gpu optims

modular_graph_and_candidates.py

@@ -130,7 +130,7 @@ def embedding_similarity_clusters(models_root: Path, missing: List[str], thr: fl
     all_embeddings = []
 
     print(f"Encoding embeddings for {len(names)} models...")
-    batch_size = 8
+    batch_size = 4  # Reduced to be more conservative
     
     for i in tqdm(range(0, len(names), batch_size), desc="Batches", leave=False):
         batch_names = names[i:i+batch_size]
@@ -142,9 +142,10 @@ def embedding_similarity_clusters(models_root: Path, missing: List[str], thr: fl
             all_embeddings.append(emb)
             print(f"✓ Completed batch of {len(batch_names)} models")
             
-            # Clear GPU cache every 5 batches to prevent memory accumulation
-            if i % (5 * batch_size) == 0 and torch.cuda.is_available():
+            # Clear GPU cache every 3 batches to prevent memory accumulation
+            if i % (3 * batch_size) == 0 and torch.cuda.is_available():
                 torch.cuda.empty_cache()
+                torch.cuda.synchronize()  # Force GPU sync
                 print(f"🧹 Cleared GPU cache after batch {i//batch_size + 1}")
                 
         except Exception as e: