From 463c4b559b6a320018c94caf178f32f167053cf6 Mon Sep 17 00:00:00 2001 From: Hansong Zhang Date: Thu, 28 Aug 2025 17:03:29 -0700 Subject: [PATCH 1/4] Remove unused line --- extension/android/jni/jni_layer_llama.cpp | 1 - 1 file changed, 1 deletion(-) diff --git a/extension/android/jni/jni_layer_llama.cpp b/extension/android/jni/jni_layer_llama.cpp index a27b8194530..c64cb516e41 100644 --- a/extension/android/jni/jni_layer_llama.cpp +++ b/extension/android/jni/jni_layer_llama.cpp @@ -15,7 +15,6 @@ #include #include -#include #include #include #include From 63e407ea85b16c14dd932c3432f5f1d144d1ce71 Mon Sep 17 00:00:00 2001 From: Hansong Zhang Date: Thu, 28 Aug 2025 18:05:14 -0700 Subject: [PATCH 2/4] test --- extension/android/jni/jni_layer_llama.cpp | 65 ++++++----------------- 1 file changed, 16 insertions(+), 49 deletions(-) diff --git a/extension/android/jni/jni_layer_llama.cpp b/extension/android/jni/jni_layer_llama.cpp index c64cb516e41..1af5f56215a 100644 --- a/extension/android/jni/jni_layer_llama.cpp +++ b/extension/android/jni/jni_layer_llama.cpp @@ -13,10 +13,12 @@ #include #include -#include -#include #include #include +#include +#include +#include +#include #include #include #include @@ -119,7 +121,7 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { float temperature_ = 0.0f; int model_type_category_; std::unique_ptr runner_; - std::unique_ptr multi_modal_runner_; + std::unique_ptr multi_modal_runner_; public: constexpr static auto kJavaDescriptor = @@ -165,19 +167,16 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { model_type_category_ = model_type_category; if (model_type_category == MODEL_TYPE_CATEGORY_MULTIMODAL) { - multi_modal_runner_ = std::make_unique( + multi_modal_runner_ = llm::create_multimodal_runner( model_path->toStdString().c_str(), - tokenizer_path->toStdString().c_str(), - temperature); + llm::load_tokenizer(tokenizer_path->toStdString())); } else if (model_type_category == MODEL_TYPE_CATEGORY_LLM) { std::optional data_path_str = data_path ? std::optional{data_path->toStdString()} : std::nullopt; - // TODO(larryliu0820): Use the API in text_llm_runner.h to create the - // runner. - runner_ = example::create_llama_runner( + runner_ = executorch::extension::llm::create_text_llm_runner( model_path->toStdString(), - tokenizer_path->toStdString(), + llm::load_tokenizer(tokenizer_path->toStdString()), data_path_str); #if defined(EXECUTORCH_BUILD_QNN) } else if (model_type_category == MODEL_TYPE_QNN_LLAMA) { @@ -260,17 +259,7 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { jint eos) { facebook::jni::local_ref tuple_result = facebook::jni::make_long_array(2); - if (model_type_category_ != MODEL_TYPE_CATEGORY_MULTIMODAL) { - tuple_result->pin()[0] = static_cast(Error::NotSupported); - return tuple_result; - } - - auto&& result = multi_modal_runner_->prefill_prompt( - prompt->toStdString(), start_pos, bos, eos); - tuple_result->pin()[0] = static_cast(Error::Ok); - if (result.ok()) { - tuple_result->pin()[1] = static_cast(start_pos); - } + tuple_result->pin()[0] = static_cast(Error::NotSupported); return tuple_result; } @@ -287,28 +276,7 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { facebook::jni::local_ref tuple_result = facebook::jni::make_long_array(2); - if (model_type_category_ != MODEL_TYPE_CATEGORY_MULTIMODAL) { - tuple_result->pin()[0] = static_cast(Error::NotSupported); - return tuple_result; - } - - auto image_size = image->size(); - std::vector images; - if (image_size != 0) { - std::vector image_data_jint(image_size); - std::vector image_data(image_size); - image->getRegion(0, image_size, image_data_jint.data()); - for (int i = 0; i < image_size; i++) { - image_data[i] = image_data_jint[i]; - } - llm::Image image_runner{image_data, width, height, channels}; - images.push_back(image_runner); - } - // TODO(hsz): make start_pos a reference