Enhance training script and evaluation logic to support all-future model target mode and improve error handling for distribution modes

evaluate_auc.py

@@ -632,13 +632,15 @@ def compute_logits_for_disease_chunk(
         device.type == "cuda" and use_amp) else torch.float32
     weight = model.risk_head.weight[disease_ids].detach().to(
         device=device, dtype=compute_dtype)
+    bias = model.risk_head.bias[disease_ids].detach().to(
+        device=device, dtype=compute_dtype)
 
     parts: List[np.ndarray] = []
     for start in tqdm(range(0, n, logit_batch_size), desc="Risk-head projection", leave=False, dynamic_ncols=True):
         end = min(start + logit_batch_size, n)
         h = torch.from_numpy(hidden_all[start:end]).to(
             device=device, dtype=compute_dtype, non_blocking=True)
-        logits = torch.matmul(h, weight.t())
+        logits = torch.matmul(h, weight.t()) + bias
         parts.append(logits.float().cpu().numpy().astype(
             np.float32, copy=False))
         del h, logits