diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index 4e36909f45e..19c7c7cae0e 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -254,7 +254,7 @@ size_t ggml_backend_get_max_size(ggml_backend_t backend) {
 void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     GGML_ASSERT(backend);
     GGML_ASSERT(tensor);
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
     if (backend->iface.set_tensor_async == NULL) {
@@ -268,7 +268,7 @@ void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor *
 void ggml_backend_tensor_get_async(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size) {
     GGML_ASSERT(backend);
     GGML_ASSERT(tensor);
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
     if (backend->iface.get_tensor_async == NULL) {
@@ -283,7 +283,7 @@ void ggml_backend_tensor_set_2d_async(ggml_backend_t backend, struct ggml_tensor
             size_t n_copies, size_t stride_tensor, size_t stride_data) {
     GGML_ASSERT(backend);
     GGML_ASSERT(tensor);
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
 
     if (n_copies <= 1 || backend->iface.set_tensor_2d_async == NULL) {
         for (size_t i = 0; i < n_copies; i++) {
@@ -295,7 +295,7 @@ void ggml_backend_tensor_set_2d_async(ggml_backend_t backend, struct ggml_tensor
         return;
     }
 
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
     backend->iface.set_tensor_2d_async(backend, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
 }
@@ -304,7 +304,7 @@ void ggml_backend_tensor_get_2d_async(ggml_backend_t backend, const struct ggml_
             size_t n_copies, size_t stride_tensor, size_t stride_data) {
     GGML_ASSERT(backend);
     GGML_ASSERT(tensor);
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
 
     if (n_copies <= 1 || backend->iface.set_tensor_2d_async == NULL) {
         for (size_t i = 0; i < n_copies; i++) {
@@ -316,7 +316,7 @@ void ggml_backend_tensor_get_2d_async(ggml_backend_t backend, const struct ggml_
         return;
     }
 
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
     backend->iface.get_tensor_2d_async(backend, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
 }
@@ -330,7 +330,7 @@ void ggml_backend_tensor_set(struct ggml_tensor * tensor, const void * data, siz
         return;
     }
 
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
     buf->iface.set_tensor(buf, tensor, data, offset, size);
@@ -345,7 +345,7 @@ void ggml_backend_tensor_get(const struct ggml_tensor * tensor, void * data, siz
         return;
     }
 
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
     buf->iface.get_tensor(buf, tensor, data, offset, size);
@@ -367,7 +367,7 @@ void ggml_backend_tensor_set_2d(struct ggml_tensor * tensor, const void * data,
         return;
     }
 
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
 
     buf->iface.set_tensor_2d(buf, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
@@ -389,7 +389,7 @@ void ggml_backend_tensor_get_2d(const struct ggml_tensor * tensor, void * data,
         return;
     }
 
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + (n_copies-1)*stride_tensor + size <= ggml_nbytes(tensor) && "tensor read out of bounds");
 
     buf->iface.get_tensor_2d(buf, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
@@ -404,7 +404,7 @@ void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size
     }
 
     GGML_ASSERT(buf != NULL && "tensor buffer not set");
-    GGML_ASSERT(tensor->data != NULL && "tensor not allocated");
+    GGML_ASSERT((tensor->data != NULL || (tensor->view_src != NULL && tensor->view_src->data != NULL)) && "tensor not allocated");
     GGML_ASSERT(offset + size <= ggml_nbytes(tensor) && "tensor write out of bounds");
     GGML_ASSERT(buf->iface.memset_tensor != NULL && "memset not implemented by backend buffer");
 
diff --git a/ggml/src/ggml-cuda/fattn-common.cuh b/ggml/src/ggml-cuda/fattn-common.cuh
index d3e95f678a5..74ce6ae9e1c 100644
--- a/ggml/src/ggml-cuda/fattn-common.cuh
+++ b/ggml/src/ggml-cuda/fattn-common.cuh
@@ -1299,14 +1299,19 @@ void launch_fattn(
 
