⚡ Bolt: Optimize RequestMetrics & SpeculateMetrics Serialization

.jules/bolt.md

@@ -0,0 +1,4 @@
+
+## 2025-02-23 - Avoid dataclasses.asdict in Hot Paths
+**Learning:** `dataclasses.asdict` does recursive deepcopy internally and is incredibly slow for large dataclasses or objects instantiated frequently. In FastDeploy, it was used in `RequestMetrics.to_dict()`, creating significant overhead.
+**Action:** When defining `to_dict()` or custom serialization methods for fast/frequent dataclasses, avoid `asdict`. Instead, iterate through `self.__dataclass_fields__` with `getattr` and do shallow copying for basic types (`int`, `float`, `str`, `bool`, `type(None)`). For nested dataclasses, ensure they also implement their own `to_dict()` method to skip the `asdict` recursive penalty.

fastdeploy/engine/request.py

@@ -897,7 +897,26 @@ def to_dict(self):
         """
         Convert the RequestMetrics object to a dictionary.
         """
-        return {k: v for k, v in asdict(self).items()}
+        import dataclasses
+
+        res = {}
+        for k in self.__dataclass_fields__:
+            v = getattr(self, k)
+            if type(v) in (int, float, str, bool, type(None)):
+                res[k] = v
+            elif isinstance(v, list):
+                res[k] = list(v)
+            elif isinstance(v, dict):
+                res[k] = dict(v)
+            else:
+                if dataclasses.is_dataclass(v):
+                    if hasattr(v, "to_dict"):
+                        res[k] = v.to_dict()
+                    else:
+                        res[k] = dataclasses.asdict(v)
+                else:
+                    res[k] = v
+        return res
 
     def record_recv_first_token(self):
         cur_time = time.time()

fastdeploy/worker/output.py

@@ -164,6 +164,20 @@ class SpeculateMetrics:
     """
     accept_ratio_per_head: list[float]
 
+    def to_dict(self):
+        return {
+            "accepted_tokens": self.accepted_tokens,
+            "rejected_tokens": self.rejected_tokens,
+            "accept_ratio": self.accept_ratio,
+            "average_accept_length": self.average_accept_length,
+            "accepted_tokens_per_head": (
+                list(self.accepted_tokens_per_head) if self.accepted_tokens_per_head is not None else None
+            ),
+            "accept_ratio_per_head": (
+                list(self.accept_ratio_per_head) if self.accept_ratio_per_head is not None else None
+            ),
+        }
+
 
 @dataclass
 class SamplerOutput: