From 0c7b3595b9e5ad2355818e259f06b0dc3f0065b3 Mon Sep 17 00:00:00 2001 From: Xuan Son Nguyen Date: Sat, 15 Jun 2024 18:53:40 +0200 Subject: [PATCH] Add `cvector-generator` example (#7514) * add control-vector-generator * calc diff * add comments * proof-of-concept stdlib implementation Implements PCA and file writing using mostly standard libraries. The output is recognized as a functional control vector, but outputs gibberish. * param parsing, refactor, comments Added basic command-line parameters for outfile and one each positive/negative prompt. Refactored some messy code in PCA computation and GGUF exporting. Left a bunch of comments regarding further work needed. * example template completions Implements an example template set built from the positive/negative prompts like the control vector Python implementation. * add multi prompts, multi-thread for PCA * fix mem error * add debugs * fix matrix transpose multiplication you have got to be kidding me * preliminary template/multiprompt support model is running out of context and that ought to be fixed (segfaulting) but other than that it looks goodish * fix zero output & param parsing, functional templating fixed a bug where the output file had no tensor data/was all zero fixed a bug where single hyphen flags were not being correctly parsed implements creation of templated prompts from input (still need to adapt based on model) * fix square_diff matmul index range and CRLF->LF line endings fixed a logic error where square_diff would not multiply all rows fixed a formatting error where the provided completions.txt had CRLF line endings * add command-line args for num threads, num completions file lines, always reload model refactored a few things and did what the commit message says on the tin * code aestheticization * fix compiler warnings * in-series multithreading for prompt embedding? added commented-out code to attempt to start implementing mutlithreading for embedding in main * remove unnecessary multithreading * interim fix memory leak * translated everything but PCA (I think) * tentatively translate the rest * fix ggml errors and make new ones at least it compiles and runs * fix cb_eval * temporary commit while I move dev environments it finally outputs a functioning control vector - "functioning" in the sense that it can be loaded and it clearly has the right idea, but makes the model incoherent * update debug statements * pre-tokenize so we can allocate correct memory to ctx_diffs_wrapped * update comments * (wip) refactor * clean up PCA ggml implementation * fix shape of v_diff_original * add n_batch for pca * working version * remember to copy back the last_eigenvector * fix n_completions * bring back n_completions * default n_pca_batch to 20 * fix macos build * add to makefile all targets * use ggml_format_name * add readme * fix .editorconfig * use ggml_backend_tensor_copy * attemp to fix compile problem on mac * fix compile warn * reuse allocr * move param parser to common * better error handling * clean up a bit * add print_usage * shorten help msg * beautify help msg * escape prompt by default * change compile target to llama-cvector-generator * typo * disable GPU for PCA * code style --------- Co-authored-by: Christian Zhou-Zheng --- .editorconfig | 3 + Makefile | 5 + common/common.cpp | 60 ++ common/common.h | 9 + examples/CMakeLists.txt | 1 + examples/cvector-generator/CMakeLists.txt | 5 + examples/cvector-generator/README.md | 34 + examples/cvector-generator/completions.txt | 582 ++++++++++++++++++ .../cvector-generator/cvector-generator.cpp | 499 +++++++++++++++ examples/cvector-generator/negative.txt | 1 + examples/cvector-generator/pca.hpp | 322 ++++++++++ examples/cvector-generator/positive.txt | 1 + 12 files changed, 1522 insertions(+) create mode 100644 examples/cvector-generator/CMakeLists.txt create mode 100644 examples/cvector-generator/README.md create mode 100644 examples/cvector-generator/completions.txt create mode 100644 examples/cvector-generator/cvector-generator.cpp