[BugFix] adjust max_tokens and min_tokens when continue to generate tokens (#5010)

* fix max and min tokens initial commit * fix double subtraction * add unit tests
2026-04-23 00:17:25 +08:00 · 2025-11-13 23:52:54 +08:00
parent 6c3d1da62f
commit 9703108c28
3 changed files with 158 additions and 3 deletions
@@ -206,7 +206,15 @@ class EngineClient:

            task["prompt_token_ids_len"] = len(task["prompt_token_ids"])
            input_ids_len = task["prompt_token_ids_len"]
-            task["max_tokens"] = min(self.max_model_len - input_ids_len, task.get("max_tokens"))
+
+            completion_token_len = len(task["completion_token_ids"]) if task.get("completion_token_ids") else 0
+            task["max_tokens"] = min(
+                self.max_model_len - input_ids_len, max(0, task.get("max_tokens") - completion_token_len)
+            )
+
+            if task.get("min_tokens") is not None:
+                task["min_tokens"] = max(1, task["min_tokens"] - completion_token_len)
+
            min_tokens = task.get("min_tokens", 1)
            if "messages" in task:
                del task["messages"]