again...

app.py

@@ -26,12 +26,12 @@ from warbler_cda.pack_loader import PackLoader
 print("🚀 Initializing Warbler CDA...")
 
 # Initialize the system components
-print("⚙️ Creating embedding provider (ZeroGPU compatible)...")
+print("⚙️ Creating embedding provider with multi-worker GPU acceleration...")
 embedding_provider = EmbeddingProviderFactory.create_provider("sentence_transformer", {
-    "num_workers": 1,  # Single worker to avoid tokenizer parallelism issues
-    "batch_size": 16,   # Small batches for ZeroGPU stability
-    "cache_dir": ".embedding_cache",
-    "device": "cpu"    # Force CPU to avoid ZeroGPU device conflicts
+    "num_workers": 4,  # Restore multi-worker for speed (device test will handle ZeroGPU)
+    "batch_size": 64,   # Larger batches for better throughput
+    "cache_dir": ".embedding_cache"
+    # Let device auto-detection handle ZeroGPU (will fall back to CPU if needed)
 })
 print(f"✅ Embedding provider: {embedding_provider.get_provider_info()['provider_id']}")
 

warbler_cda/embeddings/sentence_transformer_provider.py

@@ -60,10 +60,25 @@ class SentenceTransformerEmbeddingProvider(EmbeddingProvider):
 
             import torch
 
-            self.device = "cuda" if torch.cuda.is_available() else "cpu"
+            # Handle ZeroGPU and other GPU environments
+            has_cuda = torch.cuda.is_available()
+            if has_cuda:
+                try:
+                    # Test if we can actually use CUDA without conflicts
+                    test_tensor = torch.tensor([1.0], device="cuda")
+                    test_tensor.cpu()  # Test device transfer
+                    self.device = "cuda"
+                except RuntimeError as e:
+                    # ZeroGPU or other GPU conflicts - fall back to CPU
+                    print(f"CUDA available but has conflicts ({e}), using CPU")
+                    self.device = "cpu"
+            else:
+                self.device = "cpu"
+
             self.model = SentenceTransformer(self.model_name, device=self.device)
             if self.model is not None:
                 self.dimension = self.model.get_sentence_embedding_dimension()
+                print(f"Initialized {self.model_name} on {self.device}")
 
         except ImportError:
             self.model = None