Fix the perf regression due to additional_stop_token_ids (#1773)

python/sglang/srt/layers/sampler.py

@@ -42,11 +42,11 @@ class Sampler(nn.Module):
         logits = logits.contiguous()
 
         if self.use_nan_detectioin and torch.any(torch.isnan(logits)):
-            exit(1) if crash_on_warning else None
             logger.warning("Detected errors during sampling! NaN in the logits.")
             logits = torch.where(
                 torch.isnan(logits), torch.full_like(logits, -1e5), logits
             )
+            exit(1) if crash_on_warning else None
 
         if sampling_info.is_all_greedy:
             # Use torch.argmax if all requests use greedy sampling