fix: free GPU memory between model evaluations

evaluate.py

@@ -1,4 +1,4 @@
-import os, json, re, torch, pathlib
+import os, json, re, torch, pathlib, gc
 from unsloth import FastLanguageModel
 from deceit_env.server.environment import DeceitEnvironment
 from deceit_env.server.grader import Grader
@@ -100,6 +100,11 @@ def evaluate_model(model_name, label, n_episodes=30):
         if (i+1) % 10 == 0:
             print(f"  {i+1}/{n_episodes} done, mean reward so far: {sum(rewards)/len(rewards):.3f}")
 
+    # Free model from GPU
+    del model
+    torch.cuda.empty_cache()
+    gc.collect()
+
     return {
         "label": label,
         "mean_reward": sum(rewards)/len(rewards),
@@ -111,6 +116,14 @@ def evaluate_model(model_name, label, n_episodes=30):
 
 # Evaluate both models
 base_results = evaluate_model("unsloth/Qwen2.5-0.5B-Instruct", "Base Model (untrained)", n_episodes=30)
+
+# Free GPU memory before loading second model
+import gc
+del model
+torch.cuda.empty_cache()
+gc.collect()
+print("GPU memory freed")
+
 trained_results = evaluate_model("Ajsaxena/deceit-qwen-0.5b-full", "DECEIT Trained", n_episodes=30)
 
 # Print comparison