diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index 22f2dd33bf..81d4e2eafc 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -2529,6 +2529,9 @@ static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
                                                        vk::MemoryPropertyFlagBits::eDeviceLocal});
         } else if (device->uma) {
             // Fall back to host memory type
+            // 1. Prefer Device Local AND Host Visible (Best for UMA)
+            // 2. Fallback to Device Local
+            // 3. Fallback to Host Visible
             buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
                                                        vk::MemoryPropertyFlagBits::eDeviceLocal,
                                                        vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
@@ -2542,6 +2545,9 @@ static vk_buffer ggml_vk_create_buffer_device(vk_device& device, size_t size) {
         } else {
             // use rebar if available, otherwise fallback to device only visible memory
             if (device->allow_sysmem_fallback) {
+                // 1. Prefer Device Local AND Host Visible (ReBAR)
+                // 2. Fallback to Device Local
+                // 3. Fallback to Host Visible
                 buf = ggml_vk_create_buffer(device, size, {vk::MemoryPropertyFlagBits::eDeviceLocal | vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent,
                                                            vk::MemoryPropertyFlagBits::eDeviceLocal,
                                                            vk::MemoryPropertyFlagBits::eHostVisible | vk::MemoryPropertyFlagBits::eHostCoherent});
@@ -2936,6 +2942,10 @@ static void ggml_vk_load_shaders(vk_device& device) {
             m_warptile_mmq = m_warptile_mmq_int = { 64, 64, 64, 16, 16, 32, 2, 2, 2, 1, 16 };
             m_warptile = { 64, 64, 64, 16, 16, 32, 2, 2, 2, 1, 16 };
             m_warptile_id = m_warptile_mmqid = { 64, 64, 64, 16, 16, 32, 2, 2, 2, 1, 16 };
+
+            l_warptile_mmq = l_warptile_mmq_int = m_warptile_mmq;
+            l_warptile = m_warptile;
+            l_warptile_id = l_warptile_mmqid = m_warptile_id;
         } else if ((device->architecture == AMD_GCN) && (device->driver_id != vk::DriverId::eAmdProprietary)) {
             m_warptile_mmq = m_warptile_mmq_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
             m_warptile_mmqid = m_warptile_mmqid_int = { 256, 64, 64, 32, 16, 16, 2, 2, 2, 1, 16 };
@@ -4494,11 +4504,8 @@ static vk_device ggml_vk_get_device(size_t idx) {
         device->driver_id = driver_props.driverID;
 
         if (device->vendor_id == VK_VENDOR_ID_ARM) {
-            // Forcing FP32 path as it currently provides better performance on Mali G720.
-            // This is a simplified approach while deeper ARM-specific vec2 SIMD optimizations are investigated.
-            fp16_storage = false;
-            fp16_compute = false;
-            bfloat16_support = false;
+            // Previously forced FP32 here due to poor FP16 performance on some ARM GPUs.
+            // With adjusted l_warptile (below), FP16 is now performant and preferred.
         }
 
         // Implementing the async backend interfaces seems broken on older Intel HW,