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

Support chat conversation for StaticLLMPipeline #580

Merged
merged 12 commits into from
Jul 17, 2024
Merged

Support chat conversation for StaticLLMPipeline #580

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
321e021
Support chat mode for LLM Static pipeline - naive
TolyaTalamanov Jul 4, 2024
670b1d6
Merge branch 'master' of https://github.com/openvinotoolkit/openvino.…
TolyaTalamanov Jul 5, 2024
6e8c2e0
Fix bug with first missing generated token
TolyaTalamanov Jul 5, 2024
271e3e1
Merge branch 'master' of https://github.com/openvinotoolkit/openvino.…
TolyaTalamanov Jul 15, 2024
361aa05
Small updates
TolyaTalamanov Jul 15, 2024
7d826a3
Extend chat_sample to accept device string
TolyaTalamanov Jul 15, 2024
389dcdf
Emphasize that device is optional in chat_sample
TolyaTalamanov Jul 16, 2024
758fd05
Revert chat_sample.cpp changes
TolyaTalamanov Jul 17, 2024
b1e53a4
Change offset integer type
TolyaTalamanov Jul 17, 2024
d20e865
Update src/cpp/src/llm_pipeline_static.cpp
TolyaTalamanov Jul 17, 2024
bb47c8c
Update src/cpp/src/llm_pipeline_static.cpp
TolyaTalamanov Jul 17, 2024
6e1a665
Update llm_pipeline_static.cpp
TolyaTalamanov Jul 17, 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
56 changes: 41 additions & 15 deletions src/cpp/src/llm_pipeline_static.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,8 @@
#include "text_callback_streamer.hpp"
#include "utils.hpp"

#include <openvino/pass/stateful_to_stateless.hpp>

namespace {

std::shared_ptr<ov::Model> add_slices_to_kvcache_inputs(const std::shared_ptr<ov::Model>& model) {
Expand Down Expand Up @@ -75,18 +77,15 @@ void reshape_to_static(std::shared_ptr<ov::Model> model,
model->reshape(new_shapes);
}

void fill_tensor(ov::Tensor tensor, int64_t fill_val) {
void fill_tensor(ov::Tensor tensor, int64_t fill_val, int32_t offset = 0) {
Wovchena marked this conversation as resolved.
Show resolved Hide resolved
int64_t* tensor_data = tensor.data<int64_t>();
std::fill(tensor_data, tensor_data + tensor.get_size(), fill_val);
std::fill(tensor_data + offset, tensor_data + tensor.get_size(), fill_val);
}

void copy_with_left_offset(const ov::Tensor& orig, ov::Tensor& padded) {
const auto orig_size = orig.get_size();
const auto padded_size = padded.get_size();
const auto kLeftOffset = padded_size - orig_size;
void copy_with_offset(const ov::Tensor& orig, const int32_t offset, ov::Tensor& padded) {
int64_t* orig_data = orig.data<int64_t>();
int64_t* padded_data = padded.data<int64_t>();
std::copy(orig_data, orig_data + orig_size, padded_data + kLeftOffset);
std::copy(orig_data, orig_data + orig.get_size(), padded_data + offset);
}

ov::AnyMap extract_config_or_empty(const ov::AnyMap& config, const std::string& config_name) {
Expand Down Expand Up @@ -127,6 +126,7 @@ StaticLLMPipeline::StaticLLMPipeline(
// (1) Read the template model - this will be kvcache model
auto kvcache_model = core.read_model(path / "openvino_model.xml");
// (2) TODO: Expose KV-cache input and output layers from kvcache model
TolyaTalamanov marked this conversation as resolved.
Show resolved Hide resolved
ov::pass::StatefulToStateless().run_on_model(kvcache_model);
// (3) Clone the model - this will be prefill
auto prefill_model = kvcache_model->clone();
prefill_model->set_friendly_name(kvcache_model->get_friendly_name() + "_prefill");
Expand Down Expand Up @@ -156,6 +156,14 @@ StaticLLMPipeline::StaticLLMPipeline(
) : StaticLLMPipeline(path, path.string(), device, config) {
}

void StaticLLMPipeline::start_chat() {
m_is_chat_conversation = true;
};
void StaticLLMPipeline::finish_chat() {
TolyaTalamanov marked this conversation as resolved.
Show resolved Hide resolved
m_is_chat_conversation = false;
m_history.clear();
};

void StaticLLMPipeline::prepare_for_new_conversation() {
fill_tensor(m_prefill_request.get_tensor("input_ids"), m_tokenizer.get_pad_token_id());
fill_tensor(m_prefill_request.get_tensor("position_ids"), 0u);
Expand All @@ -175,9 +183,25 @@ DecodedResults StaticLLMPipeline::generate(
}

