llama: enable dio on linux CUDA/ROCm iGPUs (#17286)

Avoid double memory consumption by enabling direct IO for iGPUs

llm/llama_server.go

@@ -390,6 +390,15 @@ func startLlamaServer(launch llamaServerLaunchConfig, out io.Writer) (cmd *exec.
 		params = append(params, "--no-mmap")
 	}
 
+	// Direct I/O skips the page cache on load for integrated CUDA/ROCm GPUs, which
+	// share system memory with the CPU and would otherwise double-buffer weights.
+	for _, g := range launch.gpus {
+		if runtime.GOOS == "linux" && g.Integrated && (strings.EqualFold(g.Library, "CUDA") || strings.EqualFold(g.Library, "ROCm")) {
+			params = append(params, "--direct-io")
+			break
+		}
+	}
+
 	// KV cache type
 	if launch.kvCacheType != "" {
 		params = append(params, "--cache-type-k", launch.kvCacheType, "--cache-type-v", launch.kvCacheType)