upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-27 16:37:29 -05:00
Code Issues Packages Projects Releases Wiki Activity

Fixing the vulkan build issue of legacy GLSLC version that has no cooperativeMatrix API support (https://github.com/ggml-org/llama.cpp/issues/29373) (#29409)

Browse Source 
* vulkan : fix build issue of legacy glslc version by adding GGML_VULKAN_COOPMAT_GLSLC_SUPPORT macro check for Intel FA shader compiling

* vulkan : add preprocess condition to filter out unsupported FA 2 phases kernels before creation.

* vulkan : move lock_guard for Intel FA shader pointer creation under CM1 compiling preprocessor
This commit is contained in:
sliu39
2026-09-25 14:33:49 +02:00
committed by GitHub
parent b248f4a3c1
commit e85e15cf6d
2 changed files with 22 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/ggml-vulkan.cpp
+20 -12
View File
@@ -3170,6 +3170,7 @@ void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, device->pipeline_matmul_split_k_reduce, "split_k_reduce", split_k_reduce_len, split_k_reduce_data, "main", 2, 2 * sizeof(uint32_t), {256 * 4, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_flash_attn_split_k_reduce, "fa_split_k_reduce", fa_split_k_reduce_len, fa_split_k_reduce_data, "main", 3, sizeof(vk_op_flash_attn_split_k_reduce_push_constants), {1, device->subgroup_size, 1}, {device->subgroup_size}, 1, true);
#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
if (device->vendor_id == VK_VENDOR_ID_INTEL && (device->architecture == INTEL_XE2 || (device->architecture == INTEL_XE1 && device->coopmat_support && device->uma))) {
auto upper_power_of_2 = [&](uint32_t in) {
GGML_ASSERT(in != 0);
@@ -3209,6 +3210,7 @@ void ggml_vk_load_shaders(vk_device& device, vk_pipeline requested) {
ggml_vk_create_pipeline(device, pipelines.second, "xe_fa_decode_ph2", fa_decode_ph2_cm1_len, fa_decode_ph2_cm1_data, "main", 5, sizeof(vk_fa_xe_opt_push_constants), { 1, 1, 1 }, { group_sz_ph2, gqa_ratio, head_dim_pv, out_per_wg_ph2, xe_native_sub_group_size, split_p_per_iter_ph2, split_p_chunk, out_dim_per_wg }, 1, false, true, xe_native_sub_group_size);
}
}
#endif
for (auto &it : device->pipeline_fa_mask_opt) {
auto BrBc = it.first;
@@ -8219,10 +8221,11 @@ void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const
split_k = CEIL_DIV(KV, split_kv);
xe_fa_opt = xe_fa_supported_platform && xe_fa_supported_usage && xe_fa_supported_dtype;
if (xe_fa_opt) {
std::lock_guard<std::mutex> guard(ctx->device->compile_mutex);
const uint32_t split_p_size = 32;
const size_t max_dim = (nek1 + split_p_size - 1) / split_p_size;
const size_t p_dim = max_dim * split_p_size;
#if defined(VK_KHR_cooperative_matrix) && defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
std::lock_guard<std::mutex> guard(ctx->device->compile_mutex);
auto& pipelines = ctx->device->pipeline_xe_fa_decode_dual_phases;
auto it = pipelines.find({ (uint32_t)neq0, (uint32_t)nev0, qk_ratio, (uint32_t)neq1 });
if (it != pipelines.end()) {
@@ -8230,18 +8233,23 @@ void ggml_vk_flash_attn(ggml_backend_vk_context * ctx, vk_context& subctx, const
} else {
pipelines[{(uint32_t)neq0, (uint32_t)nev0, qk_ratio, (uint32_t)neq1}] = xe_fa_pipeline_dual_phases = std::make_pair(std::make_shared<vk_pipeline_struct>(), std::make_shared<vk_pipeline_struct>());
}
#endif
if (xe_fa_pipeline_dual_phases.first == nullptr || xe_fa_pipeline_dual_phases.second == nullptr) {
xe_fa_opt = false;
fa_copy_qstate = false;
} else {
size_p = neq1 * neq2 * p_dim * neq3 * sizeof(ggml_fp16_t);
size_group_max = neq1 * neq2 * max_dim * neq3 * sizeof(float);
size_t temp_size = ggml_nelements(q) * sizeof(ggml_fp16_t) + size_p + size_group_max;
fa_copy_qstate = true;
if (ctx->prealloc_size_x < temp_size) {
ctx->prealloc_size_x = temp_size;
ggml_vk_preallocate_buffers(ctx, subctx);
}
size_p = neq1 * neq2 * p_dim * neq3 * sizeof(ggml_fp16_t);
size_group_max = neq1 * neq2 * max_dim * neq3 * sizeof(float);
size_t temp_size = ggml_nelements(q) * sizeof(ggml_fp16_t) + size_p + size_group_max;
fa_copy_qstate = true;
if (ctx->prealloc_size_x < temp_size) {
ctx->prealloc_size_x = temp_size;
ggml_vk_preallocate_buffers(ctx, subctx);
}
if (ctx->prealloc_x_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
if (ctx->prealloc_x_need_sync) {
ggml_vk_sync_buffers(ctx, subctx);
}
}
}
}
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
+2
View File
@@ -929,8 +929,10 @@ void process_shaders() {
string_to_spv("fa_mask_opt", "flash_attn_mask_opt.comp", {});
#if defined(GGML_VULKAN_COOPMAT_GLSLC_SUPPORT)
string_to_spv("fa_decode_ph1", "flash_attn_decode_phase_1.comp", {}, true, true, false, false);
string_to_spv("fa_decode_ph2", "flash_attn_decode_phase_2.comp", {}, true, true, false, false);
#endif
string_to_spv("fa_sparse_compact", "flash_attn_sparse_compact.comp", {});
string_to_spv("fa_sparse_compact_subgroup", "flash_attn_sparse_compact.comp", {{"USE_SUBGROUPS", "1"}});