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

python/sglang/srt/model_executor/model_runner.py

@@ -40,6 +40,9 @@ from sglang.srt.configs.model_config import (
 )
 from sglang.srt.configs.update_config import adjust_config_with_unaligned_cpu_tp
 from sglang.srt.constants import GPU_MEMORY_TYPE_WEIGHTS
+from sglang.srt.debug_utils.tensor_dump_forward_hook import (
+    register_forward_hook_for_model,
+)
 from sglang.srt.distributed import (
     get_pp_group,
     get_tp_group,
@@ -791,6 +794,15 @@ class ModelRunner:
             f"avail mem={after_avail_memory:.2f} GB, "
             f"mem usage={self.weight_load_mem_usage:.2f} GB."
         )
+        if self.server_args.debug_tensor_dump_output_folder is not None:
+            register_forward_hook_for_model(
+                self.model,
+                self.server_args.debug_tensor_dump_output_folder,
+                self.server_args.debug_tensor_dump_layers,
+                self.tp_size,
+                self.tp_rank,
+                self.pp_rank,
+            )
 
         if self.server_args.elastic_ep_backend == "mooncake":
             # Mooncake does not support `monitored_barrier`