cache clean

modular_graph_and_candidates.py

@@ -41,6 +41,7 @@ from sentence_transformers import SentenceTransformer, util
 from tqdm import tqdm
 import numpy as np
 import spaces
+import torch
 
 # ────────────────────────────────────────────────────────────────────────────────
 # CONFIG
@@ -140,6 +141,12 @@ def embedding_similarity_clusters(models_root: Path, missing: List[str], thr: fl
             emb = model.encode(batch_texts, convert_to_numpy=True, show_progress_bar=False)
             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():
+                torch.cuda.empty_cache()
+                print(f"🧹 Cleared GPU cache after batch {i//batch_size + 1}")
+                
         except Exception as e:
             print(f"⚠️  GPU worker error for batch {batch_names}: {type(e).__name__}: {e}")
             # Create zero embeddings for all models in failed batch