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

OV Frontend supports GET_ROWS/RMS_NORM/MUL/MUL_MAT graph conversion of consecutive OPs

This commit is contained in:
zhanmyz 2025-03-04 00:05:00 +08:00 committed by Mustafa Cavus
parent 901f7347ff
commit 95ae982d59
3 changed files with 78 additions and 43 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

64
ggml/src/ggml-openvino.cpp
View File

@ -1020,39 +1020,41 @@ static enum ggml_status ggml_backend_openvino_graph_compute(ggml_backend_t backe
} }
} }
int end_node = cgraph->n_nodes - 1;
openvino_frontend_compute(backend, cgraph, 0, end_node);
// openvino_frontend_compute(backend, cgraph); // openvino_frontend_compute(backend, cgraph);
// Process nodes in order // Process nodes in order
for (int i = 0; i < cgraph->n_nodes; i++) { // for (int i = 0; i < cgraph->n_nodes; i++) {
if (std::find(permute_indices.begin(), permute_indices.end(), i) != permute_indices.end()) { // if (std::find(permute_indices.begin(), permute_indices.end(), i) != permute_indices.end()) {
ggml_backend_openvino_permute(cgraph->nodes[i]); // ggml_backend_openvino_permute(cgraph->nodes[i]);
// } else if (std::find(cont_indices.begin(), cont_indices.end(), i) != cont_indices.end()) { // // } else if (std::find(cont_indices.begin(), cont_indices.end(), i) != cont_indices.end()) {
// ggml_backend_openvino_dup_bytes(cgraph->nodes[i]); // // ggml_backend_openvino_dup_bytes(cgraph->nodes[i]);
// } else if (std::find(view_indices.begin(), view_indices.end(), i) != view_indices.end()) { // // } else if (std::find(view_indices.begin(), view_indices.end(), i) != view_indices.end()) {
// ggml_backend_openvino_view(cgraph->nodes[i]); // // ggml_backend_openvino_view(cgraph->nodes[i]);
// } else if (std::find(cpy_indices.begin(), cpy_indices.end(), i) != cpy_indices.end()) { // // } else if (std::find(cpy_indices.begin(), cpy_indices.end(), i) != cpy_indices.end()) {
// ggml_backend_openvino_cpy(cgraph->nodes[i]); // // ggml_backend_openvino_cpy(cgraph->nodes[i]);
} else if (std::find(transpose_indices.begin(), transpose_indices.end(), i) != transpose_indices.end()) { // // } else if (std::find(transpose_indices.begin(), transpose_indices.end(), i) != transpose_indices.end()) {
ggml_backend_openvino_transpose(cgraph->nodes[i]); // // ggml_backend_openvino_transpose(cgraph->nodes[i]);
} else if (std::find(reshape_indices.begin(), reshape_indices.end(), i) != reshape_indices.end()) { // // } else if (std::find(reshape_indices.begin(), reshape_indices.end(), i) != reshape_indices.end()) {
ggml_backend_openvino_reshape(cgraph->nodes[i]); // // ggml_backend_openvino_reshape(cgraph->nodes[i]);
// } else if (std::find(mul_mat_indices.begin(), mul_mat_indices.end(), i) != mul_mat_indices.end()) { // // } else if (std::find(mul_mat_indices.begin(), mul_mat_indices.end(), i) != mul_mat_indices.end()) {
// ggml_backend_openvino_mul_mat(cgraph->nodes[i]); // // ggml_backend_openvino_mul_mat(cgraph->nodes[i]);
} else { // } else {
// Process a range of nodes with openvino_frontend_compute // // Process a range of nodes with openvino_frontend_compute
int start_index = i; // int start_index = i;
while (i < cgraph->n_nodes // while (i < cgraph->n_nodes
// && std::find(view_indices.begin(), view_indices.end(), i) == view_indices.end() // // && std::find(view_indices.begin(), view_indices.end(), i) == view_indices.end()
// && std::find(cpy_indices.begin(), cpy_indices.end(), i) == cpy_indices.end() // // && std::find(cpy_indices.begin(), cpy_indices.end(), i) == cpy_indices.end()
// && std::find(cont_indices.begin(), cont_indices.end(), i) == cont_indices.end() // // && std::find(cont_indices.begin(), cont_indices.end(), i) == cont_indices.end()
// && std::find(mul_mat_indices.begin(), mul_mat_indices.end(), i) == mul_mat_indices.end() // // && std::find(mul_mat_indices.begin(), mul_mat_indices.end(), i) == mul_mat_indices.end()
) { // ) {
i++; // i++;
} // }
if (start_index < i) { // if (start_index < i) {
openvino_frontend_compute(backend, cgraph, start_index, --i); // openvino_frontend_compute(backend, cgraph, start_index, --i);
} // }
} // }
} // }
return GGML_STATUS_SUCCESS; return GGML_STATUS_SUCCESS;

