diff --git a/docs/build.md b/docs/build.md
index dcbcce7549a..c277aac82fd 100644
--- a/docs/build.md
+++ b/docs/build.md
@@ -413,6 +413,27 @@ Finally, after finishing your build, you should be able to do something like thi
 # ggml_vulkan: Using Intel(R) Graphics (ADL GT2) | uma: 1 | fp16: 1 | warp size: 32
 ```
 
+#### Targeting AMD GCN (GFX8/Polaris) Vulkan devices
+
+Older AMD GPUs such as the RX 580 expose fast FP32 wave64 execution but lack
+cooperative-matrix, integer-dot and bfloat16 shader support. You can avoid
+building unusable shader families by forcing their GLSL compiler probes off
+when configuring CMake:
+
+```bash
+cmake -B build-gfx803 -DGGML_VULKAN=ON \
+      -DGGML_VULKAN_FORCE_COOPMAT_GLSLC_SUPPORT=OFF \
+      -DGGML_VULKAN_FORCE_COOPMAT2_GLSLC_SUPPORT=OFF \
+      -DGGML_VULKAN_FORCE_INTEGER_DOT_GLSLC_SUPPORT=OFF \
+      -DGGML_VULKAN_FORCE_BFLOAT16_GLSLC_SUPPORT=OFF
+cmake --build build-gfx803 --config Release
+```
+
+At runtime, keep execution on the FP32-oriented code paths by exporting
+`GGML_VK_DISABLE_F16=1`. Polaris-class GPUs also benefit from the existing
+`GGML_VK_FORCE_MMVQ=1` flag when profiling shows that matrix–matrix–vector
+quantization remains faster than the fallback kernels.
+
 ## CANN
 This provides NPU acceleration using the AI cores of your Ascend NPU. And [CANN](https://www.hiascend.com/en/software/cann) is a hierarchical APIs to help you to quickly build AI applications and service based on Ascend NPU.
 
diff --git a/ggml/src/ggml-vulkan/CMakeLists.txt b/ggml/src/ggml-vulkan/CMakeLists.txt
index b97e7bf9955..c93d51d51e2 100644
--- a/ggml/src/ggml-vulkan/CMakeLists.txt
+++ b/ggml/src/ggml-vulkan/CMakeLists.txt
@@ -3,6 +3,40 @@ cmake_policy(SET CMP0114 NEW)
 
 find_package(Vulkan COMPONENTS glslc REQUIRED)
 
+set(GGML_VULKAN_FORCE_COOPMAT_GLSLC_SUPPORT "" CACHE STRING "Force GL_KHR_cooperative_matrix GLSL compiler support detection (use ON, OFF, or leave empty for auto)")
+set(GGML_VULKAN_FORCE_COOPMAT2_GLSLC_SUPPORT "" CACHE STRING "Force GL_NV_cooperative_matrix2 GLSL compiler support detection (use ON, OFF, or leave empty for auto)")
+set(GGML_VULKAN_FORCE_INTEGER_DOT_GLSLC_SUPPORT "" CACHE STRING "Force GL_EXT_integer_dot_product GLSL compiler support detection (use ON, OFF, or leave empty for auto)")
+set(GGML_VULKAN_FORCE_BFLOAT16_GLSLC_SUPPORT "" CACHE STRING "Force GL_EXT_bfloat16 GLSL compiler support detection (use ON, OFF, or leave empty for auto)")
+
+mark_as_advanced(
+    GGML_VULKAN_FORCE_COOPMAT_GLSLC_SUPPORT
+    GGML_VULKAN_FORCE_COOPMAT2_GLSLC_SUPPORT
+    GGML_VULKAN_FORCE_INTEGER_DOT_GLSLC_SUPPORT
+    GGML_VULKAN_FORCE_BFLOAT16_GLSLC_SUPPORT
+)
+
+function(handle_forced_shader_extension EXTENSION_NAME RESULT_VARIABLE FORCE_VARIABLE)
+    if (NOT "${${FORCE_VARIABLE}}" STREQUAL "")
+        string(TOUPPER "${${FORCE_VARIABLE}}" FORCE_VALUE_UPPER)
+        if (FORCE_VALUE_UPPER STREQUAL "ON")
+            message(STATUS "${EXTENSION_NAME} support forced ON")
+            set(${RESULT_VARIABLE} ON)
+            add_compile_definitions(${RESULT_VARIABLE})
+            list(APPEND VULKAN_SHADER_GEN_CMAKE_ARGS -D${RESULT_VARIABLE}=ON)
+            set(VULKAN_SHADER_GEN_CMAKE_ARGS "${VULKAN_SHADER_GEN_CMAKE_ARGS}" PARENT_SCOPE)
+            set(${RESULT_VARIABLE} ON PARENT_SCOPE)
+            return()
+        elseif (FORCE_VALUE_UPPER STREQUAL "OFF")
+            message(STATUS "${EXTENSION_NAME} support forced OFF")
+            set(${RESULT_VARIABLE} OFF)
+            set(${RESULT_VARIABLE} OFF PARENT_SCOPE)
+            return()
+        else()
+            message(FATAL_ERROR "Invalid value for ${FORCE_VARIABLE}: ${${FORCE_VARIABLE}}. Expected ON, OFF, or empty.")