create mode 100644 examples/cvector-generator/negative.txt create mode 100644 examples/cvector-generator/pca.hpp create mode 100644 examples/cvector-generator/positive.txt diff --git a/.editorconfig b/.editorconfig index 16d16b3b5..bd525e13f 100644 --- a/.editorconfig +++ b/.editorconfig @@ -26,3 +26,6 @@ indent_size = 2 [examples/llama.swiftui/llama.swiftui.xcodeproj/*] indent_style = tab + +[examples/cvector-generator/*.txt] +insert_final_newline = unset diff --git a/Makefile b/Makefile index 744fe5739..5ab3481fb 100644 --- a/Makefile +++ b/Makefile @@ -38,6 +38,7 @@ BUILD_TARGETS = \ llama-tokenize \ llama-train-text-from-scratch \ llama-vdot \ + llama-cvector-generator \ tests/test-c.o # Binaries only useful for tests @@ -922,6 +923,10 @@ llama-eval-callback: examples/eval-callback/eval-callback.cpp ggml.o llama.o $(C $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) +llama-cvector-generator: examples/cvector-generator/cvector-generator.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS) + $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) + $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) + llama-train-text-from-scratch: examples/train-text-from-scratch/train-text-from-scratch.cpp ggml.o llama.o $(COMMON_DEPS) train.o $(OBJS) $(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<) $(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS) diff --git a/common/common.cpp b/common/common.cpp index 1591790e6..73ff0e85b 100644 --- a/common/common.cpp +++ b/common/common.cpp @@ -1576,6 +1576,7 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa return true; } params.out_file = argv[i]; + params.cvector_outfile = argv[i]; return true; } if (arg == "-ofreq" || arg == "--output-frequency") { @@ -1610,6 +1611,55 @@ bool gpt_params_find_arg(int argc, char ** argv, const std::string & arg, gpt_pa params.i_chunk = std::stoi(argv[i]); return true; } + // cvector params + if (arg == "--completions-file") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.cvector_completions_file = argv[i]; + return true; + } + if (arg == "--positive-file") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.cvector_positive_file = argv[i]; + return true; + } + if (arg == "--negative-file") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.cvector_negative_file = argv[i]; + return true; + } + if (arg == "--completions") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.n_completions = std::stoi(argv[i]); + return true; + } + if (arg == "--pca-batch") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.n_pca_batch = std::stoi(argv[i]); + return true; + } + if (arg == "--pca-iter") { + if (++i >= argc) { + invalid_param = true; + return true; + } + params.n_pca_iterations = std::stoi(argv[i]); + return true; + } #ifndef LOG_DISABLE_LOGS // Parse args for logging parameters if (log_param_single_parse(argv[i])) { @@ -1931,6 +1981,16 @@ void gpt_params_print_usage(int /*argc*/, char ** argv, const gpt_params & param options.push_back({ "logging", " --log-append", "Don't truncate the old log file." }); #endif // LOG_DISABLE_LOGS + options.push_back({ "cvector" }); + options.push_back({ "cvector", "-o, --output FNAME", "output file (default: '%s')", params.cvector_outfile.c_str() }); + options.push_back({ "cvector", " --positive-file FNAME", "positive prompts file, one prompt per line (default: '%s')", params.cvector_positive_file.c_str() }); + options.push_back({ "cvector", " --negative-file FNAME", "negative prompts file, one prompt per line (default: '%s')", params.cvector_negative_file.c_str() }); + options.push_back({ "cvector", " --completions-file FNAME", + "completions file (default: '%s')", params.cvector_completions_file.c_str() }); + options.push_back({ "cvector", " --completions N", "number of lines of completions file to use (default: %d)", params.n_completions }); + options.push_back({ "cvector", " --batch-pca N", "batch size used for PCA. Larger batch runs faster, but uses more memory (default: %d)", params.n_pca_batch }); + options.push_back({ "cvector", " --iter-pca N", "number of iterations used for PCA (default: %d)", params.n_pca_iterations }); + printf("usage: %s [options]\n", argv[0]); for (const auto & o : options) { diff --git a/common/common.h b/common/common.h index 2345d855e..58ed72f43 100644 --- a/common/common.h +++ b/common/common.h @@ -232,6 +232,15 @@ struct gpt_params { bool process_output = false; // collect data for the output tensor bool compute_ppl = true; // whether to compute perplexity + + // cvector-generator params + int n_completions = 64; + int n_pca_batch = 20; + int n_pca_iterations = 1000; + std::string cvector_outfile = "control_vector.gguf"; + std::string cvector_completions_file = "examples/cvector-generator/completions.txt"; + std::string cvector_positive_file = "examples/cvector-generator/positive.txt"; + std::string cvector_negative_file = "examples/cvector-generator/negative.txt"; }; void gpt_params_handle_model_default(gpt_params & params); diff --git a/examples/CMakeLists.txt b/examples/CMakeLists.txt index d6ce35f4c..0b51c44c0 100644 --- a/examples/CMakeLists.txt +++ b/examples/CMakeLists.txt @@ -12,6 +12,7 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR}) if (EMSCRIPTEN) else() + add_subdirectory(cvector-generator) add_subdirectory(baby-llama) add_subdirectory(batched-bench) add_subdirectory(batched) diff --git a/examples/cvector-generator/CMakeLists.txt b/examples/cvector-generator/CMakeLists.txt new file mode 100644 index 000000000..0a559d60c --- /dev/null +++ b/examples/cvector-generator/CMakeLists.txt @@ -0,0 +1,5 @@ +set(TARGET llama-cvector-generator) +add_executable(${TARGET} cvector-generator.cpp pca.hpp) +install(TARGETS ${TARGET} RUNTIME) +target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT}) +target_compile_features(${TARGET} PRIVATE cxx_std_11) diff --git a/examples/cvector-generator/README.md b/examples/cvector-generator/README.md new file mode 100644 index 000000000..7b0e79c1f --- /dev/null +++ b/examples/cvector-generator/README.md @@ -0,0 +1,34 @@ +# cvector-generator + +This example demonstrates how to generate a control vector using gguf models. + +Related PRs: +- [Add support for control vectors](https://github.com/ggerganov/llama.cpp/pull/5970) +- (Issue) [Generate control vector using llama.cpp](https://github.com/ggerganov/llama.cpp/issues/6880) +- [Add cvector-generator example](https://github.com/ggerganov/llama.cpp/pull/7514) + +## Examples + +```sh +# CPU only +./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf + +# With GPU +./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 + +# With advanced options +./cvector-generator -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 --completions 128 --pca-iter 2000 --batch-pca 100 + +# To see help message +./cvector-generator -h +# Then, have a look at "cvector" section +``` + +## Tips and tricks + +If you have multiple lines per prompt, you can escape the newline character (change it to `\n`). For example: + +``` +<|im_start|>system\nAct like a person who is extremely happy.<|im_end|> +<|im_start|>system\nYou are in a very good mood today<|im_end|> +``` diff --git a/examples/cvector-generator/completions.txt b/examples/cvector-generator/completions.txt new file mode 100644 index 000000000..abc45ffd8 --- /dev/null +++ b/examples/cvector-generator/completions.txt @@ -0,0 +1,582 @@ + +That game +I can see +Hmm, this +I can relate to +Who is +I understand the +Ugh, +What the hell was +Hey, did anyone +Although +Thank you for choosing +What are you +Oh w +How dare you open +It was my pleasure +I'm hon +I appreciate that you +Are you k +Whoever left this +It's always +Ew, +Hey, I l +Hello? Is someone +I understand that +That poem +Aww, poor +Hey, it +Alright, who +I didn't +Well, life +The document +Oh no, this +I'm concerned +Hello, this is +This art +Hmm, this drink +Hi there! +It seems +Is +Good +I can't +Ex +Who are +I can see that +Wow, +Today is a +Hey friend +Sometimes friends +Oh, this old +The weather