Add Tokenizers #687
There are no files selected for viewing
|
There was a problem hiding this comment. Original problem is described here: https://github.com/openvinotoolkit/openvino.genai/blob/4c6c6cb9cf2b64584c797f2ff4ca0b7b658dabca/llm/llm.cpp#L85 Here's my attempt to simplify the reproducer, but it requires Debug build: #include <openvino/openvino.hpp>
int main(int argc, char* argv[]) try {
if (argc != 4) {
throw std::runtime_error(std::string{"Usage: "} + argv[0] + " <openvino_model.xml> <tokenizer.xml> '<prompt>'");
}
ov::Core core;
core.add_extension(USER_OV_EXTENSIONS_PATH); // USER_OV_EXTENSIONS_PATH is defined in CMakeLists.txt
ov::InferRequest tolenizer = core.compile_model(argv[2], "CPU").create_infer_request();
std::shared_ptr<ov::Model> model = core.read_model(argv[1]);
constexpr size_t BATCH_SIZE = 1;
std::map<std::string, ov::PartialShape> shapes = {
{"input_ids", ov::PartialShape{
BATCH_SIZE, {1, std::numeric_limits<ov::Dimension::value_type>::max()}
}},
{"attention_mask", ov::PartialShape{
BATCH_SIZE, {1, std::numeric_limits<ov::Dimension::value_type>::max()}
}}
};
for (const ov::Output<ov::Node>& input : model->inputs()) {
for (const std::string& name : input.get_names()) {
if (name.rfind("past_key_values", 0) == 0) {
ov::PartialShape shape = input.get_partial_shape();
shape[0] = BATCH_SIZE;
shapes.emplace(name, shape);
break;
}
}
}
model->reshape(shapes);
ov::preprocess::PrePostProcessor p3(model);
p3.input("input_ids").tensor().set_element_type(ov::element::i32); // cast to the type of tokenyzer's output
p3.input("attention_mask").tensor().set_element_type(ov::element::i32);
p3.input("input_ids").preprocess().convert_element_type(ov::element::i64);
p3.input("attention_mask").preprocess().convert_element_type(ov::element::i64);
model = p3.build();
ov::InferRequest ireq = core.compile_model(model, "CPU", {ov::cache_dir("llm-cache")}).create_infer_request();
for (const ov::Output<ov::Node>& input : model->inputs()) {
for (const std::string& name : input.get_names()) {
if (name.rfind("past_key_values", 0) == 0) {
ireq.get_tensor(input).set_shape(input.get_partial_shape().get_min_shape());
break;
}
}
}
float* logits;
size_t n_vocab;
int32_t out_token;
{
ov::Tensor inp{ov::element::i32, {1, 1}};
inp.data<int32_t>()[0] = 29528;
ireq.get_tensor("input_ids").set_shape({1, 1});
ireq.set_tensor("input_ids", inp);
ireq.get_tensor("attention_mask").set_shape({1, 1});
std::fill_n(ireq.get_tensor("attention_mask").data<int32_t>(), inp.get_size(), 1);
ireq.infer();
size_t n_vocab = ireq.get_tensor("logits").get_shape().back();
logits = ireq.get_tensor("logits").data<float>() + (inp.get_size() - 1) * n_vocab;
out_token = int32_t(std::max_element(logits, logits + n_vocab) - logits);
}
ireq.get_tensor("input_ids").set_shape({BATCH_SIZE, 1});
ireq.get_tensor("attention_mask").set_shape({BATCH_SIZE, 1});
ireq.get_tensor("attention_mask").data<int32_t>()[0] = 1;
constexpr int32_t SPECIAL_EOS_TOKEN = 2;
while (out_token != SPECIAL_EOS_TOKEN) {
std::cout << out_token << ' ' << std::flush;
for (const ov::Output<ov::Node>& input : model->inputs()) {
for (const std::string& name : input.get_names()) {
if (name.rfind("past_key_values", 0) == 0) {
ireq.set_tensor(input, ireq.get_tensor("present" + name.substr(15)));
break;
}
}
}
ireq.get_tensor("input_ids").data<int32_t>()[0] = 29528;
ireq.infer();
logits = ireq.get_tensor("logits").data<float>();
out_token = int32_t(std::max_element(logits, logits + n_vocab) - logits);
}
std::cout << '\n';
} catch (const std::exception& error) {
std::cerr << error.what() << '\n';
return 1;
} catch (...) {
std::cerr << "Non-exception object thrown\n";
return 1;
}
There was a problem hiding this comment. Using different Core object for llama fixes the problem, removing ov::preprocess::PrePostProcessor usage fixes the problem, storing CompiledModel instead of ireq of tokenizer fixes the problem. Moving |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
causes Segmentation fault. Removing
, {ov::cache_dir("llm-cache")}fixes the problem. StrangelyCACHE_DIRworks for python.