Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

Add Bert for JNI. #173

Merged
merged 6 commits into from
Oct 30, 2024
Merged

Add Bert for JNI. #173

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
0ffa0bc
refactor: migrate JNI to new backend.
lx200916 Oct 28, 2024
ad85229
feat: add runForResult for JNI.
lx200916 Oct 28, 2024
04f4da1
Merge branch 'main' of github.com:UbiquitousLearning/mllm into main
lx200916 Oct 30, 2024
c626eaf
Merge branch 'main' of github.com:UbiquitousLearning/mllm into main
lx200916 Oct 30, 2024
0955e0a
Add Bert.
lx200916 Oct 30, 2024
803dc30
Merge branch 'main' into dev_and
yirongjie Oct 30, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions src/Module.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
#include <memory>
#include <utility>
#include <vector>
#include <unordered_map>

namespace mllm {

Expand Down
23 changes: 23 additions & 0 deletions src/models/fuyu/processing_fuyu.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -305,6 +305,29 @@ class FuyuProcessor final : public PreProcessor {
// auto batch_size = images_.size();
get_sample_encoding(text);
}
vector<Tensor> process(const std::string &text, const std::vector<uint8_t *> &images, const std::vector<size_t> &image_length) {
image_patch_indices_per_batch.clear();
image_patch_indices_per_subseq.clear();
image_patch_input_indices_.clear();
text_lengths_.clear();
image_input_ids_.clear();
attention_mask_.clear();
image_patches_indices_.clear();
image_patches_.clear();
text_ids_.clear();
images_.clear();

PreProcessImages(images, image_length);
Process(text);
auto input_ids = image_input_ids_;
if (input_ids.empty()) {
input_ids = text_ids_;
}
vector<Tensor> result = {mllm::Tokenizer::tokens2Input(input_ids[0], "input_ids"),
vector3d2Tensor(image_patches_, "image_patches"),
vector2d2Tensor(image_patches_indices_, "image_patches_indices")};
return result;
}

vector<Tensor> process(std::string &text, vector<string> image) {
image_patch_indices_per_batch.clear();
Expand Down
238 changes: 78 additions & 160 deletions tools/jni/LibHelper.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,101 +1,61 @@
//
// Created by Xiang Li on 2023/12/16.
//
#include "helper.hpp"
#include "processor/FuyuPreProcess.hpp"

#ifdef ANDROID_API
// #ifdef ANDROID_API

#include "LibHelper.hpp"
#include <iostream>
#include <Types.hpp>
#include <memory>
#include <utility>
#include <valarray>
#include "Net.hpp"
#include "Executor.hpp"
// #include "NetParameter.hpp"
#include "express/Express.hpp"
#include "tokenizers/BPE/Bpe.hpp"
#include "modeling_llama.hpp"
#include "modeling_fuyu.hpp"
#include <vector>
#include "models/bert/modeling_bert.hpp"
#include "models/bert/tokenization_bert.hpp"
#include "models/fuyu/configuration_fuyu.hpp"
#include "models/fuyu/modeling_fuyu.hpp"
#include "models/qwen/configuration_qwen.hpp"
#include "models/qwen/modeling_qwen.hpp"
#include "models/qwen/tokenization_qwen.hpp"
#include "tokenizers/Unigram/Unigram.hpp"
using namespace mllm;
#include "models/fuyu/processing_fuyu.hpp"

inline bool exists_test(const std::string &name) {
std::ifstream f(name.c_str());
return f.good();
}

unsigned int LibHelper::postProcessing(shared_ptr<Tensor> result, shared_ptr<Tensor> &out_result) const {
switch (model_) {
case LLAMA: {
return postProcessing_llama(result, out_result);
}
case FUYU: {
return postProcessing_Fuyu(result, out_result);
}
default: return 0;
}
// switch (model_) {
// case LLAMA: {
// return 0;
// }
// case FUYU: {
// // return chatPostProcessing(unsigned int token_idx, Tensor &tokens_tensor, const int &clean_tensors);
// }
// default: return 0;
// }
}

bool LibHelper::setUp(const std::string &base_path, std::string weights_path, std::string vocab_path, PreDefinedModel model, MLLMBackendType backend_type) {
c = new Context();
BackendConfig bn;
weights_path = base_path + weights_path;
vocab_path = base_path + vocab_path;
LOGI("Setup!");
// check path exists
if (!exists_test(weights_path) || !exists_test(vocab_path)) {
return false;
}

const auto param_loader = new ParamLoader(std::move(weights_path));
executor_ = new Executor(param_loader);
net_ = new Net(bn);
if (net_ == nullptr || executor_ == nullptr || !param_loader->isAvailible()) {
return false;
}
auto size = param_loader->getParamSize();
LOGI("param size:%d", size);
model_ = model;
LOGI("MODEL TYPE:%d", model_);