46
ggml/src/ggml-openvino/ggml-decoder.cpp
View File

@ -20,6 +20,7 @@ void GgmlOvDecoder::set_input_output(ggml_tensor* node, std::map<std::string, gg
std::string src0_name = std::string(node->src[0]->name); std::string src0_name = std::string(node->src[0]->name);
std::string node_name = std::string(node->name); std::string node_name = std::string(node->name);
switch (node->op) { switch (node->op) {
// Unary OPs // Unary OPs
case GGML_OP_UNARY: case GGML_OP_UNARY:
@ -110,7 +111,7 @@ void GgmlOvDecoder::set_input_output(ggml_tensor* node, std::map<std::string, gg
m_output_names.push_back(node_name); m_output_names.push_back(node_name);
m_continuous = true; m_continuous = true;
ov::Shape src_shape(node->src[0]->ne, node->src[0]->ne + 4); ov::Shape src_shape(node->src[0]->ne, node->src[0]->ne + 3);
auto input_param = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, src_shape); auto input_param = std::make_shared<ov::op::v0::Parameter>(ov::element::f32, src_shape);
m_params.push_back(input_param); m_params.push_back(input_param);
break; break;
@ -217,6 +218,7 @@ void GgmlOvDecoder::set_input_output(ggml_tensor* node, std::map<std::string, gg
m_node_op_name[src0_name] = ggml_op_name(node->op); m_node_op_name[src0_name] = ggml_op_name(node->op);
m_output_names.push_back(node_name); m_output_names.push_back(node_name);
if (node->src[1]) { if (node->src[1]) {
// std::string src1_name = std::string(node->src[1]->name) + "_" + std::to_string(node->src[1]->view_offs) + "_input_" + ggml_op_name(node->src[1]->op);
// std::string src1_name = std::string(node->src[1]->name) + "_" + std::to_string(node->src[1]->view_offs); // std::string src1_name = std::string(node->src[1]->name) + "_" + std::to_string(node->src[1]->view_offs);
std::string src1_name = std::string(node->src[1]->name); std::string src1_name = std::string(node->src[1]->name);
inputs[src1_name] = node->src[1]; inputs[src1_name] = node->src[1];
@ -228,6 +230,7 @@ void GgmlOvDecoder::set_input_output(ggml_tensor* node, std::map<std::string, gg
// OPs with 3 inputs: // OPs with 3 inputs:
case GGML_OP_ROPE: case GGML_OP_ROPE:
{ {
// std::string src1_name = std::string(node->src[1]->name) + "_" + std::to_string(node->src[1]->view_offs) + "_input_" + ggml_op_name(node->src[1]->op);
// std::string src1_name = std::string(node->src[1]->name) + "_" + std::to_string(node->src[1]->view_offs); // std::string src1_name = std::string(node->src[1]->name) + "_" + std::to_string(node->src[1]->view_offs);
std::string src1_name = std::string(node->src[1]->name); std::string src1_name = std::string(node->src[1]->name);
inputs[src0_name] = node->src[0]; inputs[src0_name] = node->src[0];
@ -239,6 +242,7 @@ void GgmlOvDecoder::set_input_output(ggml_tensor* node, std::map<std::string, gg
outputs[node_name] = node; outputs[node_name] = node;
m_output_names.push_back(node_name); m_output_names.push_back(node_name);
if (node->src[2]) { if (node->src[2]) {
// std::string src2_name = std::string(node->src[2]->name) + "_" + std::to_string(node->src[2]->view_offs) + "_input_" + ggml_op_name(node->src[2]->op);
// std::string src2_name = std::string(node->src[2]->name) + "_" + std::to_string(node->src[2]->view_offs); // std::string src2_name = std::string(node->src[2]->name) + "_" + std::to_string(node->src[2]->view_offs);
std::string src2_name = std::string(node->src[2]->name); std::string src2_name = std::string(node->src[2]->name);
inputs[src2_name] = node->src[2]; inputs[src2_name] = node->src[2];
@ -253,7 +257,7 @@ void GgmlOvDecoder::set_input_output(ggml_tensor* node, std::map<std::string, gg
} }
void ggml_graph_op_print(const struct ggml_cgraph * cgraph) { void ggml_graph_op_print(const struct ggml_cgraph * cgraph) {
std::ofstream file("2_graph_node_src_op_name.txt"); std::ofstream file("01_nodes.txt");
if (!file.is_open()) { if (!file.is_open()) {
std::cerr << "Failed to open file" << std::endl; std::cerr << "Failed to open file" << std::endl;
return; return;
