perf: pad small attention heads to 64 for MMA Flash Attention (#1992)

CMakeLists.txt

@@ -343,6 +343,11 @@ add_subdirectory(thirdparty)
 target_sources(${SD_LIB} PRIVATE $<TARGET_OBJECTS:zip>)
 target_link_libraries(${SD_LIB} PUBLIC ggml)
 target_link_libraries(${SD_LIB} PRIVATE onig sd-utf8proc)
+if (SD_CUDA)
+    find_package(CUDAToolkit REQUIRED)
+    target_link_libraries(${SD_LIB} PRIVATE CUDA::cuda_driver)
+    set_property(SOURCE src/core/ggml_extend_backend.cpp APPEND PROPERTY COMPILE_DEFINITIONS SD_USE_CUDA)
+endif()
 target_include_directories(${SD_LIB} PUBLIC . src include)
 target_include_directories(${SD_LIB} PRIVATE src/core)
 target_include_directories(${SD_LIB} PUBLIC . thirdparty)

cmake/stable-diffusion-config.cmake.in

@@ -10,6 +10,9 @@ set(SD_BIN_DIR "@PACKAGE_SD_BIN_INSTALL_DIR@")
 
 include(CMakeFindDependencyMacro)
 find_dependency(ggml REQUIRED HINTS "${SD_LIB_DIR}/cmake")
+if(@SD_CUDA@ AND NOT SD_SHARED_LIB)
+    find_dependency(CUDAToolkit REQUIRED)
+endif()
 
 if(NOT TARGET stable-diffusion)
     find_library(stable-diffusion_LIBRARY stable-diffusion
@@ -28,6 +31,10 @@ if(NOT TARGET stable-diffusion)
             INTERFACE_COMPILE_FEATURES "c_std_11;cxx_std_17"
             POSITION_INDEPENDENT_CODE ON)
 
+    if(@SD_CUDA@ AND NOT SD_SHARED_LIB)
+        set_property(TARGET stable-diffusion APPEND PROPERTY INTERFACE_LINK_LIBRARIES CUDA::cuda_driver)
+    endif()
+
     if(SD_SHARED_LIB)
         target_compile_definitions(stable-diffusion
             INTERFACE SD_BUILD_SHARED_LIB)

src/core/ggml_extend.cpp

@@ -643,6 +643,14 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
     ggml_tensor* kqv = nullptr;
 
     auto build_kqv = [&](ggml_tensor* q_in, ggml_tensor* k_in, ggml_tensor* v_in, ggml_tensor* mask_in) -> ggml_tensor* {
+        const bool pad_head = d_head > 0 && d_head < 64 && q_in->ne[0] == d_head && k_in->ne[0] == d_head &&
+                              q_in->type == GGML_TYPE_F32 && k_in->type == GGML_TYPE_F32 &&
+                              v_in->type == GGML_TYPE_F32 && sd_backend_supports_cuda_mma(backend);
+        if (pad_head) {
+            // CUDA FA MMA starts at 64 channels; keep the original head's attention scale.
+            q_in = ggml_pad(ctx, q_in, 64 - d_head, 0, 0, 0);
+            k_in = ggml_pad(ctx, k_in, 64 - d_head, 0, 0, 0);
+        }
         if (kv_scale != 1.0f) {
             k_in = ggml_ext_scale(ctx, k_in, kv_scale);
         }
@@ -650,6 +658,9 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
 
         v_in = ggml_ext_cont(ctx, ggml_permute(ctx, v_in, 0, 2, 1, 3));
         v_in = ggml_reshape_3d(ctx, v_in, d_head, L_k, n_kv_head * N);
+        if (pad_head) {
+            v_in = ggml_pad(ctx, v_in, 64 - d_head, 0, 0, 0);
+        }
         if (kv_scale != 1.0f) {
             v_in = ggml_ext_scale(ctx, v_in, kv_scale);
         }
@@ -679,6 +690,9 @@ ggml_tensor* ggml_ext_attention_ext(ggml_context* ctx,
         if (kv_scale != 1.0f) {
             out = ggml_ext_scale(ctx, out, 1.0f / kv_scale);
         }
+        if (pad_head) {
+            out = ggml_ext_slice(ctx, out, 0, 0, d_head);
+        }
         return out;
     };
 

