Configure and call dumper in main SGLang logic (#19093)

python/sglang/srt/debug_utils/dump_comparator.py

@@ -20,9 +20,7 @@ def main(args):
     )
     if args.filter:
         df_target = df_target.filter(pl.col("filename").str.contains(args.filter))
-    assert all(
-        c in df_target.columns for c in ["rank", "step", "dump_index", "name"]
-    )
+    assert all(c in df_target.columns for c in ["rank", "step", "dump_index", "name"])
 
     df_baseline = read_meta(args.baseline_path)
     print("df_target", df_target)
@@ -41,9 +39,7 @@ def main(args):
             baseline_step = location_info.baseline_step
             baseline_token_slice = location_info.baseline_token_slice
         else:
-            baseline_step = (
-                row["step"] - args.start_id + args.baseline_start_id
-            )
+            baseline_step = row["step"] - args.start_id + args.baseline_start_id
             baseline_token_slice = None
 
         tensor_dim_desc = None

python/sglang/srt/debug_utils/dumper.py

@@ -168,6 +168,10 @@ class _Dumper:
 
     # ------------------------------- public :: core ---------------------------------
 
+    @property
+    def may_enable(self) -> bool:
+        return self._config.enable or self._config.server_port_parsed is not None
+
     def step(self):
         """This should be called on all ranks at the end of each iteration."""
 
@@ -237,6 +241,7 @@ class _Dumper:
         self,
         model: "torch.nn.Module",
     ) -> Optional["_NonIntrusiveDumper"]:
+        self._ensure_http_server()
         mode = self._config.non_intrusive_mode
         if mode == "off":
             return None

python/sglang/srt/model_executor/model_runner.py

@@ -50,6 +50,7 @@ from sglang.srt.configs.load_config import LoadConfig, LoadFormat
 from sglang.srt.configs.model_config import AttentionArch, ModelConfig, ModelImpl
 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.dumper import dumper
 from sglang.srt.debug_utils.tensor_dump_forward_hook import (
     register_forward_hook_for_model,
 )
@@ -1055,6 +1056,9 @@ class ModelRunner(ModelRunnerKVCacheMixin):
                 self.pp_rank,
             )
 
+        if dumper.may_enable:
+            dumper.register_non_intrusive_dumper(self.model)
+
         # Pre-expand RoPE cache before CUDA Graph capture
         reserve_rope_cache_for_long_sequences(
             self.model,
@@ -2438,6 +2442,9 @@ class ModelRunner(ModelRunnerKVCacheMixin):
         if self.eplb_manager is not None:
             self.eplb_manager.on_forward_pass_end()
 
+        if dumper.may_enable:
+            dumper.step()
+
         return output
 
     def _forward_raw(