 #ifdef GGML_USE_HIP
     // HIP/ROCm: bypass the memory pool for f16 temp buffers.
-    // The legacy pool (ggml_cuda_pool_leg) retains peak-sized allocations permanently.
-    // For quantized KV dequant, this means the f16 temp buffer stays allocated,
-    // consuming more VRAM than the quantized KV compression saves — causing OOM.
-    // Using raw alloc+free ensures the memory is released after the kernel completes.
+    // The legacy pool (ggml_cuda_pool_leg) retains peak-sized allocations permanently
+    // because free() stores buffers for reuse rather than releasing them.
+    // On HIP without VMM support (RDNA 3/4), this means the f16 dequant temp buffers
+    // for quantized KV stay allocated after use, consuming more VRAM than the KV
+    // compression saves — causing OOM before f16 at equivalent context lengths.
+    // Using raw cudaMalloc/cudaFree ensures memory is released after the kernel completes.
+    // Ref: https://github.com/ggml-org/llama.cpp/issues/22107
     struct hip_f16_alloc {
         half * ptr = nullptr;
         cudaStream_t stream;
         hip_f16_alloc(cudaStream_t s) : stream(s) {}
+        hip_f16_alloc(const hip_f16_alloc &) = delete;
+        hip_f16_alloc & operator=(const hip_f16_alloc &) = delete;
         ~hip_f16_alloc() {
             if (ptr) {
                 // Cast to void: hipStreamSynchronize / hipFree are [[nodiscard]] under
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index a296d0ab446..831ff31ffa7 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -4600,6 +4600,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
         ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q5_0], "set_rows_q5_0" #itype, set_rows_q5_0 ## itype ## _len, set_rows_q5_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
         ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q5_1], "set_rows_q5_1" #itype, set_rows_q5_1 ## itype ## _len, set_rows_q5_1 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
         ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_Q8_0], "set_rows_q8_0" #itype, set_rows_q8_0 ## itype ## _len, set_rows_q8_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_TURBO3_0], "set_rows_turbo3_0" #itype, set_rows_turbo3_0 ## itype ## _len, set_rows_turbo3_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
+        ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_TURBO4_0], "set_rows_turbo4_0" #itype, set_rows_turbo4_0 ## itype ## _len, set_rows_turbo4_0 ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true); \
         ggml_vk_create_pipeline(device, device->pipeline_set_rows ## itype [GGML_TYPE_IQ4_NL], "set_rows_iq4_nl" #itype, set_rows_iq4_nl ## itype ## _len, set_rows_iq4_nl ## itype ## _data, "main", 3, sizeof(vk_op_binary_push_constants), {1, 1, 1}, {1}, 1, true);
 
     SET_ROWS(_i32)
@@ -15814,6 +15816,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_Q5_0:
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_TURBO3_0:
+                    case GGML_TYPE_TURBO4_0:
                     case GGML_TYPE_IQ4_NL:
                         return true;
                     default:
@@ -15838,6 +15842,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_Q5_0:
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_TURBO3_0:
+                    case GGML_TYPE_TURBO4_0:
                     case GGML_TYPE_IQ4_NL:
                         return true;
                     default:
@@ -15853,6 +15859,8 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
                     case GGML_TYPE_Q5_0:
                     case GGML_TYPE_Q5_1:
                     case GGML_TYPE_Q8_0:
+                    case GGML_TYPE_TURBO3_0:
+                    case GGML_TYPE_TURBO4_0:
                     case GGML_TYPE_IQ4_NL:
                         return true;
                     default:
diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
index d99b2b5d802..91ab4c00427 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
@@ -749,7 +749,7 @@ void process_shaders() {
         string_to_spv("cpy_" + t + "_f32", "copy_from_quant.comp", {{"DATA_A_" + to_uppercase(t), "1"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }
 
-    for (std::string t : {"f32", "f16", "bf16", "q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl"}) {
+    for (std::string t : {"f32", "f16", "bf16", "q1_0", "q4_0", "q4_1", "q5_0", "q5_1", "q8_0", "iq4_nl", "turbo3_0", "turbo4_0"}) {
         string_to_spv("set_rows_" + t + "_i32", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uint"}, {"B_SIZE", "32"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
         string_to_spv("set_rows_" + t + "_i64", "copy_to_quant.comp", {{"SET_ROWS", "1"}, {"DATA_A_" + to_uppercase(t), "1"}, {"B_TYPE", "uvec2"}, {"B_SIZE", "64"}, {"D_TYPE", "float"}, {"FLOAT_TYPE", "float"}});
     }