outside +This place is sur +I appreciate your input +Thank you for the +Look at +I'm disappoint +To my +How dare you +That's an +This piece of art +Eww +This park is +This is incredible +Oh no, someone +Exc +Well, it' +I warned +Hey, I understand +Hey, I saw +How dare you go +What the he +Hey +It's +Hello? Hello? +It +Oh no! +This is the perfect +Good morning, +Oh no, there +It's so +Yeah +Uh, +Hello everyone +Who turned off +The weather +Who' +Hey, this +Wait, +Eww, gross +Excuse +It seems like you +Thank you so +What happened? +Oh my g +I am deeply sad +I war +Okay, let' +Hey, that +That was a beautiful +Oh no! That +What happened +Hey there +The artist' +What?! +Hey, it' +I am disappoint +It seems like +Oh no! The +This park is a +If you +Yes! I did +It sounds +What +Who is it +Hmm, that +That's strange +Yeah, that was +That's interesting +This park +What the hell +Who is that +I feel like my +Oh well +What the hell is +Hello? Hello +To my dearest +Bless you!\" +Thank you for +Oh, looks like +Can you please +This place is +Eww, what +Bless you +Is everything +Hey, I just +Whoever left these +Well, that' +I feel +Hey, do you +It's sad +Oh no, it +Hey, that' +Oh my god, +Thank you, +Hello little one, +I apolog +Hey team, I +How dare you read +Who is this and +Whoever left +Hi there! W +A +If you have +I was +U +Bless +Well, this +Oh, I' +It's a +Eww, +Is everything okay? +Oh, I +Hello, can you +Al +That was a great +What are +I understand that not +Oh no, not +Who is it?\" +Hey, can we +Whoever is taking +I would love to +Hey, I noticed +Hey, could +I understand that there +Hello? +D +Oh man, I +Thank you so much +Oh no, my +Dear [Name +Uh +I remember +Hey, who +Well, it +Are you +I understand that it +Hey, is +I would +Who is this +Excuse me +Alright +I am thrilled +Sometimes friends have +Who the +It's interesting +I would love +E +Hello? Is anyone +Well, this is +This place +Well, +I warned you +Hey, watch where +Oh my +That' +Sometimes friends have different +I understand that everyone +What? +What do these notes +I can relate +I'm not +I understand +To my dear +Guys +Well +Hey, I appreciate +Wow, what +Dear +That melody +Who the hell +Today is +Hello little +Wow, look +That's great +Love is never wrong +I'm having +Whoa, did +Ugh +Can you please provide +I miss you, +I feel uncom +I know +Ugh, this +Hey, watch +Oh great, a +I didn +Okay +That game of char +Oh +I appreciate +Who's there +I am so +Oh great, someone +Hey, could you +I remember wondering +Wait, what? +What do +Hello? Can +Hey there, +That game of +This is incred +Oh my gosh +Oh great, f +I appreciate your +It sounds like +What the heck +Okay, I understand +Ew +I understand that this +Uh, hi +Hi everyone! +What the hell? +Thank you for your +Oh no, the +Wow, I +Who turned +Dear [ +Whoever +This is a +Whoa, he +What in the world +Although the physical +Hello, who is +That's amaz +Hey, I know +Okay, that +Hi everyone +Hey, is everything +I understand your fr +Oh no, poor +Oh, look +Good morning +Ew, gross +Oh no, did +Look at the family +Hey team +Yes! +Hey, can I +Okay, that' +It's great +Love is +Hey, what +Good morning, world +Who is it? +That poem really reson +I +That's +I understand the task +Gu +Hello? Who' +This postcard is +Whoa, +Oh, that +I understand that I +Whoever is +Hello? Who is +I'm really +Wow, this +Can +This artwork really +This is a shame +I miss you too +Who are you? +Today is a difficult +Hey, just +Are you okay +I am +Hi, +Wow, that +Hey there! Can +Okay, stay +Oh great, just +Yeah, +Hello? Can you +Oh, looks +Thank you for sharing +I'm glad +Hey, is that +Hmm +It was my +It sounds like you +Wow, your +I was promised certain +That was such a +Thank +Excuse you +That was +Hey team, +I feel un +It was +What' +Hey friend, I +How +Saying goodbye +That +It's heart +How dare +Oh, +Hello, may +What's this +Thank you for recogn +Aww, that +Oh, I remember +Hmm, that' +I miss +I know this +Wait +Is everything okay +Who is that person +Wow, you +Oh great +I'm