More debugging

app.py

@@ -41,7 +41,7 @@ tokenizer = AutoTokenizer.from_pretrained("bigcode/starcoderbase-3b")
 vardecoder_model = AutoModelForCausalLM.from_pretrained(
     "ejschwartz/resym-vardecoder", 
     torch_dtype=torch.bfloat16,
-    device_map="cuda",
+    device_map="cpu",
 )
 print("Loaded vardecoder model successfully.")
 
@@ -50,12 +50,26 @@ print(f"Model dtype: {next(vardecoder_model.parameters()).dtype}")
 print(f"Model is meta: {next(vardecoder_model.parameters()).is_meta}")
 print(f"Model parameters: {sum(p.numel() for p in vardecoder_model.parameters() if p.requires_grad):,}")
 
+# Check if parameters actually have data
+sample_param = next(vardecoder_model.parameters())
+print(f"Sample parameter shape: {sample_param.shape}")
+print(f"Sample parameter requires_grad: {sample_param.requires_grad}")
+print(f"Sample parameter data type: {type(sample_param.data)}")
+print(f"Sample parameter storage: {sample_param.storage()}")
 
 # Check memory after first model
 print(f"GPU memory after vardecoder:")
 print(f"Allocated: {torch.cuda.memory_allocated() / 1024**3:.2f} GB")
 print(f"Reserved: {torch.cuda.memory_reserved() / 1024**3:.2f} GB")
 
+# Try manual memory allocation test
+test_tensor = torch.randn(1000, 1000, dtype=torch.bfloat16, device='cuda')
+print(f"GPU memory after test tensor:")
+print(f"Allocated: {torch.cuda.memory_allocated() / 1024**3:.2f} GB")
+print(f"Reserved: {torch.cuda.memory_reserved() / 1024**3:.2f} GB")
+del test_tensor
+torch.cuda.empty_cache()
+
 # Add more detailed debugging before loading the second model
 try:
     logger.info("Loading fielddecoder model...")
@@ -67,7 +81,7 @@ try:
     fielddecoder_model = AutoModelForCausalLM.from_pretrained(
         "ejschwartz/resym-fielddecoder", 
         torch_dtype=torch.bfloat16,
-        device_map="cuda",
+        device_map="cpu",
     )
     logger.info("Successfully loaded fielddecoder model")
 except Exception as e: