happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

vulkan: disable large mmv subgroups on older Nvidia GPUs (#15717)

This commit is contained in:
Ruben Ortlam 2025-09-01 20:58:35 +02:00 committed by GitHub
parent 078ce23ea7
commit fec7911f8f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 1 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -4662,7 +4662,7 @@ static vk_pipeline ggml_vk_get_dequantize_mul_mat_vec(ggml_backend_vk_context *
// heuristic to choose workgroup size
uint32_t dmmv_wg = DMMV_WG_SIZE_SUBGROUP;
if (ctx->device->vendor_id == VK_VENDOR_ID_NVIDIA || ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
if ((ctx->device->vendor_id == VK_VENDOR_ID_NVIDIA && ctx->device->architecture != vk_device_architecture::NVIDIA_PRE_TURING) || ctx->device->vendor_id == VK_VENDOR_ID_INTEL) {
// Prefer larger workgroups when M is small, to spread the work out more
// and keep more SMs busy.
// q6_k seems to prefer small workgroup size even for "medium" values of M.