+        endif()
+    endif()
+endfunction()
+
 function(detect_host_compiler)
     if (CMAKE_HOST_SYSTEM_NAME STREQUAL "Windows")
         find_program(HOST_C_COMPILER NAMES cl gcc clang NO_CMAKE_FIND_ROOT_PATH)
@@ -52,29 +86,61 @@ if (Vulkan_FOUND)
     set(VULKAN_SHADER_GEN_CMAKE_ARGS "")
 
     # Test all shader extensions
-    test_shader_extension_support(
-        "GL_KHR_cooperative_matrix"
-        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp"
-        "GGML_VULKAN_COOPMAT_GLSLC_SUPPORT"
-    )
+    if (NOT "${GGML_VULKAN_FORCE_COOPMAT_GLSLC_SUPPORT}" MATCHES "^(|[Aa][Uu][Tt][Oo])$")
+        handle_forced_shader_extension(
+            "GL_KHR_cooperative_matrix"
+            "GGML_VULKAN_COOPMAT_GLSLC_SUPPORT"
+            "GGML_VULKAN_FORCE_COOPMAT_GLSLC_SUPPORT"
+        )
+    else()
+        test_shader_extension_support(
+            "GL_KHR_cooperative_matrix"
+            "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat_support.comp"
+            "GGML_VULKAN_COOPMAT_GLSLC_SUPPORT"
+        )
+    endif()
 
-    test_shader_extension_support(
-        "GL_NV_cooperative_matrix2"
-        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
-        "GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT"
-    )
+    if (NOT "${GGML_VULKAN_FORCE_COOPMAT2_GLSLC_SUPPORT}" MATCHES "^(|[Aa][Uu][Tt][Oo])$")
+        handle_forced_shader_extension(
+            "GL_NV_cooperative_matrix2"
+            "GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT"
+            "GGML_VULKAN_FORCE_COOPMAT2_GLSLC_SUPPORT"
+        )
+    else()
+        test_shader_extension_support(
+            "GL_NV_cooperative_matrix2"
+            "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_coopmat2_support.comp"
+            "GGML_VULKAN_COOPMAT2_GLSLC_SUPPORT"
+        )
+    endif()
 
-    test_shader_extension_support(
-        "GL_EXT_integer_dot_product"
-        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_integer_dot_support.comp"
-        "GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT"
-    )
+    if (NOT "${GGML_VULKAN_FORCE_INTEGER_DOT_GLSLC_SUPPORT}" MATCHES "^(|[Aa][Uu][Tt][Oo])$")
+        handle_forced_shader_extension(
+            "GL_EXT_integer_dot_product"
+            "GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT"
+            "GGML_VULKAN_FORCE_INTEGER_DOT_GLSLC_SUPPORT"
+        )
+    else()
+        test_shader_extension_support(
+            "GL_EXT_integer_dot_product"
+            "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_integer_dot_support.comp"
+            "GGML_VULKAN_INTEGER_DOT_GLSLC_SUPPORT"
+        )
+    endif()
 
-    test_shader_extension_support(
-        "GL_EXT_bfloat16"
-        "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp"
-        "GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT"
-    )
+    if (NOT "${GGML_VULKAN_FORCE_BFLOAT16_GLSLC_SUPPORT}" MATCHES "^(|[Aa][Uu][Tt][Oo])$")
+        handle_forced_shader_extension(
+            "GL_EXT_bfloat16"
+            "GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT"
+            "GGML_VULKAN_FORCE_BFLOAT16_GLSLC_SUPPORT"
+        )
+    else()
+        test_shader_extension_support(
+            "GL_EXT_bfloat16"
+            "${CMAKE_CURRENT_SOURCE_DIR}/vulkan-shaders/test_bfloat16_support.comp"
