Store preferences by model profile

src/server/request.rs

@@ -499,8 +499,8 @@ pub(super) fn parse_chat_request(
         return Err((400, format!("model is not installed and verified: {model}")));
     }
 
-    let mut generation = config.generation.clone();
-    generation.system_prompt.clear();
+    let reasoning_mode = request_reasoning(&request, requested_id)?;
+    let mut generation = config.generation_for(model, reasoning_mode);
     generation.max_generated_tokens = request
         .max_completion_tokens
         .or(request.max_tokens)
@@ -512,7 +512,6 @@ pub(super) fn parse_chat_request(
     generation.top_p = request.top_p.or(Some(1.0));
     generation.min_p = request.min_p.or(Some(0.05));
     generation.seed = request.seed.filter(|seed| *seed > 0);
-    generation.reasoning_mode = request_reasoning(&request, requested_id)?;
 
     let tools_enabled = (!request.tools.is_empty() || !request.tool_schemas.is_empty())
         && request.tool_choice.as_ref().and_then(Value::as_str) != Some("none");
@@ -525,7 +524,8 @@ pub(super) fn parse_chat_request(
         protocol,
     )?;
     generation.system_prompt = system;
-    let mut effective = effective_settings(model, &generation, &config.runtime, &state.models_path)
+    let runtime = config.runtime_for(model);
+    let mut effective = effective_settings(model, &generation, &runtime, &state.models_path)
         .map_err(|error| (400, error))?;
     effective.turn.top_k = request.top_k.unwrap_or(0);
     if effective.turn.top_k < 0 {