fix: ensure max_thinking_tokens is respected in OpenAI text mode (#9604)

Signed-off-by: weizhoublue <weizhou.lan@daocloud.io>
Co-authored-by: Ryan McCormick <rmccormick@nvidia.com>
Co-authored-by: Dmitry Tokarev <dtokarev@nvidia.com>

Author: 3 people

components/src/dynamo/vllm/handlers.py

@@ -875,6 +875,12 @@ def build_sampling_params_openai(
     if "min_tokens" in request and request["min_tokens"] is not None:
         sampling_params.min_tokens = request["min_tokens"]
 
+    nvext_max_thinking_tokens = (request.get("nvext") or {}).get("max_thinking_tokens")
+    if nvext_max_thinking_tokens is not None and hasattr(
+        sampling_params, "thinking_token_budget"
+    ):
+        sampling_params.thinking_token_budget = nvext_max_thinking_tokens
+
     return sampling_params
 
 

components/src/dynamo/vllm/tests/test_vllm_unit.py

@@ -1215,3 +1215,16 @@ def test_rejects_multimodal_combo(self, mock_vllm_cli):
             ValueError, match="--embedding-worker cannot be combined with multimodal"
         ):
             parse_args()
+
+
+def test_build_sampling_params_openai_maps_max_thinking_tokens():
+    from dynamo.vllm.handlers import build_sampling_params_openai
+
+    request = {
+        "model": "test-model",
+        "prompt": "Solve: 1+1.",
+        "max_tokens": 32,
+        "nvext": {"max_thinking_tokens": 1024},
+    }
+    sp = build_sampling_params_openai(request, default_sampling_params={})
+    assert sp.thinking_token_budget == 1024