support discrete profiling for mindspeed

Author: tardis-key

docs/ascend_tutorial/ascend_profiling_en.rst

@@ -1,10 +1,10 @@
-Data collection based on FSDP backend on Ascend devices(en)
+Data collection based on FSDP or MindSpeed(Megatron) on Ascend devices(en)
 ==========================================================================================
 
 Last updated: 08/14/2025.
 
 This is a tutorial for data collection using the GRPO or DAPO algorithm
-based on FSDP on Ascend devices.
+based on FSDP or MindSpeed(Megatron) on Ascend devices.
 
 Configuration
 -------------

docs/ascend_tutorial/ascend_profiling_zh.rst

@@ -1,11 +1,11 @@
-Data collection based on FSDP backend on Ascend devices(zh)
+Data collection based on FSDP or MindSpeed(Megatron) on Ascend devices(zh)
 ====================================
 
-在昇腾设备上基于FSDP后端进行数据采集
+在昇腾设备上基于FSDP或MindSpeed(Megatron)后端进行数据采集
 
 Last updated: 08/14/2025.
 
-这是一份在昇腾设备上基于FSDP后端使用GRPO或DAPO算法进行数据采集的教程。
+这是一份在昇腾设备上基于FSDP或MindSpeed(Megatron)后端，使用GRPO或DAPO算法进行数据采集的教程。
 
 配置
 ----

verl/workers/fsdp_workers.py

@@ -1483,7 +1483,7 @@ def init_model(self):
         )
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="critic"))
-    @DistProfiler.annotate(color="cyan")
+    @DistProfiler.annotate(color="cyan", role="compute_values")
     def compute_values(self, data: DataProto):
         if self._is_offload_param:
             load_fsdp_model_to_gpu(self.critic_module)
@@ -1503,7 +1503,7 @@ def compute_values(self, data: DataProto):
         return output
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="critic"))
-    @DistProfiler.annotate(color="pink")
+    @DistProfiler.annotate(color="pink", role="update_critic")
     def update_critic(self, data: DataProto):
         if self._is_offload_param:
             load_fsdp_model_to_gpu(self.critic_module)
@@ -1867,7 +1867,7 @@ def _switch_chat_template(self, data: DataProto):
         return DataProto.from_dict(rm_inputs)
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="reward"))
-    @DistProfiler.annotate(color="brown")
+    @DistProfiler.annotate(color="brown", role="compute_rm_score")
     def compute_rm_score(self, data: DataProto):
         import itertools
 

verl/workers/megatron_workers.py

@@ -631,7 +631,7 @@ async def trainer_mode(self):
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="actor"))
     @GPUMemoryLogger(role="update_actor", logger=logger)
-    @DistProfiler.annotate(color="red")
+    @DistProfiler.annotate(color="red", role="actor_update")
     def update_actor(self, data: DataProto):
         assert self._is_actor
         if self._is_offload_param:
@@ -674,7 +674,7 @@ def update_actor(self, data: DataProto):
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="rollout"))
     @GPUMemoryLogger(role="generate_sequences", logger=logger)
-    @DistProfiler.annotate(color="red")
+    @DistProfiler.annotate(color="red", role="rollout_generate")
     def generate_sequences(self, prompts: DataProto):
         assert self._is_rollout
         prompts = prompts.to(get_device_name())
@@ -724,7 +724,7 @@ def generate_sequences(self, prompts: DataProto):
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="actor"))
     @GPUMemoryLogger(role="compute_ref_log_prob", logger=logger)
-    @DistProfiler.annotate(color="olive")
+    @DistProfiler.annotate(color="olive", role="ref_compute_log_prob")
     def compute_ref_log_prob(self, data: DataProto):
         assert self._is_ref
         if self._ref_is_offload_param:
@@ -746,7 +746,7 @@ def compute_ref_log_prob(self, data: DataProto):
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="actor"))
     @GPUMemoryLogger(role="compute_log_prob", logger=logger)
-    @DistProfiler.annotate(color="blue")
+    @DistProfiler.annotate(color="blue", role="actor_compute_log_prob")
     def compute_log_prob(self, data: DataProto):
         assert self._is_actor
         if self._is_offload_param:
@@ -1079,7 +1079,7 @@ def init_model(self):
         )
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="critic"))
-    @DistProfiler.annotate(color="cyan")
+    @DistProfiler.annotate(color="cyan", role="compute_values")
     def compute_values(self, data: DataProto):
         micro_batch_size = self.config.ppo_micro_batch_size_per_gpu
         data.meta_info["micro_batch_size"] = micro_batch_size
@@ -1096,7 +1096,7 @@ def compute_values(self, data: DataProto):
         return output
 
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="critic"))
-    @DistProfiler.annotate(color="pink")
+    @DistProfiler.annotate(color="pink", role="update_critic")
     def update_critic(self, data: DataProto):
         data = data.to(get_device_id())
 
@@ -1313,7 +1313,7 @@ def init_model(self):
     # TODO: reward model use itself tokenizer instead of sft tokenizer
     # the input_ids, responses, attention_mask and position_ids may be different!
     @register(dispatch_mode=make_nd_compute_dataproto_dispatch_fn(mesh_name="reward"))
-    @DistProfiler.annotate(color="brown")
+    @DistProfiler.annotate(color="brown", role="compute_rm_score")
     def compute_rm_score(self, data: DataProto):
         data.meta_info["micro_batch_size"] = self.config.micro_batch_size_per_gpu
         data.meta_info["max_token_len"] = self.config.forward_max_token_len_per_gpu