sad +Wow, the +I am very disappoint +Who turned off the +I understand that things +I'm very +Hi +That's very +Okay, I +Oh no, +Wow, there +What's wrong +I apologize for +Hey, I +Can I help you +Oh, I didn +Alright, +Oh wow, +Oh my goodness +I know this event +What in the +Saying +Yeah, that +Guys, I +Hey, this v +This post +Are +Hey, can +Hello? Is +I can only imagine +Oh, that sounds +Hey, is anyone +I am disappointed +Hello, +Hey everyone, I +That was such +It's okay +The artist +Whoa +I understand that mistakes +Can I help +Who +Hi everyone! I +Hey, can you +Wow, how +Today +Oh no, I +Oh well, I +Well, that +This is the +Yes! I finally +Hey there little +Hello everyone! +Love is never +Look at the +This postcard +Oh great, +Can I +Hmm, this is +I understand your +Oh, look at +B +I'm so +Whoa, this +W +Oh, this +Sometimes +This piece of +What the +That was a +Hey, do +Oh no +Whoa, what +I feel like I +The documentary +Hello +Hello little one +I understand that my +Eww, that +Wow, an +Yes! Finally, +Although the physical location +Whoever is watching +That movie +I remember wondering about +Hey there, little +Who's +Hello, who +Hello everyone! Thank +Hello, can +That's too +Hey, just wanted +Hey there, I +Saying good +Hey there! +Who is there? +Oh my good +I am very +Oh no, what +Wow, thank +I was promised +Hi, is +Hey, I' +Guys, the +Oh no, that +Who is there +Hello, this +That movie really touched +If you have something +The documentary was +I'm starting +Are you kidd +That movie really +Hey everyone, +Thank you for considering +I didn' +Yes! I +Can you +Oh my god +Hey, whoever +That melody really +Thank you, little +Hello, may I +Look +Wow, we +It looks +What do these +Oh wow +I apologize +What are you all +It's such +It's clear +Hey, I was +Hey friend, +I can only +The weather outside is +Eww, this +I miss you +Wow +Aww, +Hi, is there +This artwork +Okay, +Oh well, +This +I' +Say +Hey there little gu +Hmm, +Whoa, who +I am thr +Oh man +Okay, stay calm +I'm happy +Oh, this cur +Oh man, +I'm sorry +Hello? Who +What?! That +This piece +Hey everyone +That's so +Are you okay? +What happened? Where +Hi there +The +Who the hell entered +I can +Guys, +What's +What in +It's important +I'm +I'm coming +It' +Yes! Finally +Wait, what +Wow, reading +I'm surprised +Hey, did +Hey, +Okay, let +I understand that you +Who the hell threw +Eww, who +Thank you for thinking +Who is this?\" +I am deeply +Thank you for including +Oh no, an +It looks like you +Aww +I'm confused +Wow, it +That poem really +Yes +Hey there, is +Hey, what' +Thank you for remember +To +This is +Thank you for making +I can' +That mel +Wow, they +I feel like +Although the +Who are you +Love +If +What the hell are +I am so sad +Oh, I found +Thank you +It looks like +Well, life is +I appreciate that +The artist's +Whoa, that +It's never \ No newline at end of file diff --git a/examples/cvector-generator/cvector-generator.cpp b/examples/cvector-generator/cvector-generator.cpp new file mode 100644 index 000000000..9941683db --- /dev/null +++ b/examples/cvector-generator/cvector-generator.cpp @@ -0,0 +1,499 @@ +#include "common.h" +#include "llama.h" +#include "ggml.h" +#include "pca.hpp" + +#ifdef GGML_USE_CUDA +#include "ggml-cuda.h" +#endif + +#ifdef GGML_USE_METAL +#include "ggml-metal.h" +#endif + +#include +#include +#include +#include +#include +#include +#include +#include + + +////////////////////////////////////////////////// +// utils + +template +static std::string tokens_to_str(llama_context * ctx, Iter begin, Iter end) { + std::string ret; + for (; begin != end; ++begin) { + ret += llama_token_to_piece(ctx, *begin); + } + + return ret; +} + +static void print_usage(int argc, char ** argv, const gpt_params & params) { + gpt_params_print_usage(argc, argv, params); + + printf("\nexample usage:\n"); + printf("\n CPU only: %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf\n", argv[0]); + printf("\n with GPU: %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99\n", argv[0]); + printf("\n