discrete profiler support agent loop

Co-authored-by: tardis-key <[email protected]>

Author: mengchengTang

recipe/fully_async_policy/agent_loop/agent_loop.py

@@ -325,3 +325,9 @@ async def sleep(self):
 
     async def clear_kv_cache(self):
         await asyncio.gather(*[replica.clear_kv_cache() for replica in self.rollout_replicas])
+
+    async def start_profile(self, **kwargs):
+        await asyncio.gather(*[replica.start_profile(**kwargs) for replica in self.rollout_replicas])
+
+    async def stop_profile(self):
+        await asyncio.gather(*[replica.stop_profile() for replica in self.rollout_replicas])

recipe/one_step_off_policy/agent_loop/agent_loop.py

@@ -62,3 +62,9 @@ async def sleep(self):
 
     async def clear_kv_cache(self):
         await asyncio.gather(*[replica.clear_kv_cache() for replica in self.rollout_replicas])
+
+    async def start_profile(self, **kwargs):
+        await asyncio.gather(*[replica.start_profile(**kwargs) for replica in self.rollout_replicas])
+
+    async def stop_profile(self):
+        await asyncio.gather(*[replica.stop_profile() for replica in self.rollout_replicas])

verl/experimental/agent_loop/agent_loop.py

@@ -819,6 +819,7 @@ def generate_sequences(self, prompts: DataProto) -> DataProto:
 
         # Fix for Issue #4147: Always call wake_up() to ensure weight sync
         # The wake_up()/sleep() methods internally check free_cache_engine
+        self.start_profile(role="agent_loop_rollout_generate")
         self.wake_up()
         if self.reward_model_manager:
             self.reward_model_manager.wake_up()
@@ -841,6 +842,7 @@ def generate_sequences(self, prompts: DataProto) -> DataProto:
         timing = self._performance_metrics(metrics, output)
 
         output.meta_info = {"timing": timing, **outputs[0].meta_info}
+        self.stop_profile()
         return output
 
     def _performance_metrics(self, metrics: list[list[dict[str, str]]], output: DataProto) -> dict[str, float]:
@@ -877,6 +879,14 @@ def clear_kv_cache(self):
         """Clear all rollout kv cache, but don`t sleep."""
         self._run_all([replica.clear_kv_cache() for replica in self.rollout_replicas])
 
+    def start_profile(self, **kwargs):
+        """Start profiling on all rollout replicas."""
+        self._run_all([replica.start_profile(**kwargs) for replica in self.rollout_replicas])
+
+    def stop_profile(self):
+        """Stop profiling on all rollout replicas."""
+        self._run_all([replica.stop_profile() for replica in self.rollout_replicas])
+
     def _run_all(self, tasks: list[asyncio.Task]):
         async def run_all():
             await asyncio.gather(*tasks)

verl/utils/profiler/mstx_profile.py

@@ -173,46 +173,59 @@ def __init__(self, rank: int, config: ProfilerConfig, tool_config: NPUToolConfig
             config = ProfilerConfig(ranks=[], enable=False)
         if not tool_config:
             assert not config.enable, "tool_config must be set when profiler is enabled"
-        self.enable: bool = config.enable
-        if not config.enable:
-            return
-        self.this_step: bool = False
         self.discrete: bool = tool_config.discrete
-        self.this_rank: bool = False
-        self.profile_npu = None
+        self.e2e_profile_npu = None
+        self.capture_profiler_npu = None
         self.profile_contents = tool_config.contents
         self.profile_level = tool_config.level
         self.profile_save_path = config.save_path
         self.analysis = tool_config.analysis
-        if config.all_ranks:
-            self.this_rank = True
-        elif config.ranks:
-            self.this_rank = rank in config.ranks
-
-    def start(self, **kwargs):
-        role, profile_step = kwargs.get("role", None), kwargs.get("profile_step", None)
-        profile_step = str(profile_step) if profile_step is not None else None
-        if self.enable and self.this_rank:
-            self.this_step = True
-            if not self.discrete and NPUProfiler._define_count == 0:
-                self.profile_npu = get_npu_profiler(
-                    contents=self.profile_contents,
-                    profile_level=self.profile_level,
-                    profile_save_path=self.profile_save_path,
-                    analysis=self.analysis,
-                    role=role,
-                    profile_step=profile_step,
-                )
-                self.profile_npu.start()
-                NPUProfiler._define_count += 1
-
-    def stop(self):
-        if self.enable and self.this_rank:
-            self.this_step = False
-            if not self.discrete and NPUProfiler._define_count == 1:
-                self.profile_npu.step()
-                self.profile_npu.stop()
-                NPUProfiler._define_count -= 1
+
+    def start_e2e_profiler(self, **kwargs):
+        role = kwargs.get("role", None)
+        if not self.discrete and NPUProfiler._define_count == 0:
+            self.profile_npu = get_npu_profiler(
+                contents=self.profile_contents,
+                profile_level=self.profile_level,
+                profile_save_path=self.profile_save_path,
+                analysis=self.analysis,
+                role=role,
+            )
+            self.profile_npu.start()
+            NPUProfiler._define_count += 1
+
+    def stop_e2e_profiler(self):
+        if not self.discrete and NPUProfiler._define_count == 1:
+            self.profile_npu.step()
+            self.profile_npu.stop()
+            NPUProfiler._define_count -= 1
+
+    def start_capture_profiler(self, **kwargs):
+        """Start an on-demand profiling segment."""
+        role = kwargs.get("role", "")
+
+        if self.discrete:
+            self.capture_profiler_npu = get_npu_profiler(
+                contents=self.profile_contents,
+                profile_level=self.profile_level,
+                profile_save_path=self.profile_save_path,
+                analysis=self.analysis,
+                role=role,
+            )
+            self.capture_profiler_npu.start()
+
+        self._capture_range_id = mark_start_range(message=role)
+
+    def stop_capture_profiler(self):
+        """Stop the on-demand profiling segment."""
+        if hasattr(self, "_capture_range_id"):
+            mark_end_range(self._capture_range_id)
+            del self._capture_range_id
+
+        if self.discrete and getattr(self, "capture_profiler_npu", None):
+            self.capture_profiler_npu.step()
+            self.capture_profiler_npu.stop()
+            del self.capture_profiler_npu
 
