tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 20:05:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: support noncontiguous rms_norm (#13031)

Browse Source
This commit is contained in:
Jeff Bolz
2025-04-20 03:50:02 -05:00
committed by GitHub
parent 4ba9d711ba
commit 66168204be
2 changed files with 39 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
ggml/src/ggml-vulkan/ggml-vulkan.cpp
View File

@ -2397,7 +2397,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
ggml_vk_create_pipeline(device, device->pipeline_norm_f32, "norm_f32", norm_f32_len, norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_group_norm_f32, "group_norm_f32", group_norm_f32_len, group_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rms_norm_f32, "rms_norm_f32", rms_norm_f32_len, rms_norm_f32_data, "main", 2, sizeof(vk_op_unary_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_rms_norm_back_f32, "rms_norm_back_f32", rms_norm_back_f32_len, rms_norm_back_f32_data, "main", 3, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
ggml_vk_create_pipeline(device, device->pipeline_l2_norm_f32, "l2_norm_f32", l2_norm_f32_len, l2_norm_f32_data, "main", 2, sizeof(vk_op_push_constants), {1, 1, 1}, {}, 1);
@ -6006,6 +6006,7 @@ static bool ggml_vk_op_supports_incontiguous(ggml_op op) {
case GGML_OP_REPEAT:
case GGML_OP_REPEAT_BACK:
case GGML_OP_ROPE:
case GGML_OP_RMS_NORM:
return true;
default:
return false;
@ -6216,7 +6217,6 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
switch (op) {
case GGML_OP_NORM:
case GGML_OP_RMS_NORM:
case GGML_OP_RMS_NORM_BACK:
case GGML_OP_L2_NORM:
case GGML_OP_SOFT_MAX:
@ -6233,6 +6233,10 @@ static void ggml_vk_op_f32(ggml_backend_vk_context * ctx, vk_context& subctx, co
elements = { nr, 1, 1 };
}
} break;
case GGML_OP_RMS_NORM:
elements = { (uint32_t)ne01, (uint32_t)ne02, (uint32_t)ne03 };
break;
case GGML_OP_SUM:
// We use GGML_OP_SUM_ROWS with 1 row.
elements = { 1, 1, 1 };
@ -6883,7 +6887,17 @@ static void ggml_vk_group_norm(ggml_backend_vk_context * ctx, vk_context& subctx
static void ggml_vk_rms_norm(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, ggml_tensor * dst, bool dryrun = false) {
float * op_params = (float *)dst->op_params;
ggml_vk_op_f32<vk_op_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, { (uint32_t)src0->ne[0], (uint32_t)src0->ne[1], op_params[0], 0.0f }, dryrun);
const uint32_t src0_type_size = ggml_type_size(src0->type);
const uint32_t dst_type_size = ggml_type_size(dst->type);
ggml_vk_op_f32<vk_op_unary_push_constants>(ctx, subctx, src0, nullptr, nullptr, dst, GGML_OP_RMS_NORM, {
(uint32_t)ggml_nelements(src0),
(uint32_t)src0->ne[0], (uint32_t)src0->ne[1], (uint32_t)src0->ne[2], (uint32_t)src0->ne[3], (uint32_t)src0->nb[0] / src0_type_size, (uint32_t)src0->nb[1] / src0_type_size, (uint32_t)src0->nb[2] / src0_type_size, (uint32_t)src0->nb[3] / src0_type_size,
(uint32_t) dst->ne[0], (uint32_t) dst->ne[1], (uint32_t) dst->ne[2], (uint32_t) dst->ne[3], (uint32_t) dst->nb[0] / dst_type_size, (uint32_t) dst->nb[1] / dst_type_size, (uint32_t) dst->nb[2] / dst_type_size, (uint32_t) dst->nb[3] / dst_type_size,
0,
op_params[0], 0.0f,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
}, dryrun);
}
static void ggml_vk_rms_norm_back(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
@ -9388,10 +9402,10 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
case GGML_OP_VIEW:
case GGML_OP_PERMUTE:
case GGML_OP_TRANSPOSE:
case GGML_OP_RMS_NORM:
return true;
case GGML_OP_NORM:
case GGML_OP_GROUP_NORM:
case GGML_OP_RMS_NORM:
case GGML_OP_L2_NORM:
return ggml_is_contiguous(op->src[0]);
case GGML_OP_ADD: