = FastLanguageModel.get_peft_model(
+            model,
+            r=16,  # Choose any number > 0 ! Suggested 8, 16, 32, 64, 128
+            target_modules=[
+                "q_proj",
+                "k_proj",
+                "v_proj",
+                "o_proj",
+                "gate_proj",
+                "up_proj",
+                "down_proj",
+            ],
+            lora_alpha=16,
+            lora_dropout=0,  # Supports any, but = 0 is optimized
+            bias="none",  # Supports any, but = "none" is optimized
+            use_gradient_checkpointing=True,
+            random_state=3407,
+            use_rslora=False,  # We support rank stabilized LoRA
+            loftq_config=None,  # And LoftQ
+        )
 
     EOS_TOKEN = tokenizer.eos_token  # Must add EOS_TOKEN