bool LibHelper::setUp(const std::string &base_path, std::string weights_path, std::string vocab_path, std::string merge_path, PreDefinedModel model, MLLMBackendType backend_type) {
FuyuConfig fuyuconfig(tokens_limit, "8B");
QWenConfig qwconfig(tokens_limit, "1.5B");
switch (model) {
case LLAMA: {
int vocab_size = 32000;
int hidden_dim = 4096;
int ffn_hidden_dim = 11008;
int mutil_head_size = 32;
llama2(c, vocab_size, hidden_dim, ffn_hidden_dim, mutil_head_size);
net_->convert(c->sub_param_, BackendType::MLLM_CPU);
tokenizer_ = new BPETokenizer(vocab_path);
eos_id_ = 2;
case LLAMA:
tokenizer_ = make_shared<QWenTokenizer>(vocab_path, merge_path);
module_ = make_shared<QWenForCausalLM>(qwconfig);
break;
}
case FUYU: {
int vocab_size = 262144;
int hidden_dim = 4096;
int ffn_hidden_dim = 4096 * 4;
int mutil_head_size = 64;
int patch_size = 30;
Fuyu(c, vocab_size, patch_size, 3, hidden_dim, ffn_hidden_dim, mutil_head_size);
net_->convert(c->sub_param_, BackendType::MLLM_CPU);
tokenizer_ = new UnigramTokenizer(vocab_path);
eos_id_ = 71013;

case FUYU:
processor_ = new FuyuProcessor(vocab_path);
module_ = make_shared<FuyuModel>(fuyuconfig);
break;
case Bert:
tokenizer_ = make_shared<BertTokenizer>(vocab_path, true);
module_ = make_shared<BertModel>(BertConfig());
break;
}
default: {
return false;
}
}
size = tokenizer_->getVocabSize();
LOGI("tokenizer size:%d", size);
if (!tokenizer_->isAvailible()) {
return false;
}

executor_->setup(net_);

module_->load(weights_path);
is_first_run_cond_ = true;
return true;
}
Expand All @@ -105,95 +65,53 @@ void LibHelper::setCallback(callback_t callback) {
}

void LibHelper::run(std::string &input_str, uint8_t *image, unsigned max_step, unsigned int image_length) {
// PreProcess
switch (model_) {
case LLAMA: {
if (input_str[0] != ' ') {
input_str = ' ' + input_str;
if (model_ == LLAMA) {
auto in_str = tokenizer_->apply_chat_template(input_str);
auto input_tensor = tokenizer_->tokenize(in_str);
LlmTextGeneratorOpts opt{
.max_new_tokens = max_step,
.do_sample = true,
.temperature = 0.3F,
.top_k = 50,
.top_p = 0.F,
};
module_->generate(input_tensor, opt, [&](unsigned int out_token) -> bool {
auto out_string = tokenizer_->detokenize({out_token});
auto [not_end, output_string] = tokenizer_->postprocess(out_string);
callback_(output_string, !not_end);
if (!not_end) { return false; }
return true;
});
module_->clear_kvcache();
} else if (model_ == FUYU) {
auto processor = dynamic_cast<FuyuProcessor *>(processor_);
auto input_tensors = processor->process(input_str, {image}, {image_length});
for (int step = 0; step < max_step; step++) {
auto result = (*module_)({input_tensors[0], input_tensors[1], input_tensors[2]});
auto outputs = processor->detokenize(result[0]);
auto out_string = outputs.first;
auto out_token = outputs.second;
auto [end, string] = processor->postprocess(out_string);
callback_(string, !end);
if (!end) { break; }
}
break;
} else if (model_ == Bert) {
LOGE("Bert model is not supported in this version.");
}
case FUYU: {
if (input_str[input_str.length() - 1] != '\n') {
input_str = input_str + '\n';
}
}
}
auto tokens_id = vector<token_id_t>();
shared_ptr<Tensor> input = std::make_shared<Tensor>();
shared_ptr<Tensor> input_id = std::make_shared<Tensor>();
shared_ptr<Tensor> img_patch = std::make_shared<Tensor>();
shared_ptr<Tensor> img_patch_id = std::make_shared<Tensor>();

if (model_ == FUYU) {
LOGI("Image Found!");
LOGI("%s\n", input_str.c_str());
auto pre_processor = FuyuPreProcess(tokenizer_);
pre_processor.images_.clear();
pre_processor.image_input_ids_.clear();
pre_processor.image_patches_indices_.clear();
pre_processor.image_patches_.clear();
if (image != nullptr && image_length > 0) {
pre_processor.PreProcessImages({image}, {image_length});
} else {
pre_processor.PreProcessImages({}, {0});
}
pre_processor.Process(input_str);
auto input_ids = pre_processor.image_input_ids_;
if (input_ids.empty()) {
input_ids = pre_processor.text_ids_;
}

