Enable user marker for vllm profiling (ROCm#357)

* Enable user marker for vllm profiling

---------

Co-authored-by: Gregory Shtrasberg <[email protected]>

Author: Lzy17

vllm/utils.py

@@ -282,6 +282,37 @@ def inner(*args, **kwds):
             return func
 
 
+class rpd_user_marker:
+
+    def __init__(self, name=None):
+        self.name = name
+        self.marker = None
+
+    def __enter__(self):
+        if is_hipScopedMarker_available():
+            from hipScopedMarker import hipScopedMarker
+            marker_name = self.name if self.name else "UserMarker Undefined"
+            self.marker = hipScopedMarker(f"{marker_name}")
+            self.marker.__enter__()
+        return self
+
+    def __exit__(self, exc_type, exc_val, exc_tb):
+        if is_hipScopedMarker_available() and self.marker:
+            self.marker.__exit__(exc_type, exc_val, exc_tb)
+
+    def start(self):
+        if is_hipScopedMarker_available():
+            from hipScopedMarker import hipScopedMarker
+            marker_name = self.name if self.name else "UserMarker Undefined"
+            self.marker = hipScopedMarker(f"{marker_name}")
+            self.marker.__enter__()
+        return self
+
+    def end(self, exc_type=0, exc_val=0, exc_tb=0):
+        if is_hipScopedMarker_available() and self.marker:
+            self.marker.__exit__(exc_type, exc_val, exc_tb)
+
+
 class Device(enum.Enum):
     GPU = enum.auto()
     CPU = enum.auto()

vllm/worker/model_runner.py

@@ -47,8 +47,8 @@
 from vllm.sequence import IntermediateTensors, SequenceGroupMetadata
 from vllm.utils import (DeviceMemoryProfiler, GiB_bytes, PyObjectCache,
                         async_tensor_h2d, flatten_2d_lists,
-                        is_pin_memory_available, rpd_mark, supports_dynamo,
-                        weak_ref_tensor)
+                        is_pin_memory_available, rpd_mark, rpd_user_marker,
+                        supports_dynamo, weak_ref_tensor)
 from vllm.worker.model_runner_base import (
     ModelRunnerBase, ModelRunnerInputBase, ModelRunnerInputBuilderBase,
     _add_attn_metadata_broadcastable_dict,
@@ -1630,6 +1630,12 @@ def execute_model(
         assert model_input.attn_metadata is not None
         prefill_meta = model_input.attn_metadata.prefill_metadata
         decode_meta = model_input.attn_metadata.decode_metadata
+        if prefill_meta:
+            marker_instance = rpd_user_marker(name="Prefill")
+        else:
+            marker_instance = rpd_user_marker(name="Decode")
+
+        marker_instance.start()
         # TODO(andoorve): We can remove this once all
         # virtual engines share the same kv cache.
         virtual_engine = model_input.virtual_engine
@@ -1765,6 +1771,7 @@ def execute_model(
 
             output.hidden_states = hidden_states
 
+        marker_instance.end()
         return [output]
 
     def need_recv_kv(self, model_input, kv_caches) -> bool: