Make reasoning levels enforce real token budgets

dev/test_router_support.py

@@ -140,26 +140,27 @@ class LlamaCppReasoningTests(unittest.TestCase):
                     normalized["chat_template_kwargs"],
                     {"enable_thinking": False},
                 )
+                self.assertEqual(normalized["thinking_budget_tokens"], 0)
 
-    def test_low_and_medium_reach_chat_template(self) -> None:
-        for effort in ("low", "medium"):
+    def test_reasoning_levels_receive_real_per_request_budgets(self) -> None:
+        expected = {
+            "minimal": ("low", 256),
+            "low": ("low", 768),
+            "medium": ("medium", 2048),
+            "high": ("xhigh", 4096),
+            "xhigh": ("xhigh", 8192),
+            "max": ("xhigh", 8192),
+            "ultra": ("xhigh", 8192),
+        }
+        for effort, (template_effort, budget) in expected.items():
             with self.subTest(effort=effort):
                 request = {"reasoning_effort": effort, "messages": []}
                 normalized = _normalize_llamacpp_reasoning(request)
                 self.assertEqual(normalized["chat_template_kwargs"], {
                     "enable_thinking": True,
-                    "reasoning_effort": effort,
+                    "reasoning_effort": template_effort,
                 })
-
-    def test_unsupported_high_levels_are_clamped_to_xhigh(self) -> None:
-        for effort in ("high", "xhigh", "max", "ultra"):
-            with self.subTest(effort=effort):
-                request = {"reasoning_effort": effort, "messages": []}
-                normalized = _normalize_llamacpp_reasoning(request)
-                self.assertEqual(
-                    normalized["chat_template_kwargs"]["reasoning_effort"],
-                    "xhigh",
-                )
+                self.assertEqual(normalized["thinking_budget_tokens"], budget)
 
     def test_existing_template_kwargs_are_preserved(self) -> None:
         request = {
@@ -173,6 +174,7 @@ class LlamaCppReasoningTests(unittest.TestCase):
             "enable_thinking": True,
             "reasoning_effort": "low",
         })
+        self.assertEqual(normalized["thinking_budget_tokens"], 768)
 
     def test_request_without_effort_uses_safe_off_default(self) -> None:
         request = {"messages": []}
@@ -181,6 +183,13 @@ class LlamaCppReasoningTests(unittest.TestCase):
             request["chat_template_kwargs"],
             {"enable_thinking": False},
         )
+        self.assertEqual(request["thinking_budget_tokens"], 0)
+
+    def test_native_thinking_budget_is_preserved_without_openai_effort(self) -> None:
+        request = {"thinking_budget_tokens": 1234, "messages": []}
+        self.assertIs(_normalize_llamacpp_reasoning(request), request)
+        self.assertEqual(request["thinking_budget_tokens"], 1234)
+        self.assertNotIn("chat_template_kwargs", request)
 
 
 class ChatGenerationLimitTests(unittest.TestCase):