     def annotate(self, message: Optional[str] = None, role: Optional[str] = None, **kwargs_outer) -> Callable:
         """Decorate a Worker member function to profile the current rank in the current training step.
@@ -230,42 +243,33 @@ def annotate(self, message: Optional[str] = None, role: Optional[str] = None, **
         def decorator(func):
             @functools.wraps(func)
             def wrapper(*args, **kwargs_inner):
-                if not self.enable:
-                    return func(*args, **kwargs_inner)
-
                 profile_name = message or func.__name__
                 discrete_mode = self.discrete
-                profile_enable = self.this_step and self.enable
-
-                if not profile_enable:
-                    return func(*args, **kwargs_inner)
-
-                if profile_enable:
-                    if not discrete_mode:
-                        mark_range = mark_start_range(message=profile_name)
-                    else:
-                        profile_npu = get_npu_profiler(
-                            contents=self.profile_contents,
-                            profile_level=self.profile_level,
-                            profile_save_path=self.profile_save_path,
-                            analysis=self.analysis,
-                            role=role,
-                        )
-                        profile_npu.start()
-                        mark_range = mark_start_range(message=profile_name)
+
+                if not discrete_mode:
+                    mark_range = mark_start_range(message=profile_name)
+                else:
+                    profile_npu = get_npu_profiler(
+                        contents=self.profile_contents,
+                        profile_level=self.profile_level,
+                        profile_save_path=self.profile_save_path,
+                        analysis=self.analysis,
+                        role=role,
+                    )
+                    profile_npu.start()
+                    mark_range = mark_start_range(message=profile_name)
 
                 result = func(*args, **kwargs_inner)
 
-                if profile_enable:
-                    if not discrete_mode:
-                        mark_end_range(mark_range)
-                    else:
-                        mark_end_range(mark_range)
-                        profile_npu.step()
-                        profile_npu.stop()
+                if not discrete_mode:
+                    mark_end_range(mark_range)
+                else:
+                    mark_end_range(mark_range)
+                    profile_npu.step()
+                    profile_npu.stop()
 
                 return result
 
             return wrapper
 
-        return decorator
+        return decorator

verl/utils/profiler/nvtx_profile.py

@@ -129,25 +129,23 @@ def __init__(self, rank: int, config: Optional[ProfilerConfig], tool_config: Opt
         self.enable = config.enable
         if not config.enable:
             return
-        self.this_step: bool = False
         self.discrete: bool = tool_config.discrete
-        self.this_rank: bool = False
-        if config.all_ranks:
-            self.this_rank = True
-        elif config.ranks:
-            self.this_rank = rank in config.ranks
-
-    def start(self, **kwargs):
-        if self.enable and self.this_rank:
-            self.this_step = True
-            if not self.discrete:
-                torch.cuda.profiler.start()
-
-    def stop(self):
-        if self.enable and self.this_rank:
-            self.this_step = False
-            if not self.discrete:
-                torch.cuda.profiler.stop()
+
+    def start_e2e_profiler(self, **kwargs):
+        if not self.discrete:
+            torch.cuda.profiler.start()
+
+    def stop_e2e_profiler(self):
+        if not self.discrete:
+            torch.cuda.profiler.stop()
+
+    def start_capture_profiler(self, **kwargs):
+        if self.discrete:
+            torch.cuda.profiler.start()
+
+    def stop_capture_profiler(self):
+        if self.discrete:
+            torch.cuda.profiler.stop()
 
     def annotate(
         self,
@@ -176,22 +174,17 @@ def annotate(
         def decorator(func):
             @functools.wraps(func)
             def wrapper(*args, **kwargs_inner):
-                if not self.enable:
-                    return func(*args, **kwargs_inner)
-
                 profile_name = message or func.__name__
 
