feat(profiling): add extra_nsys_args for optional nsys CLI flags

docs/config-reference.md

@@ -661,6 +661,9 @@ Profiling configuration for nsys or torch profiler.
 profiling:
   type: "nsys"                       # "none", "nsys", or "torch"
 
+  # Extra arguments for nsys profile (when type is nsys or nsys-time)
+  extra_nsys_args: ["--stats=true"]       # Optional: list of strings
+
   # Phase-specific profiling step configs
   prefill:
     start_step: 10                   # Step to start profiling
@@ -677,6 +680,7 @@ profiling:
 | Field         | Type   | Required | Default | Description                              |
 | ------------- | ------ | -------- | ------- | ---------------------------------------- |
 | `type`        | string | No       | "none"  | Profiling type: "none", "nsys", "torch"  |
+| `extra_nsys_args` | list[string] | No | null | Extra args for nsys profile (when type is `nsys` or `nsys-time`) |
 | `prefill`     | object | Disaggregated | null | Prefill phase config                   |
 | `decode`      | object | Disaggregated | null | Decode phase config                    |
 | `aggregated`  | object | Aggregated | null | Aggregated phase config                  |
@@ -730,6 +734,7 @@ resources:
 
 profiling:
   type: "nsys"
+  extra_nsys_args: ["--stats=true", "--trace=osrt"]
   aggregated:
     start_step: 10
     stop_step: 25

docs/profiling.md

@@ -56,6 +56,9 @@ profiling:
 profiling:
   type: "torch" # Required: "none", "torch", or "nsys"
 
+  # nsys / nsys-time: extra arguments for nsys profile (e.g. ["--stats=true"])
+  extra_nsys_args: []  # Optional
+
   # Disaggregated mode: must set both prefill and decode sections
   prefill:
     start_step: 0 # Step to start profiling for prefill workers
@@ -111,10 +114,13 @@ When using `nsys`, workers are wrapped with:
 ```bash
 nsys profile -t cuda,nvtx --cuda-graph-trace=node \
   -c cudaProfilerApi --capture-range-end stop \
+  [extra_nsys_args...] \
   -o /logs/profiles/{mode}/{name} \
   python3 -m sglang.launch_server ...
 ```
 
+You can pass extra arguments via `profiling.extra_nsys_args` (e.g. `["--stats=true", "--trace=osrt"]`).
+
 ## Example Configurations
 
 ### Torch Profiler (Recommended for Python analysis)

src/srtctl/core/schema.py

@@ -655,6 +655,9 @@ class ProfilingConfig:
 
     type: str = "none"  # "none", "nsys", "nsys-time", or "torch"
 
+    # Extra arguments passed to nsys profile (appended before `-o`; see get_nsys_prefix)
+    extra_nsys_args: list[str] | None = None
+
     # Phase-specific profiling step configs (not used for nsys-time)
     prefill: ProfilingPhaseConfig | None = None
     decode: ProfilingPhaseConfig | None = None
@@ -768,6 +771,9 @@ def _get_nsys_prefix_trtllm(self, output_file: str) -> list[str]:
                 "stop",
             ]
 
+        if self.extra_nsys_args:
+            cmd.extend(self.extra_nsys_args)
+
         cmd += [
             "--kill",
             "none",
@@ -814,10 +820,13 @@ def get_nsys_prefix(
             "stop",
             "--force-overwrite",
             "true",
-            "-o",
-            output_file,
         ]
 
+        if self.extra_nsys_args:
+            cmd.extend(self.extra_nsys_args)
+
+        cmd.extend(["-o", output_file])
+
         if frontend_type == "dynamo":
             cmd.insert(-2, "--trace-fork-before-exec=true")
 

tests/test_profiling.py

@@ -47,6 +47,38 @@ def test_nsys_profiling(self):
         prefix_router = profiling.get_nsys_prefix("/output/test", frontend_type="sglangrouter")
         assert "--trace-fork-before-exec=true" not in prefix_router
 
+    def test_nsys_profiling_with_extra_args(self):
+        """Test nsys profiling with custom extra_nsys_args."""
+        from srtctl.core.schema import ProfilingConfig
+
+        profiling = ProfilingConfig(
+            type="nsys",
+            extra_nsys_args=["--stats=true", "--trace=osrt"],
+        )
+
+        prefix = profiling.get_nsys_prefix("/output/test")
+        assert "nsys" in prefix
+        assert "profile" in prefix
+        assert "/output/test" in prefix
+        assert "--stats=true" in prefix
+        assert "--trace=osrt" in prefix
+        # Extra args appear before -o output
+        o_idx = prefix.index("-o")
+        stats_idx = prefix.index("--stats=true")
+        assert stats_idx < o_idx
+
+    def test_nsys_trtllm_prefix_includes_extra_args(self):
+        """TRTLLM nsys wrap should honor extra_nsys_args (same ordering as default path: before -o)."""
+        from srtctl.core.schema import ProfilingConfig
+
+        profiling = ProfilingConfig(
+            type="nsys",
+            extra_nsys_args=["--stats=true"],
+        )
+        prefix = profiling.get_nsys_prefix("/out/rank", backend_type="trtllm")
+        assert "--stats=true" in prefix
+        assert prefix.index("--stats=true") < prefix.index("-o")
+
     def test_torch_profiling(self):
         """Test torch profiling configuration."""
         from srtctl.core.schema import ProfilingConfig, ProfilingPhaseConfig