+            "GGML_VULKAN_BFLOAT16_GLSLC_SUPPORT"
+        )
+    endif()
 
     target_link_libraries(ggml-vulkan PRIVATE Vulkan::Vulkan)
     target_include_directories(ggml-vulkan PRIVATE ${CMAKE_CURRENT_BINARY_DIR})
diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
index ebbb412e55f..e3f6c2a78cb 100644
--- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp
+++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp
@@ -257,30 +257,58 @@ static vk_device_architecture get_device_architecture(const vk::PhysicalDevice&
             }
         }
 
-        if (!amd_shader_core_properties || !integer_dot_product || !subgroup_size_control) {
+        if (!amd_shader_core_properties) {
             return vk_device_architecture::OTHER;
         }
 
         vk::PhysicalDeviceProperties2 props2;
         vk::PhysicalDeviceShaderCorePropertiesAMD shader_core_props_amd;
-        vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props;
-        vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+        vk::PhysicalDeviceSubgroupProperties subgroup_props;
 
+        shader_core_props_amd.pNext = &subgroup_props;
+        subgroup_props.pNext = nullptr;
         props2.pNext = &shader_core_props_amd;
-        shader_core_props_amd.pNext = &integer_dot_props;
-        integer_dot_props.pNext = &subgroup_size_control_props;
 
         device.getProperties2(&props2);
 
-        if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 64) {
+        uint32_t max_subgroup_size = subgroup_props.subgroupSize;
+
+        if (subgroup_size_control) {
+            vk::PhysicalDeviceProperties2 subgroup_size_props;
+            vk::PhysicalDeviceSubgroupSizeControlPropertiesEXT subgroup_size_control_props;
+
+            subgroup_size_control_props.pNext = nullptr;
+            subgroup_size_props.pNext = &subgroup_size_control_props;
+
+            device.getProperties2(&subgroup_size_props);
+
+            max_subgroup_size = subgroup_size_control_props.maxSubgroupSize;
+        }
+
+        bool integer_dot_mixed_signedness = false;
+        if (integer_dot_product) {
+            vk::PhysicalDeviceProperties2 integer_dot_props2;
+            vk::PhysicalDeviceShaderIntegerDotProductPropertiesKHR integer_dot_props;
+
+            integer_dot_props.pNext = nullptr;
+            integer_dot_props2.pNext = &integer_dot_props;
+
+            device.getProperties2(&integer_dot_props2);
+
+            integer_dot_mixed_signedness = integer_dot_props.integerDotProduct4x8BitPackedMixedSignednessAccelerated;
+        }
+
+        const uint32_t wavefront_size = shader_core_props_amd.wavefrontSize;
+
+        if (wavefront_size == 64 && max_subgroup_size >= 64) {
             return vk_device_architecture::AMD_GCN;
         }
-        if (subgroup_size_control_props.maxSubgroupSize == 64 && subgroup_size_control_props.minSubgroupSize == 32) {
-            // RDNA
+
+        if (wavefront_size == 32) {
             if (shader_core_props_amd.wavefrontsPerSimd == 20) {
                 return vk_device_architecture::AMD_RDNA1;
             }
-            if (integer_dot_props.integerDotProduct4x8BitPackedMixedSignednessAccelerated) {
+            if (integer_dot_mixed_signedness) {
                 return vk_device_architecture::AMD_RDNA3;
             }
             return vk_device_architecture::AMD_RDNA2;