-                if self.this_step:
-                    if self.discrete:
-                        torch.cuda.profiler.start()
-                    mark_range = mark_start_range(message=profile_name, color=color, domain=domain, category=category)
+                if self.discrete:
+                    torch.cuda.profiler.start()
+                mark_range = mark_start_range(message=profile_name, color=color, domain=domain, category=category)
 
                 result = func(*args, **kwargs_inner)
 
-                if self.this_step:
-                    mark_end_range(mark_range)
-                    if self.discrete:
-                        torch.cuda.profiler.stop()
+                mark_end_range(mark_range)
+                if self.discrete:
+                    torch.cuda.profiler.stop()
 
                 return result
 

verl/utils/profiler/profile.py

@@ -14,6 +14,7 @@
 
 import functools
 import os
+from tabnanny import check
 from typing import Callable, Optional
 
 import torch
@@ -183,6 +184,8 @@ class DistProfiler:
     - torch_memory: Torch CUDA memory snapshot dump
     """
 
+    _this_step = False
+
     def __init__(
         self, rank: int, config: Optional[ProfilerConfig] = None, tool_config: Optional[object] = None, **kwargs
     ):
@@ -192,6 +195,7 @@ def __init__(
 
         self._impl = None
         self._tool = getattr(config, "tool", None)
+        self._enable = config.enable
 
         # Normalize rank selection
         self._this_rank = False
@@ -201,7 +205,7 @@ def __init__(
             self._this_rank = rank in config.ranks
         else:
             # default rank 0 if enabled but ranks unspecified
-            self._this_rank = (rank == 0) if config.enable else False
+            self._this_rank = (rank == 0) if self._enable else False
 
         # Lazy import to avoid circular deps
         if self._tool == "nsys":
@@ -221,11 +225,24 @@ def __init__(
             # Fallback to a no-op impl
             self._impl = _NoOpProfiler()
 
-    def start(self, **kwargs):
-        return getattr(self._impl, "start", lambda **_: None)(**kwargs)
+    def check_enable(self):
+        return self._enable and self._this_rank and DistProfiler._this_step
 
-    def stop(self):
-        return getattr(self._impl, "stop", lambda: None)()
+    def start_e2e_profiler(self, **kwargs):
+        if self.check_enable():
+            return getattr(self._impl, "start_e2e_profiler", lambda **_: None)(**kwargs)
+
+    def stop_e2e_profiler(self):
+        if self.check_enable():
+            return getattr(self._impl, "stop_e2e_profiler", lambda: None)()
+
+    def start_capture_profiler(self, **kwargs):
+        if self.check_enable():
+            return getattr(self._impl, "start_capture_profiler", lambda **_: None)(**kwargs)
+
+    def stop_capture_profiler(self):
+        if self.check_enable():
+            return getattr(self._impl, "stop_capture_profiler", lambda: None)()
 
     @classmethod
     def annotate(
@@ -242,6 +259,8 @@ def wrapper(self_instance, *args, **kwargs_inner):
                 profiler = getattr(self_instance, "profiler", None)
                 if not profiler:
                     return func(self_instance, *args, **kwargs_inner)
+                elif not profiler.check_enable():
+                    return func(self_instance, *args, **kwargs_inner)
 
                 impl = profiler._impl
                 if hasattr(impl, "annotate"):
@@ -361,11 +380,33 @@ def __init__(self, profiler: DistProfiler):
     from verl.single_controller.base.decorator import Dispatch, register
 
     @register(dispatch_mode=Dispatch.ONE_TO_ALL)
-    def start_profile(self, **kwargs) -> None:
+    def start_e2e_profile(self, **kwargs) -> None:
+        """Start profiling for the current rank in the current training step."""
+        self.profiler.start_e2e_profiler(**kwargs)
+
+    @register(dispatch_mode=Dispatch.ONE_TO_ALL)
+    def stop_e2e_profile(self) -> None:
+        """Stop profiling for the current rank in the current training step."""
+        self.profiler.stop_e2e_profiler()
+
+    @register(dispatch_mode=Dispatch.ONE_TO_ALL)
+    def start_capture_profile(self, **kwargs) -> None:
         """Start profiling for the current rank in the current training step."""
         self.profiler.start(**kwargs)
 
     @register(dispatch_mode=Dispatch.ONE_TO_ALL)
-    def stop_profile(self) -> None:
+    def stop_capture_profile(self) -> None:
         """Stop profiling for the current rank in the current training step."""
-        self.profiler.stop()
+        self.profiler.stop_capture_profiler()
+
+    @register(dispatch_mode=Dispatch.DIRECT_ROLLOUT_METHOD)
+    async def start_async_rollout_profile(self, **kwargs):
+        """Start an async rollout profiling segment."""
+        self.rollout.start_async_rollout_profile(**kwargs)
+        return True
+
+    @register(dispatch_mode=Dispatch.DIRECT_ROLLOUT_METHOD)
+    async def stop_async_rollout_profile(self):
+        """Stop the async rollout profiling segment."""
+        self.rollout.stop_async_rollout_profile()
+        return True