Update run_eval.py

run_eval.py

@@ -2,6 +2,7 @@
 import datetime, os, subprocess, tempfile
 from pathlib import Path
 
+import gc
 import pandas as pd, yaml, torch
 from huggingface_hub import HfApi, login, hf_hub_download, model_info
 from lm_eval import evaluator
@@ -11,8 +12,10 @@ from transformers import (
     AutoModelForCausalLM,
     AutoModelForSequenceClassification,
     AutoTokenizer,
+    BitsAndBytesConfig 
 )
 
+
 CONFIGS = []
 
 # ───── Load all configs ─────
@@ -67,6 +70,8 @@ for cfg in CONFIGS:
     try:
         base_model = AutoModelForCausalLM.from_pretrained(
             base_model_id,
+            device_map="auto", 
+            torch_dtype=torch.float16,
             trust_remote_code=True,
             use_safetensors=True
         )
@@ -76,6 +81,8 @@ for cfg in CONFIGS:
         print(f"⚠️ Failed to load causal LM: {e}")
         base_model = AutoModelForSequenceClassification.from_pretrained(
             base_model_id,
+            device_map="auto", 
+            torch_dtype=torch.float16,
             trust_remote_code=True,
             use_safetensors=True
         )
@@ -93,14 +100,17 @@ for cfg in CONFIGS:
         continue
 
     try:
-        peft_model = PeftModel.from_pretrained(base_model, adapter_repo)
+        peft_model = PeftModel.from_pretrained(
+            base_model,
+            adapter_repo,
+            device_map="auto",
+            torch_dtype=torch.float16,
+        )        
         merged_model = peft_model.merge_and_unload()
     except Exception as e:
         print(f"Failed to apply adapter {adapter_repo}: {e}")
         continue
 
-    device = "cuda" if torch.cuda.is_available() else "cpu"
-    merged_model.to(device)
     merged_model.eval()
 
     with tempfile.TemporaryDirectory() as td: