From b0eab053eb25ce35500a98688d959bb58a048e45 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Fri, 2 Jan 2026 13:31:34 -0600 Subject: [PATCH] address feedback --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 6ad1ce70c8..e1045bac85 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -4172,8 +4172,8 @@ static void ggml_vk_load_shaders(vk_device& device) { ggml_vk_create_pipeline(device, device->pipeline_sum_rows_f32, "sum_rows_f32", sum_rows_f32_len, sum_rows_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { device->subgroup_size }, 1); - const uint32_t CUMSUM_ELEM_PER_THREAD = device->vendor_id == VK_VENDOR_ID_NVIDIA ? 4 : 2; - ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 256, device->subgroup_size, CUMSUM_ELEM_PER_THREAD }, 1, true, true, device->subgroup_size); + const uint32_t cumsum_elem_per_thread = (device->vendor_id == VK_VENDOR_ID_AMD || device->vendor_id == VK_VENDOR_ID_INTEL) ? 2 : 4; + ggml_vk_create_pipeline(device, device->pipeline_cumsum_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 256, device->subgroup_size, cumsum_elem_per_thread }, 1, true, true, device->subgroup_size); ggml_vk_create_pipeline(device, device->pipeline_cumsum_small_f32, "cumsum_f32", cumsum_f32_len, cumsum_f32_data, "main", 2, sizeof(vk_op_sum_rows_push_constants), {1, 1, 1}, { 128, device->subgroup_size, 1 }, 1, true, true, device->subgroup_size); ggml_vk_create_pipeline(device, device->pipeline_cumsum_multipass1_f32, "cumsum_multipass1_f32", cumsum_multipass1_f32_len, cumsum_multipass1_f32_data, "main", 3, sizeof(vk_op_sum_rows_push_constants), {256, 1, 1}, { 256, device->subgroup_size }, 1, true, true, device->subgroup_size); ggml_vk_create_pipeline(device, device->pipeline_cumsum_multipass2_f32, "cumsum_multipass2_f32", cumsum_multipass2_f32_len, cumsum_multipass2_f32_data, "main", 3, sizeof(vk_op_sum_rows_push_constants), {256, 1, 1}, { 256, device->subgroup_size }, 1, true, true, device->subgroup_size);