perf: slim turbo4b runtime profile

crates/inference-core/src/profile.rs

@@ -15,9 +15,9 @@ impl RuntimeProfile {
         match self {
             Self::Turbo4B => ProfileDefaults {
                 n_gpu_layers: 999,
-                n_ctx: 4096,
+                n_ctx: 2048,
-                n_batch: 512,
+                n_batch: 256,
-                n_ubatch: 128,
+                n_ubatch: 64,
                 flash_attn: true,
                 n_threads: 8,
             },