and update it here - jint result = static_cast( - multi_modal_runner_->prefill_images(images, start_pos)); - tuple_result->pin()[0] = result; - tuple_result->pin()[1] = static_cast(start_pos); + tuple_result->pin()[0] = static_cast(Error::NotSupported); return tuple_result; } @@ -319,13 +287,12 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { facebook::jni::alias_ref callback, jboolean echo) { if (model_type_category_ == MODEL_TYPE_CATEGORY_MULTIMODAL) { - return static_cast(multi_modal_runner_->generate_from_pos( - prompt->toStdString(), - seq_len, - start_pos, + + return static_cast(multi_modal_runner_->generate( + std::vector{llm::MultimodalInput{prompt->toStdString()}}, + llm::GenerationConfig {.echo = static_cast(echo), .seq_len = seq_len}, [callback](const std::string& result) { callback->onResult(result); }, - [callback](const llm::Stats& stats) { callback->onStats(stats); }, - echo)); + [callback](const llm::Stats& stats) { callback->onStats(stats); })); } else if (model_type_category_ == MODEL_TYPE_CATEGORY_LLM) { executorch::extension::llm::GenerationConfig config{ .echo = static_cast(echo), From 43d8e5edc09fa8f1e3b2f5566f5f279a4f194063 Mon Sep 17 00:00:00 2001 From: Hansong Zhang Date: Wed, 3 Sep 2025 16:08:17 -0700 Subject: [PATCH 3/4] Prefill --- extension/android/jni/jni_layer_llama.cpp | 52 ++++++++++++++++------ extension/llm/runner/multimodal_runner.cpp | 4 +- extension/llm/runner/multimodal_runner.h | 4 +- 3 files changed, 43 insertions(+), 17 deletions(-) diff --git a/extension/android/jni/jni_layer_llama.cpp b/extension/android/jni/jni_layer_llama.cpp index 1af5f56215a..0c3550f151a 100644 --- a/extension/android/jni/jni_layer_llama.cpp +++ b/extension/android/jni/jni_layer_llama.cpp @@ -121,7 +121,9 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { float temperature_ = 0.0f; int model_type_category_; std::unique_ptr runner_; - std::unique_ptr multi_modal_runner_; + std::unique_ptr + multi_modal_runner_; + std::vector prefill_inputs_; public: constexpr static auto kJavaDescriptor = @@ -215,6 +217,9 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { facebook::jni::alias_ref callback, jboolean echo) { if (model_type_category_ == MODEL_TYPE_CATEGORY_MULTIMODAL) { + std::vector inputs = prefill_inputs_; + prefill_inputs_.clear(); + inputs.emplace_back(llm::MultimodalInput{prompt->toStdString()}); auto image_size = image->size(); std::vector images; if (image_size != 0) { @@ -225,15 +230,18 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { image_data[i] = image_data_jint[i]; } llm::Image image_runner{image_data, width, height, channels}; - images.push_back(image_runner); + inputs.emplace_back(llm::MultimodalInput{std::move(image_runner)}); } + executorch::extension::llm::GenerationConfig config{ + .echo = static_cast(echo), + .seq_len = seq_len, + .temperature = temperature_, + }; multi_modal_runner_->generate( - std::move(images), - prompt->toStdString(), - seq_len, - [callback](std::string result) { callback->onResult(result); }, - [callback](const llm::Stats& result) { callback->onStats(result); }, - echo); + std::move(inputs), + config, + [callback](const std::string& result) { callback->onResult(result); }, + [callback](const llm::Stats& result) { callback->onStats(result); }); } else if (model_type_category_ == MODEL_TYPE_CATEGORY_LLM) { executorch::extension::llm::GenerationConfig config{ .echo = static_cast(echo), @@ -257,9 +265,10 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { jlong start_pos, jint bos, jint eos) { + prefill_inputs_.emplace_back(llm::MultimodalInput{prompt->toStdString()}); facebook::jni::local_ref tuple_result = facebook::jni::make_long_array(2); - tuple_result->pin()[0] = static_cast(Error::NotSupported); + tuple_result->pin()[0] = static_cast(Error::Ok); return tuple_result; } @@ -273,10 +282,24 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { jint height, jint channels, jlong start_pos) { + std::vector images; + auto image_size = image->size(); + if (image_size != 0) { + std::vector image_data_jint(image_size); + std::vector image_data(image_size); + image->getRegion(0, image_size, image_data_jint.data()); + for (int i = 0; i < image_size; i++) { + image_data[i] = image_data_jint[i]; + } + llm::Image image_runner{image_data, width, height, channels}; + prefill_inputs_.emplace_back( + llm::MultimodalInput{std::move(image_runner)}); + } + facebook::jni::local_ref tuple_result = facebook::jni::make_long_array(2); - tuple_result->pin()[0] = static_cast(Error::NotSupported); + tuple_result->pin()[0] = static_cast(Error::Ok); return tuple_result; } @@ -287,10 +310,13 @@ class ExecuTorchLlmJni : public facebook::jni::HybridClass { facebook::jni::alias_ref callback, jboolean echo) { if (model_type_category_ == MODEL_TYPE_CATEGORY_MULTIMODAL) { - + std::vector inputs = prefill_inputs_; + prefill_inputs_.clear(); + inputs.emplace_back(llm::MultimodalInput{prompt->toStdString()}); return static_cast(multi_modal_runner_->generate( - std::vector{llm::MultimodalInput{prompt->toStdString()}}, - llm::GenerationConfig {.echo = static_cast(echo), .seq_len = seq_len}, + inputs, + llm::GenerationConfig{ + .echo = static_cast(echo), .seq_len = seq_len}, [callback](const std::string& result) { callback->onResult(result); }, [callback](const llm::Stats& stats) { callback->onStats(stats); })); } else if (model_type_category_ == MODEL_TYPE_CATEGORY_LLM) { diff --git a/extension/llm/runner/multimodal_runner.cpp b/extension/llm/runner/multimodal_runner.cpp index 2bc658692da..f6b29d42c09 100644 --- a/extension/llm/runner/multimodal_runner.cpp +++ b/extension/llm/runner/multimodal_runner.cpp @@ -65,8 +65,8 @@ Error MultimodalRunner::load() { Error MultimodalRunner::generate( const std::vector& inputs, const GenerationConfig& config, - std::function& token_callback, - std::function& stats_callback) { + std::function token_callback, + std::function stats_callback) { if (inputs.empty()) { ET_LOG(Error, "MultimodalInput vector cannot be empty"); return Error::InvalidArgument; diff --git a/extension/llm/runner/multimodal_runner.h b/extension/llm/runner/multimodal_runner.h index 186a5bf70e4..fc87a9ab18a 100644 --- a/extension/llm/runner/multimodal_runner.h +++ b/extension/llm/runner/multimodal_runner.h @@ -116,8 +116,8 @@ class ET_EXPERIMENTAL MultimodalRunner { virtual ::executorch::runtime::Error generate( const std::vector& inputs, const GenerationConfig& config, - std::function& token_callback, - std::function& stats_callback); + std::function token_callback, + std::function stats_callback); inline void stop() { text_token_generator_->stop(); From b84bf31be971079abcd53bb58f7d8d015f1d4a3d Mon Sep 17 00:00:00 2001 From: Hansong Zhang Date: Fri, 5 Sep 2025 15:36:26 -0700 Subject: [PATCH 4/4] debug --- extension/llm/runner/llm_runner_helper.cpp | 4 ++-- extension/llm/runner/multimodal_prefiller.cpp | 7 +++++++ extension/llm/runner/multimodal_runner.cpp | 8 ++++++-- 3 files changed, 15 insertions(+), 4 deletions(-) diff --git a/extension/llm/runner/llm_runner_helper.cpp b/extension/llm/runner/llm_runner_helper.cpp index f12de5f1d87..c5ce85fcf9b 100644 --- a/extension/llm/runner/llm_runner_helper.cpp +++ b/extension/llm/runner/llm_runner_helper.cpp @@ -111,9 +111,9 @@ get_llm_metadata(tokenizers::Tokenizer* tokenizer, Module* module) { if (!method_names.count(llm::kMaxSeqLen)) { ET_LOG( Error, - "Required metadata method %s not found in model", + "Required metadata method %s not found in model. Bypass", llm::kMaxSeqLen); - return ::executorch::runtime::Error::InvalidArgument; + // return ::executorch::runtime::Error::InvalidArgument; } for (auto& pair : metadata) { diff --git a/extension/llm/runner/multimodal_prefiller.cpp