vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml : remove ggml_task_type and GGML_PERF (#8017)

Browse source 
* ggml : remove ggml_task_type and GGML_PERF

* check abort_callback on main thread only

* vulkan : remove usage of ggml_compute_params

* remove LLAMA_PERF
This commit is contained in:
slaren 2024-06-24 03:07:59 +02:00 • committed by GitHub
parent e112b610a1
commit 95f57bb5d5
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
8 changed files with 402 additions and 1082 deletions
Download patch file Download diff file Expand all files Collapse all files

36
ggml-vulkan.cpp
View file

@ -513,8 +513,8 @@ static size_t vk_skip_checks;
static size_t vk_output_tensor;
static void ggml_vk_print_tensor(ggml_backend * ctx, const ggml_tensor * tensor, const char * name);
static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor);
static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor);
static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * tensor);
static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * tensor);
#endif
typedef void (*ggml_vk_func_t)(ggml_backend_vk_context * ctx, vk_context * subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst);
@ -5644,7 +5644,7 @@ static void ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
}
}
static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor){
static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor * tensor){
ggml_tensor_extra_gpu * extra = nullptr;
switch (tensor->op) {
@ -5697,17 +5697,10 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_compute_
return false;
}
if (params->ith != 0) {
return true;
}
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE) {
return true;
}
VK_LOG_DEBUG("ggml_vk_compute_forward(" << tensor << ", name=" << tensor->name << ", op=" << ggml_op_name(tensor->op) << ", type=" << tensor->type << ", ne0=" << tensor->ne[0] << ", ne1=" << tensor->ne[1] << ", ne2=" << tensor->ne[2] << ", ne3=" << tensor->ne[3] << ", nb0=" << tensor->nb[0] << ", nb1=" << tensor->nb[1] << ", nb2=" << tensor->nb[2] << ", nb3=" << tensor->nb[3] << ", view_src=" << tensor->view_src << ", view_offs=" << tensor->view_offs << ")");
#ifdef GGML_VULKAN_CHECK_RESULTS
ggml_vk_check_results_0(ctx, params, tensor);
ggml_vk_check_results_0(ctx, tensor);
#endif
vk_context& subctx = ctx->gc.contexts[extra->ctx_idx];
@ -6214,9 +6207,6 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
ggml_vk_build_graph(ctx,cgraph->nodes[i], i == last_node);
}
ggml_compute_params params = {};
params.type = GGML_TASK_TYPE_COMPUTE;
params.ith = 0;
for (int i = 0; i < cgraph->n_nodes; i++) {
ggml_tensor * node = cgraph->nodes[i];
@ -6224,13 +6214,13 @@ GGML_CALL static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backen
continue;
}
bool ok = ggml_vk_compute_forward(ctx, &params, node);
bool ok = ggml_vk_compute_forward(ctx, node);
if (!ok) {
fprintf(stderr, "%s: error: op not supported %s (%s)\n", __func__, node->name, ggml_op_name(node->op));
}
#ifdef GGML_VULKAN_CHECK_RESULTS
else {
ggml_vk_check_results_1(ctx, &params, node);
ggml_vk_check_results_1(ctx, node);
}
#endif
GGML_ASSERT(ok);
@ -6600,11 +6590,8 @@ void * comp_result;
size_t comp_size;
size_t comp_nb[GGML_MAX_DIMS];
size_t check_counter = 0;
static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor) {
if (params->ith != 0) {
return;
}
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_tensor * tensor) {
if (tensor->op == GGML_OP_TRANSPOSE) {
return;
}
@ -6908,11 +6895,8 @@ static void ggml_vk_check_results_0(ggml_backend_vk_context * ctx, ggml_compute_
ggml_free(ggml_ctx);
}
static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_compute_params * params, ggml_tensor * tensor) {
if (params->ith != 0) {
return;
}
if (params->type == GGML_TASK_TYPE_INIT || params->type == GGML_TASK_TYPE_FINALIZE || tensor->op == GGML_OP_TRANSPOSE) {
static void ggml_vk_check_results_1(ggml_backend_vk_context * ctx, ggml_tensor * tensor) {
if (tensor->op == GGML_OP_TRANSPOSE) {
return;
}
if (!(vk_output_tensor > 0 && vk_output_tensor == check_counter) && check_counter <= vk_skip_checks) {