AIX: server-side Top-P

enricoros · Dec 25, 2024 · f52e350 · f52e350
1 parent b2d66af
commit f52e350
Show file tree

Hide file tree

Showing 3 changed files with 18 additions and 0 deletions.
diff --git a/src/modules/aix/server/dispatch/chatGenerate/adapters/anthropic.messageCreate.ts b/src/modules/aix/server/dispatch/chatGenerate/adapters/anthropic.messageCreate.ts
@@ -88,6 +88,12 @@ export function aixToAnthropicMessageCreate(model: AixAPI_Model, chatGenerate: A
     // top_p: undefined,
   };
 
+  // Top-P instead of temperature
+  if (model.topP !== undefined) {
+    payload.top_p = model.topP;
+    delete payload.temperature
+  }
+
   // Preemptive error detection with server-side payload validation before sending it upstream
   const validated = AnthropicWire_API_Message_Create.Request_schema.safeParse(payload);
   if (!validated.success) {

diff --git a/src/modules/aix/server/dispatch/chatGenerate/adapters/gemini.generateContent.ts b/src/modules/aix/server/dispatch/chatGenerate/adapters/gemini.generateContent.ts
@@ -51,6 +51,12 @@ export function aixToGeminiGenerateContent(model: AixAPI_Model, chatGenerate: Ai
     },
   };
 
+  // Top-P instead of temperature
+  if (model.topP !== undefined) {
+    delete payload.generationConfig!.temperature;
+    payload.generationConfig!.topP = model.topP;
+  }
+
   // Preemptive error detection with server-side payload validation before sending it upstream
   const validated = GeminiWire_API_Generate_Content.Request_schema.safeParse(payload);
   if (!validated.success) {

diff --git a/src/modules/aix/server/dispatch/chatGenerate/adapters/openai.chatCompletions.ts b/src/modules/aix/server/dispatch/chatGenerate/adapters/openai.chatCompletions.ts
@@ -79,6 +79,12 @@ export function aixToOpenAIChatCompletions(openAIDialect: OpenAIDialects, model:
     user: undefined,
   };
 
+  // Top-P instead of temperature
+  if (model.topP !== undefined) {
+    delete payload.temperature;
+    payload.top_p = model.topP;
+  }
+
   if (hotFixOpenAIo1Family)
     payload = _fixRequestForOpenAIO1_maxCompletionTokens(payload);