feat: Add Non-intrusive Tensor Dumping for Model Inference (#10566)

python/sglang/srt/server_args.py

@@ -511,6 +511,9 @@ class ServerArgs:
 
     # Debug tensor dumps
     debug_tensor_dump_output_folder: Optional[str] = None
+    # -1 mean dump all layers.
+    debug_tensor_dump_layers: int = -1
+    # TODO(guoyuhong): clean the old dumper code.
     debug_tensor_dump_input_file: Optional[str] = None
     debug_tensor_dump_inject: bool = False
 
@@ -1784,7 +1787,13 @@ class ServerArgs:
                 )
 
     def _handle_other_validations(self):
-        pass
+        # Handle model inference tensor dump.
+        if self.debug_tensor_dump_output_folder is not None:
+            logger.warning(
+                "Cuda graph and server warmup are disabled because of using tensor dump mode"
+            )
+            self.disable_cuda_graph = True
+            self.skip_server_warmup = True
 
     @staticmethod
     def add_cli_args(parser: argparse.ArgumentParser):
@@ -3375,6 +3384,12 @@ class ServerArgs:
             default=ServerArgs.debug_tensor_dump_output_folder,
             help="The output folder for dumping tensors.",
         )
+        parser.add_argument(
+            "--debug-tensor-dump-layers",
+            type=int,
+            default=-1,
+            help="The layer number for dumping tensors.",
+        )
         parser.add_argument(
             "--debug-tensor-dump-input-file",
             type=str,