b/extension/llm/runner/multimodal_prefiller.cpp index 1d9a0c8fdfc..60039708d01 100644 --- a/extension/llm/runner/multimodal_prefiller.cpp +++ b/extension/llm/runner/multimodal_prefiller.cpp @@ -37,6 +37,8 @@ MultimodalPrefiller::MultimodalPrefiller( Result MultimodalPrefiller::prefill( const MultimodalInput& input, int64_t& start_pos) { + ET_LOG(Error, "Here 000000000000000000000000000000000000000000000000000000000000000000 %d", (int) (start_pos)); + ET_LOG(Error, "Here 000000000000000000000000000000000000000000000000000000000000000000 %s", input.get_text().c_str()); // 1. Run encoder model. ::executorch::runtime::EValue encoder_output; if (input.is_image()) { @@ -73,12 +75,14 @@ Result MultimodalPrefiller::prefill( auto& text = input.get_text(); std::vector tokens = ET_UNWRAP_TOKENIZER(tokenizer_->encode(text)); + ET_LOG(Error, "Here 666666666666666666666666666666666666666666666666666666666666666666666666666666"); auto text_tensor = executorch::extension::from_blob( tokens.data(), {1, static_cast(tokens.size())}, ::executorch::aten::ScalarType::Long); + ET_LOG(Error, "Here 555555555555555555555555555555555555555555555555555555555555555555555555555"); // Run text encoder (token embeddings) auto token_embedding_outputs = ET_UNWRAP(module_->execute(kTokenEmbeddingMethod, text_tensor)); @@ -89,6 +93,7 @@ Result MultimodalPrefiller::prefill( // For any other input types, return error return ::executorch::runtime::Error::NotSupported; } + ET_LOG(Error, "Here 000000000000000000000000000000000000000000000000000000000000000000000000000"); // 2. Run decoder model for prefill. // `cache_position` goes from start_pos to start_pos + encoder_output.size(1). @@ -107,6 +112,7 @@ Result MultimodalPrefiller::prefill( cache_positions.data(), {static_cast(seq_len)}, executorch::aten::ScalarType::Long); + ET_LOG(Error, "Here 111111111111111111111111111111111111111111111111111111111111111111111111111111111111111"); auto prefill_result = module_->execute( kTextModelMethod, {cache_position_tensor, encoder_output}); if (prefill_result.error() != ::executorch::runtime::Error::Ok) { @@ -121,6 +127,7 @@ Result MultimodalPrefiller::prefill( return ::executorch::runtime::Error::InvalidState; } auto outputs_res = prefill_outputs[0].toTensor(); + ET_LOG(Error, "Here 222222222222222222222222222222222222222222222222222222222222222222222222222222"); // Update start_pos, tracking the current cache position. start_pos += seq_len; diff --git a/extension/llm/runner/multimodal_runner.cpp b/extension/llm/runner/multimodal_runner.cpp index f6b29d42c09..2c82f4ee7ea 100644 --- a/extension/llm/runner/multimodal_runner.cpp +++ b/extension/llm/runner/multimodal_runner.cpp @@ -57,9 +57,9 @@ Error MultimodalRunner::load() { // Don't print with the same priority during warmup #define RUNNER_ET_LOG(warmup, format, ...) \ if (warmup) { \ - ET_LOG(Debug, format, __VA_ARGS__); \ + ET_LOG(Error, format, __VA_ARGS__); \ } else { \ - ET_LOG(Info, format, __VA_ARGS__); \ + ET_LOG(Error, format, __VA_ARGS__); \ } Error MultimodalRunner::generate( @@ -104,9 +104,12 @@ Error MultimodalRunner::generate( uint64_t prefill_next_token = 0; // Process multimodal inputs in order + ET_LOG(Error, "0000000000000000000000000000000000000000000000000000SIZE%d", inputs.size()); for (const MultimodalInput& input : inputs) { + ET_LOG(Error, "00000000000000000000000000000000123321451345143100"); prefill_next_token = ET_UNWRAP(multimodal_prefiller_->prefill(input, pos_)); } + ET_LOG(Error, "1111111111111111111111111111111111111111111111111111"); stats_->first_token_ms = time_in_ms(); stats_->prompt_eval_end_ms = time_in_ms(); @@ -114,6 +117,7 @@ Error MultimodalRunner::generate( wrapped_callback(ET_UNWRAP_TOKENIZER( tokenizer_->decode(prefill_next_token, prefill_next_token))); + ET_LOG(Info, "2222222222222222222222222222222222222222222222222222"); RUNNER_ET_LOG( config.warming,