UnigramTokenizer::token2Tensor(net_, input_ids[0], input_id);
const auto image_patches = pre_processor.image_patches_;
const auto image_patch_indices = pre_processor.image_patches_indices_;
patches2Tensor(img_patch, net_, image_patches);
patchIdx2Tensor(img_patch_id, net_, image_patch_indices);

} else if (model_ == LLAMA) {
tokenizer_->tokenize(input_str, tokens_id, true);
LOGI("is_first_run_cond_:%d", is_first_run_cond_);
if (is_first_run_cond_) {
is_first_run_cond_ = false;
} else {
LOGI("Keep Speaker!");
if (tokens_id[0] > 0) {
tokens_id[0] = 13;
}
}
mllm::Tokenizer::token2Tensor(net_, tokens_id, input);
}

auto out_string = std::string();

for (int step = 0; step < max_step; step++) {
unsigned int token_idx;
if (model_ == FUYU) {
LOGI("Image Patch!");
executor_->run(net_, {input_id, img_patch, img_patch_id});
auto result = executor_->result();
token_idx = postProcessing(result[0], input_id);
fullTensor(img_patch, net_, {0, 0, 0, 0}, 1.0F);
fullTensor(img_patch_id, net_, {0, 0, 0, 0}, 1.0F);
} else {
executor_->run(net_, {input});
auto result = executor_->result();
token_idx = postProcessing(result[0], input);
}
const auto out_token = tokenizer_->detokenize({token_idx});
if (out_token == "</s>" || token_idx == eos_id_) {
callback_(out_string, true);
break;
}
out_string += out_token;
callback_(out_string, step == max_step - 1);
}
std::vector<float> LibHelper::runForResult(std::string &input_str) {
if (model_ == Bert) {
auto bert_tokenizer = dynamic_pointer_cast<BertTokenizer>(tokenizer_);
auto [token_ids, type_ids, position_ids] = bert_tokenizer->process(input_str);
auto result = (*module_)({token_ids, type_ids, position_ids})[0];
auto output_arr = result.hostPtr<float>();
return std::vector<float>(output_arr, output_arr + result.count());
} else {
return {};
}
}

LibHelper::~LibHelper() {
delete c;
delete net_;
delete executor_;
delete tokenizer_;
delete processor_;
}
#endif
// #endif
38 changes: 19 additions & 19 deletions tools/jni/LibHelper.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,10 +4,11 @@

#ifndef LIBHELPER_HPP
#define LIBHELPER_HPP

#include <string>
#ifdef ANDROID_API
#include <android/asset_manager.h>
// #ifdef ANDROID_API
#include <android/log.h>
#include <vector>
#define TAG "MLLM"
#define LOGD(...) __android_log_print(ANDROID_LOG_DEBUG,TAG,__VA_ARGS__)
#define LOGE(...) __android_log_print(ANDROID_LOG_ERROR,TAG,__VA_ARGS__)
Expand All @@ -17,20 +18,12 @@
namespace mllm {
class Tokenizer;
class PreProcessor;
class FuyuPreProcess;
struct Context;
}


namespace mllm {
class Backend;
class Net;
class Executor;
class Module;
class Tensor;

enum PreDefinedModel {
LLAMA = 0,
FUYU,
Bert,
};

enum MLLMBackendType {
Expand All @@ -42,22 +35,29 @@ enum MLLMBackendType {
typedef std::function<void(std::string, bool)> callback_t;

class LibHelper {
Context *c = nullptr;
Net *net_ = nullptr;
Executor *executor_ = nullptr;
// Context *c = nullptr;
// Net *net_ = nullptr;
// Executor *executor_ = nullptr;
callback_t callback_ = [](std::string, bool) {
};
Tokenizer *tokenizer_ = nullptr;

std::shared_ptr<Tokenizer> tokenizer_;
PreProcessor *processor_;
std::shared_ptr<Module> module_;

// Tokenizer *tokenizer_ = nullptr;
unsigned int eos_id_ = 2;
PreDefinedModel model_ = PreDefinedModel::LLAMA;
bool is_first_run_cond_ = true;
int tokens_limit = 1000;
unsigned postProcessing(std::shared_ptr<Tensor> result, std::shared_ptr<Tensor> &out_result) const;
public:
bool setUp(const std::string &base_path, std::string weights_path, std::string vocab_path, PreDefinedModel model, MLLMBackendType backend_type = MLLMBackendType::CPU);
bool setUp(const std::string &base_path, std::string weights_path, std::string vocab_path, std::string merge_path, PreDefinedModel model, MLLMBackendType backend_type = MLLMBackendType::CPU);
void setCallback(callback_t callback);
void run(std::string &input_str, uint8_t *image, unsigned max_step, unsigned image_length) ;
std::vector<float> runForResult(std::string &input_str);
~LibHelper();
};
} // mllm
} // namespace mllm
#endif
#endif //LIBHELPER_HPP
// #endif //LIBHELPER_HPP
Loading