src/core/ggml_extend_backend.cpp

@@ -8,6 +8,10 @@
 #include <stdexcept>
 #include <vector>
 
+#ifdef SD_USE_CUDA
+#include <cuda.h>
+#endif
+
 #include "core/util.h"
 #include "ggml/src/ggml-impl.h"
 #include "stable-diffusion.h"
@@ -429,6 +433,70 @@ bool sd_backend_is_cpu(ggml_backend_t backend) {
     return dev != nullptr && ggml_backend_dev_type(dev) == GGML_BACKEND_DEVICE_TYPE_CPU;
 }
 
+bool sd_backend_supports_cuda_mma(ggml_backend_t backend) {
+#ifdef SD_USE_CUDA
+    if (!sd_backend_is(backend, "CUDA")) {
+        return false;
+    }
+    auto dev = ggml_backend_get_device(backend);
+    if (dev == nullptr) {
+        return false;
+    }
+    static std::mutex mutex;
+    static std::unordered_map<ggml_backend_dev_t, bool> cache;
+    std::lock_guard<std::mutex> lock(mutex);
+    auto it = cache.find(dev);
+    if (it != cache.end()) {
+        return it->second;
+    }
+    const bool supported = [&]() {
+        ggml_backend_dev_props props{};
+        ggml_backend_dev_get_props(dev, &props);
+        CUdevice device;
+        int major = 0, minor = 0;
+        if (props.device_id == nullptr || cuInit(0) != CUDA_SUCCESS ||
+            cuDeviceGetByPCIBusId(&device, props.device_id) != CUDA_SUCCESS ||
+            cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, device) != CUDA_SUCCESS ||
+            cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, device) != CUDA_SUCCESS) {
+            return false;
+        }
+        auto reg          = ggml_backend_dev_backend_reg(dev);
+        auto get_features = reinterpret_cast<ggml_backend_get_features_t>(
+            ggml_backend_reg_get_proc_address(reg, "ggml_backend_get_features"));
+        if (get_features == nullptr) {
+            return false;
+        }
+        // Match ggml's highest compiled architecture for this device, including PTX fallback.
+        const int cc      = 100 * major + 10 * minor;
+        int compiled_arch = 0;
+        for (auto feature = get_features(reg); feature != nullptr && feature->name != nullptr; ++feature) {
+            if (std::strcmp(feature->name, "ARCHS") != 0 || feature->value == nullptr) {
+                continue;
+            }
+            const char* arch = feature->value;
+            while (*arch != '\0') {
+                char* end        = nullptr;
+                const long value = std::strtol(arch, &end, 10);
+                if (end == arch) {
+                    ++arch;
+                    continue;
+                }
+                if (value <= cc && value > compiled_arch) {
+                    compiled_arch = static_cast<int>(value);
+                }
+                arch = end;
+            }
+        }
+        return compiled_arch == 700 || compiled_arch >= 750;
+    }();
+    cache.emplace(dev, supported);
+    return supported;
+#else
+    (void)backend;
+    return false;
+#endif
+}
+
 ggml_backend_t sd_backend_cpu_init() {
     ggml_backend_load_all_once();
     return ggml_backend_init_by_type(GGML_BACKEND_DEVICE_TYPE_CPU, nullptr);

src/core/ggml_extend_backend.h

@@ -87,6 +87,7 @@ private:
 
 bool sd_backend_is(ggml_backend_t backend, const std::string& name);
 bool sd_backend_is_cpu(ggml_backend_t backend);
+bool sd_backend_supports_cuda_mma(ggml_backend_t backend);
 ggml_backend_t sd_backend_cpu_init();
 bool sd_backend_cpu_set_n_threads(ggml_backend_t backend_cpu, int n_threads);
 ggml_status sd_backend_graph_compute_with_eval_callback(ggml_backend_t backend,