OPENVINO_ASSERT(std::holds_alternative<std::string>(inputs));
auto tokenized_input = m_tokenizer.encode(std::get<std::string>(inputs));
auto& prompt = std::get<std::string>(inputs);

if (m_is_chat_conversation) {
m_history.push_back({{"role", "user"}, {"content", prompt}});
constexpr bool add_generation_prompt = true;
prompt = m_tokenizer.apply_chat_template(m_history, add_generation_prompt);
}

auto tokenized_input = m_tokenizer.encode(prompt);
auto encoded_results = generate(tokenized_input, config, streamer);
return {m_tokenizer.decode(encoded_results.tokens), encoded_results.scores};
DecodedResults decoded_results = {m_tokenizer.decode(encoded_results.tokens), encoded_results.scores};
pavel-esir marked this conversation as resolved.
Show resolved Hide resolved

if (m_is_chat_conversation) {
// Tail of chat template is missing in KV cache.
TolyaTalamanov marked this conversation as resolved.
Show resolved Hide resolved
// Find the tail to concatenate it with the next input prompt.
auto answer = decoded_results.texts[0];
m_history.push_back({{"role", "assistant"}, {"content", answer}});
}
return decoded_results;
}

EncodedResults StaticLLMPipeline::generate(
Expand Down Expand Up @@ -222,22 +246,25 @@ EncodedResults StaticLLMPipeline::generate(
ov::genai::EncodedResults results;
// NB: Only batch=1 is supported now
results.scores.resize(1u);
results.scores[0] = 0u;
TolyaTalamanov marked this conversation as resolved.
Show resolved Hide resolved
results.tokens.resize(1u);

// NB: Check if input prompt less than maximum size
// NB: Check if there is enough space in KV-cache to process input prompt
auto prompt_len = input_ids.get_size();
if (prompt_len > m_kvcache_desc.total_size) {
OPENVINO_THROW("Currently static pipeline only process up to " + std::to_string(m_kvcache_desc.total_size) + " tokens");
}

// NB: Reset tensors on every generate call - chat conversation isn't supported yet!
// NB: From the "generate" perspective, every prompt is treated as start of new conversation,
// but in case the real chat, prompt contains information about past conversation context
prepare_for_new_conversation();

auto padded_input_ids = m_prefill_request.get_tensor("input_ids");
copy_with_left_offset(input_ids, padded_input_ids);
const auto offset = padded_input_ids.get_size() - input_ids.get_size();
copy_with_offset(input_ids, offset, padded_input_ids);

auto padded_attention_mask = m_prefill_request.get_tensor("attention_mask");
copy_with_left_offset(attention_mask, padded_attention_mask);
fill_tensor(padded_attention_mask, 1u, offset);

auto padded_position_ids = m_prefill_request.get_tensor("position_ids");
auto* padded_pos_data = padded_position_ids.data<int64_t>();
Expand All @@ -248,13 +275,13 @@ EncodedResults StaticLLMPipeline::generate(
// NB: Now there are prompt_len tokens in KV-cache
m_kvcache_desc.num_stored_tokens += prompt_len;
int64_t last_token = utils::argmax(m_prefill_request.get_tensor("logits"), 0);
results.tokens[0].push_back(last_token);
TolyaTalamanov marked this conversation as resolved.
Show resolved Hide resolved
if (streamer_ptr && streamer_ptr->put(last_token)) {
return results;
}

padded_attention_mask.copy_to(m_kvcache_request.get_tensor("attention_mask"));


// Inputs: input_ids, attention_mask, position_ids, ...
// Outputs: logits, ...
const auto kStartInputKVCacheLayers = 3u;
Expand Down Expand Up @@ -286,7 +313,6 @@ EncodedResults StaticLLMPipeline::generate(

last_token = utils::argmax(m_kvcache_request.get_tensor("logits"), 0);
results.tokens[0].push_back(last_token);
results.scores[0] = 0u;
TolyaTalamanov marked this conversation as resolved.
Show resolved Hide resolved

if (streamer_ptr && streamer_ptr->put(last_token)) {
break;
Expand Down
11 changes: 5 additions & 6 deletions src/cpp/src/llm_pipeline_static.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,12 +35,8 @@ class StaticLLMPipeline final : public LLMPipelineImplBase {
StreamerVariant streamer
) override;

void start_chat() override {
OPENVINO_THROW("Currently chat conversation mode isn't supported");
};
void finish_chat() override {
OPENVINO_THROW("Currently chat conversation mode isn't supported");
};
void start_chat() override;
void finish_chat() override;

private:
void prepare_for_new_conversation();
Expand All @@ -54,6 +50,9 @@ class StaticLLMPipeline final : public LLMPipelineImplBase {
KVCacheDesc m_kvcache_desc;
ov::InferRequest m_kvcache_request;
ov::InferRequest m_prefill_request;

bool m_is_chat_conversation = false;
ChatHistory m_history;
};

} // namespace genai
Expand Down
Loading