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

address feedback

This commit is contained in:
Jeff Bolz 2026-02-13 09:46:44 -06:00
parent 66d7c14359
commit feb3eef27e
1 changed files with 7 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -6768,12 +6768,12 @@ static void ggml_vk_matmul(
uint32_t padded_n) {
VK_LOG_DEBUG("ggml_vk_matmul(a: (" << a.buffer->buffer << ", " << a.offset << ", " << a.size << "), b: (" << b.buffer->buffer << ", " << b.offset << ", " << b.size << "), d: (" << d.buffer->buffer << ", " << d.offset << ", " << d.size << "), split_k: (" << (split_k_buffer.buffer != nullptr ? split_k_buffer.buffer->buffer : VK_NULL_HANDLE) << ", " << split_k_buffer.offset << ", " << split_k_buffer.size << "), m: " << m << ", n: " << n << ", k: " << k << ", stride_a: " << stride_a << ", stride_b: " << stride_b << ", stride_d: " << stride_d << ", batch_stride_a: " << batch_stride_a << ", batch_stride_b: " << batch_stride_b << ", batch_stride_d: " << batch_stride_d << ", split_k: " << split_k << ", batch: " << batch << ", ne02: " << ne02 << ", ne12: " << ne12 << ", broadcast2: " << broadcast2 << ", broadcast3: " << broadcast3 << ", padded_n: " << padded_n << ")");
if (split_k == 1) {
ggml_pipeline_request_descriptor_sets(ctx, pipeline, CEIL_DIV(batch, ctx->device->properties.limits.maxComputeWorkGroupCount[2]));
uint32_t base_work_group_z = 0;
while (base_work_group_z < batch) {
uint32_t groups_z = std::min(batch - base_work_group_z, ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
const vk_mat_mat_push_constants pc = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, base_work_group_z, batch, k, ne02, ne12, broadcast2, broadcast3, padded_n };
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, d }, pc, { m, n, groups_z });
base_work_group_z += groups_z;
@ -6791,15 +6791,16 @@ static void ggml_vk_matmul(
uint32_t k_split = CEIL_DIV(k, split_k);
k_split = ROUNDUP_POW2(k_split, 256);
ggml_pipeline_request_descriptor_sets(ctx, pipeline, CEIL_DIV(batch, ctx->device->properties.limits.maxComputeWorkGroupCount[2]));
uint32_t base_work_group_z = 0;
while (base_work_group_z < batch) {
uint32_t groups_z = std::min(batch - base_work_group_z, ctx->device->properties.limits.maxComputeWorkGroupCount[2]);
ggml_pipeline_request_descriptor_sets(ctx, pipeline, 1);
const vk_mat_mat_push_constants pc1 = { m, n, k, stride_a, stride_b, stride_d, batch_stride_a, batch_stride_b, batch_stride_d, base_work_group_z, batch, k_split, ne02, ne12, broadcast2, broadcast3, padded_n };
// Make sure enough workgroups get assigned for split k to work
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, { a, b, split_k_buffer }, pc1, { (CEIL_DIV(m, pipeline->wg_denoms[0]) * pipeline->wg_denoms[0]) * split_k, n, groups_z });
base_work_group_z += groups_z;
}
ggml_vk_sync_buffers(ctx, subctx);
const std::array<uint32_t, 2> pc2 = { (uint32_t)(m * n * batch), split_k };
@ -7587,9 +7588,10 @@ static void ggml_vk_mul_mat_vec_q_f16(ggml_backend_vk_context * ctx, vk_context&
fusion_flags |= MAT_VEC_FUSION_FLAGS_BIAS1;
}
ggml_pipeline_request_descriptor_sets(ctx, dmmv, CEIL_DIV(ne12 * ne13, ctx->device->properties.limits.maxComputeWorkGroupCount[1]));
uint32_t base_work_group_y = 0;
while (base_work_group_y < ne12 * ne13) {
ggml_pipeline_request_descriptor_sets(ctx, dmmv, 1);
uint32_t groups_y = std::min((uint32_t)(ne12 * ne13) - base_work_group_y, ctx->device->properties.limits.maxComputeWorkGroupCount[1]);
const vk_mat_vec_push_constants pc = {