Accept validated thinking template options in chat compatibility layer

api_compat.py

@@ -5,7 +5,7 @@ positive levels to the model template; it handles 'none' as thinking disabled.
 """
 import math
 
-CHAT_FIELDS=frozenset({'model','messages','stream','stream_options','temperature','top_p','top_k','max_tokens','max_completion_tokens','stop','seed','tools','tool_choice','parallel_tool_calls','response_format','repeat_penalty','presence_penalty','frequency_penalty','logprobs','top_logprobs','user','n','reasoning_effort'})
+CHAT_FIELDS=frozenset({'model','messages','stream','stream_options','temperature','top_p','top_k','max_tokens','max_completion_tokens','stop','seed','tools','tool_choice','parallel_tool_calls','response_format','repeat_penalty','presence_penalty','frequency_penalty','logprobs','top_logprobs','user','n','reasoning_effort','chat_template_kwargs'})
 LLAMA_EFFORTS=frozenset({'none','low','medium','high','xhigh'})
 # The installed llama.cpp server rejects minimal and max. Use nearest supported hints.
 EFFORT_ALIASES={'minimal':'low','max':'xhigh','ultra':'xhigh'}
@@ -17,6 +17,13 @@ def normalize_chat(data):
  unknown=set(data)-CHAT_FIELDS
  if unknown:raise CompatibilityError('Nicht unterstützte Chat-Felder: '+', '.join(sorted(unknown)))
  body=dict(data)
+ if 'chat_template_kwargs' in body:
+  kwargs=body['chat_template_kwargs']
+  if not isinstance(kwargs,dict) or set(kwargs)-{'enable_thinking','preserve_thinking'}:
+   raise CompatibilityError('chat_template_kwargs unterstützt enable_thinking und preserve_thinking.')
+  if any(type(v) is not bool for v in kwargs.values()):
+   raise CompatibilityError('Thinking-Template-Optionen müssen true oder false sein.')
+  body['chat_template_kwargs']=dict(kwargs)
  for key,lo,hi in [('repeat_penalty',0,2),('presence_penalty',-2,2),('frequency_penalty',-2,2)]:
   if key in body:
    value=body[key]