From ac84e40f884e4df24fd39a6037a5fcddbdab22c5 Mon Sep 17 00:00:00 2001
From: project6-dev <dev@project6.ai>
Date: Wed, 12 Aug 2026 06:16:44 +0000
Subject: [PATCH] fix: accept max_completion_tokens + reasoning_effort + n=2
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Sub655 replay log shows 881 requests, majority failing with:
  'Extra inputs are not permitted', 'input': 8192  (max_completion_tokens)
  'Extra inputs are not permitted', 'input': 'low' (reasoning_effort)

Root cause: protocol.py OpenAIBaseModel has extra='forbid',
and ChatCompletionRequest lacks these two OpenAI API fields.

Sub168 has the SAME bug — both reject max_completion_tokens.

Fixes:
1. protocol.py: add max_completion_tokens (Optional[int]) field
2. protocol.py: add reasoning_effort (Optional[str]) field
3. serving_chat.py: merge max_completion_tokens into max_tokens
4. computility-run.yaml: max_num_seqs=2 (fixes t2_n_2 HTTP 400)
---
 computility-run.yaml            | 2 +-
 qwen3_6_scripts/protocol.py     | 3 +++
 qwen3_6_scripts/serving_chat.py | 3 +++
 3 files changed, 7 insertions(+), 1 deletion(-)

diff --git a/computility-run.yaml b/computility-run.yaml
index 41351f46..ac5b8ec2 100644
--- a/computility-run.yaml
+++ b/computility-run.yaml
@@ -15,7 +15,7 @@ command:
     - -tp
     - '4'
     - --max-num-seqs
-    - '1'
+    - '2'
     - --disable-log-requests
     - --disable-frontend-multiprocessing
     - --max-num-batched-tokens
diff --git a/qwen3_6_scripts/protocol.py b/qwen3_6_scripts/protocol.py
index a51a1c57..3a041b44 100644
--- a/qwen3_6_scripts/protocol.py
+++ b/qwen3_6_scripts/protocol.py
@@ -180,6 +180,8 @@ class ChatCompletionRequest(OpenAIBaseModel):
     logprobs: Optional[bool] = False
     top_logprobs: Optional[int] = 0
     max_tokens: Optional[int] = None
+    # OpenAI newer API field — treat as alias for max_tokens
+    max_completion_tokens: Optional[int] = None
     n: Optional[int] = 1
     presence_penalty: Optional[float] = 0.0
     response_format: Optional[ResponseFormat] = None
@@ -193,6 +195,7 @@ class ChatCompletionRequest(OpenAIBaseModel):
     tool_choice: Optional[Union[Literal["none"], Literal["auto"],
                                 ChatCompletionNamedToolChoiceParam]] = "none"
     thinking: Optional[Union[bool, str, Dict[str, Any]]] = None
+    reasoning_effort: Optional[str] = None
 
     # NOTE this will be ignored by VLLM -- the model determines the behavior
     parallel_tool_calls: Optional[bool] = False
diff --git a/qwen3_6_scripts/serving_chat.py b/qwen3_6_scripts/serving_chat.py
index 4ed841b5..366253b1 100644
--- a/qwen3_6_scripts/serving_chat.py
+++ b/qwen3_6_scripts/serving_chat.py
@@ -394,6 +394,9 @@ class OpenAIServingChat(OpenAIServing):
             assert prompt_inputs is not None
 
             sampling_params: Union[SamplingParams, BeamSearchParams]
+            # OpenAI API: max_completion_tokens takes precedence over max_tokens
+            if request.max_completion_tokens is not None and request.max_tokens is None:
+                request.max_tokens = request.max_completion_tokens
             default_max_tokens = self.max_model_len - len(
                 prompt_inputs["prompt_token_ids"])
             if request.use_beam_search: