fix: clamp max_tokens and inference parameters to valid Gradio slider ranges

qflash_test.go

@@ -626,4 +626,27 @@ func TestQwenServiceAutoFailover(t *testing.T) {
   }
 }
 
+func TestResolveMaxTokens(t *testing.T) {
+  // Default when empty
+  if v := ResolveMaxTokens(ChatCompletionRequest{}); v != 4096 {
+    t.Fatalf("expected default 4096, got %d", v)
+  }
+
+  // Explicit max_tokens
+  if v := ResolveMaxTokens(ChatCompletionRequest{MaxTokens: 2048}); v != 2048 {
+    t.Fatalf("expected 2048, got %d", v)
+  }
+
+  // max_completion_tokens precedence when max_tokens is 0
+  if v := ResolveMaxTokens(ChatCompletionRequest{MaxCompletionTokens: 1024}); v != 1024 {
+    t.Fatalf("expected 1024, got %d", v)
+  }
+
+  // Cap at 32768
+  if v := ResolveMaxTokens(ChatCompletionRequest{MaxTokens: 65536}); v != 32768 {
+    t.Fatalf("expected cap at 32768, got %d", v)
+  }
+}
+
+