@ -262,6 +266,13 @@ void ggml_graph_op_print(const struct ggml_cgraph * cgraph) {
file << "=== GRAPH ===\n"; file << "=== GRAPH ===\n";
file << "n_nodes = " << cgraph->n_nodes << "\n"; file << "n_nodes = " << cgraph->n_nodes << "\n";
file << " " << std::setw(3) << "nodes"
<< std::setw(15) << "shape"
<< std::setw(16) << "op"
<< std::setw(20) << "name"
<< std::setw(3) << " "
<< std::setw(50) << "stride"
<< "\n";
for (int i = 0; i < cgraph->n_nodes; i++) { for (int i = 0; i < cgraph->n_nodes; i++) {
struct ggml_tensor * node = cgraph->nodes[i]; struct ggml_tensor * node = cgraph->nodes[i];
@ -269,9 +280,14 @@ void ggml_graph_op_print(const struct ggml_cgraph * cgraph) {
<< std::setw(5) << node->ne[0] << ", " << std::setw(5) << node->ne[0] << ", "
<< std::setw(5) << node->ne[1] << ", " << std::setw(5) << node->ne[1] << ", "
<< std::setw(5) << node->ne[2] << "] " << std::setw(5) << node->ne[2] << "] "
<< std::left << std::setw(16) << ggml_op_name(node->op) << std::right << " " << std::left << std::setw(20) << ggml_op_name(node->op) << std::right << " "
<< " " << node->name << std::left << std::setw(44) << node->name << std::right
<< ((node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" : node->grad ? "g" : " ") << "\n"; << ((node->flags & GGML_TENSOR_FLAG_PARAM) ? "x" : node->grad ? "g" : " ")
<< std::setw(2) << "[ "
<< std::setw(0) << node->nb[0] << ", "
<< std::setw(5) << node->nb[1] << ", "
<< std::setw(5) << node->nb[2] << "] "
<< "\n";
if (node->src[0]) { if (node->src[0]) {
file << std::setw(10) << " [ " file << std::setw(10) << " [ "
@ -279,15 +295,19 @@ void ggml_graph_op_print(const struct ggml_cgraph * cgraph) {
<< std::setw(5) << node->src[0]->ne[1] << ", " << std::setw(5) << node->src[0]->ne[1] << ", "
<< std::setw(5) << node->src[0]->ne[2] << "] " << std::setw(5) << node->src[0]->ne[2] << "] "
<< std::setw(12) << std::setw(12)
<< "0: " << ggml_op_name(node->src[0]->op) << " "; << "0: " << std::left << std::setw(12) << ggml_op_name(node->src[0]->op) << std::right;
// // Custom logic to handle '\000' // // Custom logic to handle '\000'
// const char* name_ptr = node->src[0]->name; // const char* name_ptr = node->src[0]->name;
// while (*name_ptr != '\0' || *(name_ptr + 1) != '\0' || *(name_ptr + 2) != '\0') { // while (*name_ptr != '\0' || *(name_ptr + 1) != '\0' || *(name_ptr + 2) != '\0') {
// file << *name_ptr; // file << *name_ptr;
// name_ptr++; // name_ptr++;
// } // }
file << node->src[0]->name; file << std::left << std::setw(30) << node->src[0]->name << std::right
file << "\n"; << std::setw(16) << "[ "
<< std::setw(0) << node->src[0]->nb[0] << ", "
<< std::setw(5) << node->src[0]->nb[1] << ", "
<< std::setw(5) << node->src[0]->nb[2] << "] "
<< "\n";
} }
if (node->src[1]) { if (node->src[1]) {
file << std::setw(10) << " [ " file << std::setw(10) << " [ "
@ -295,15 +315,19 @@ void ggml_graph_op_print(const struct ggml_cgraph * cgraph) {
<< std::setw(5) << node->src[1]->ne[1] << ", " << std::setw(5) << node->src[1]->ne[1] << ", "
<< std::setw(5) << node->src[1]->ne[2] << "] " << std::setw(5) << node->src[1]->ne[2] << "] "
<< std::setw(12) << std::setw(12)
<< "1: " << ggml_op_name(node->src[1]->op) << " "; << "1: " << std::left << std::setw(12) << ggml_op_name(node->src[1]->op) << std::right;
// // Custom logic to handle '\000' // // Custom logic to handle '\000'
// const char* name_ptr = node->src[1]->name; // const char* name_ptr = node->src[1]->name;
// while (*name_ptr != '\0' || *(name_ptr + 1) != '\0' || *(name_ptr + 2) != '\0') { // while (*name_ptr != '\0' || *(name_ptr + 1) != '\0' || *(name_ptr + 2) != '\0') {
// file << *name_ptr; // file << *name_ptr;
// name_ptr++; // name_ptr++;
// } // }
file << node->src[1]->name; file << std::left << std::setw(30) << node->src[1]->name << std::right
file << "\n"; << std::setw(16) << "[ "
<< std::setw(0) << node->src[1]->nb[0] << ", "
<< std::setw(5) << node->src[1]->nb[1] << ", "
<< std::setw(5) << node->src[1]->nb[2] << "] "
<< "\n";
} }
} }

11
ggml/src/ggml-openvino/utils.cpp
View File

@ -121,7 +121,7 @@ enum ggml_status openvino_frontend_compute(ggml_backend_t backend, struct ggml_c
// Convert InputModel -> ov::Model // Convert InputModel -> ov::Model
std::shared_ptr<ov::Model> model = front_end->convert(input_model); std::shared_ptr<ov::Model> model = front_end->convert(input_model);
// ov::save_model(model, "/home/user/zhan/merge_git_commits/llama.cpp-ov/001_model.xml"); ov::save_model(model, "/home/user/zhan/merge_git_commits/llama.cpp-ov/001_model.xml");
if (!model) { if (!model) {
GGML_LOG_ERROR("Model is not converted \n"); GGML_LOG_ERROR("Model is not converted \n");
@ -145,6 +145,14 @@ enum ggml_status openvino_frontend_compute(ggml_backend_t backend, struct ggml_c
// Set input tensor // Set input tensor
for (size_t i = 0; i < input_names.size(); i++) { for (size_t i = 0; i < input_names.size(); i++) {
infer_request.set_input_tensor(i, input_tensors[input_names[i]]); infer_request.set_input_tensor(i, input_tensors[input_names[i]]);
// auto input_tensor = infer_request.get_input_tensor(i);
// auto input_shape = input_tensor.get_shape();
// std::cout << "Input tensor " << i << " shape: ";
// for (const auto& dim : input_shape) {
// std::cout << dim << " ";
// }
// std::cout << std::endl;
} }
infer_request.infer(); infer_request.infer();
@ -155,6 +163,7 @@ enum ggml_status openvino_frontend_compute(ggml_backend_t backend, struct ggml_c
for (size_t i = 0; i < output_names.size(); i++) { for (size_t i = 0; i < output_names.size(); i++) {
// std::string op_name = ggml_decoder->get_node_op_name(output_names[i]); // std::string op_name = ggml_decoder->get_node_op_name(output_names[i]);
auto output_tensor = infer_request.get_output_tensor(i); auto output_tensor = infer_request.get_output_tensor(i);
// output_tensor.get_shape();
std::memcpy(output_tensors[output_names[i]], output_tensor.data(), output_tensor.get_byte_size()); std::memcpy(output_tensors[output_names[i]], output_tensor.data(), output_tensor.get_byte_size());
#ifdef GGML_OPENVINO_DEBUG #ifdef GGML_OPENVINO_DEBUG
printf("Output %s after: %g\n", output_names[i].c_str(), *(double*)(output_tensor.data())); printf("Output %s after: %g\n", output_names[i].c_str(), *(double*)(output_tensor.data()));