enable nsys

Signed-off-by: Leo Tian <lctian@nvidia.com>

Author: leo-cf-tian

src/srtctl/backends/vllm.py

@@ -331,6 +331,9 @@ def build_worker_command(
             ]
         )
 
+        if nsys_prefix is not None:
+            cmd.extend(["--profiler-config.profiler", "cuda"])
+
         # Disaggregation mode (dynamo 1.0.0+: --is-prefill-worker/--is-decode-worker are deprecated)
         if mode in ("prefill", "decode"):
             cmd.extend(["--disaggregation-mode", mode])

src/srtctl/core/schema.py

@@ -644,8 +644,10 @@ def get_env_vars(self, mode: str, profile_dir: str) -> dict[str, str]:
             phase_key = mode.upper() if mode != "agg" else "AGG"
             if phase_config.start_step is not None:
                 env[f"PROFILE_{phase_key}_START_STEP"] = str(phase_config.start_step)
+                env["VLLM_PROFILER_DELAY_ITERS"] = str(phase_config.start_step)
             if phase_config.stop_step is not None:
                 env[f"PROFILE_{phase_key}_STOP_STEP"] = str(phase_config.stop_step)
+                env["VLLM_PROFILER_MAX_ITERS"] = str(phase_config.stop_step - phase_config.start_step)
 
         if self.is_torch:
             env["SGLANG_TORCH_PROFILER_DIR"] = f"{profile_dir}/{mode}"