advanced: %s -m ./dolphin-2.0-mistral-7b.Q4_K_M.gguf -ngl 99 --completions 128 --pca-iter 2000 --batch-pca 100\n", argv[0]); + printf("\n"); +} + +////////////////////////////////////////////////// + + +// cb_eval is reused for each pair of positive - negative prompt +struct callback_data { + ggml_context * ctx_ggml = nullptr; // holds v_pos, v_neg, v_diff_filtered + + int n_layers = 0; + int n_tokens = 0; + bool is_eval_pos = true; + + // each element of the vector correspond to one layer + std::vector v_pos; // vector of matrices of size [n_embd, n_tokens] + std::vector v_neg; // vector of matrices of size [n_embd, n_tokens] + std::vector v_diff_filtered; // vector of matrices of size [n_embd, n_nonzero_rows]. NOTE: n_nonzero_rows maybe different for each layer + + // save a tensor into either v_pos or v_neg (decided by is_eval_pos) + void save_tensor_for_layer(struct ggml_tensor * t) { + GGML_ASSERT(t->type == GGML_TYPE_F32); + + if (ctx_ggml == nullptr) { + // alloc a new ctx_ggml if needed + struct ggml_init_params params_ggml = { + /*.mem_size =*/ ggml_tensor_overhead() * n_layers * 3u, + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + ctx_ggml = ggml_init(params_ggml); + } + + // copy tensor data + auto n_bytes = ggml_nbytes(t); + struct ggml_tensor * t_layer = ggml_new_tensor_2d(ctx_ggml, t->type, t->ne[0], t->ne[1]); + t_layer->data = malloc(n_bytes); // TODO @ngxson : get rid of this malloc somehow + ggml_backend_tensor_get(t, t_layer->data, 0, n_bytes); + ggml_set_name(t_layer, ggml_get_name(t)); + //print_debug_tensor(t_layer); + + if (is_eval_pos) { + v_pos.push_back(t_layer); + } else { + v_neg.push_back(t_layer); + } + } + + // calculate diff (v_pos - v_neg) and place the result back to v_pos + // all zero rows in the diff tensor will also be removed + // NOTE: final layer is ignored. we only have (n_layers - 1) to process + std::vector calc_diff() { + for (float il = 0; il < v_pos.size(); il++) { + float * a = (float *) v_pos[il]->data; + float * b = (float *) v_neg[il]->data; + size_t n_elem = ggml_nelements(v_pos[il]); + for (size_t j = 0; j < n_elem; j++) { + a[j] -= b[j]; + } + //print_debug_tensor(v_pos[i]); + auto diff_filtered = filter_nonzero_rows(v_pos[il]); + v_diff_filtered.push_back(diff_filtered); + } + return v_diff_filtered; // for convinient, we return the result std::vector + } + + // delete zero rows from a given 2D tensor + struct ggml_tensor * filter_nonzero_rows(struct ggml_tensor * a) { + //printf("filter_nonzero_rows\n"); + auto is_row_all_zeros = [](struct ggml_tensor * t, int row, float eps) -> bool { + // check if given row containing all zero elements + int n_cols = t->ne[0]; // hint: should be equal to n_embd + for (int col = 0; col < n_cols; ++col) { + if (ggml_get_f32_nd(t, col, row, 0, 0) > eps) { + return false; + } + } + return true; + }; + std::vector rows_to_copy; // the idx of non-zero cols (to be copied to row of diff_filtered) + for (int i_row = 0; i_row < a->ne[1]; i_row++) { + if (!is_row_all_zeros(a, i_row, 1e-6)) { + rows_to_copy.push_back(i_row); + } + } + + // get "n_nonzero_rows" for the output "diff_filtered" + int n_nonzero_rows = rows_to_copy.size(); + //printf("n_nonzero_rows: %d\n", n_nonzero_rows); + int n_embd = a->ne[0]; + GGML_ASSERT(n_nonzero_rows > 0); + + // diff_filtered: [n_embd, n_nonzero_rows] + struct ggml_tensor * diff_filtered = ggml_new_tensor_2d( + ctx_ggml, GGML_TYPE_F32, n_embd, n_nonzero_rows); + ggml_format_name(diff_filtered, "diff_filtered_%s", a->name); + diff_filtered->data = malloc(ggml_nbytes(diff_filtered)); + + // copy non-zero rows + for (int dest_row = 0; dest_row < n_nonzero_rows; dest_row++) { + int src_row = rows_to_copy[dest_row]; + for (int i = 0; i < n_embd; i++) { + float src_elem = ggml_get_f32_nd(a, i, src_row, 0, 0); + ggml_set_f32_nd(diff_filtered, i, dest_row, 0, 0, src_elem); + } + } + + //print_debug_tensor(diff_filtered); + + return diff_filtered; + } + + // we don't implement destructor, because we want to reuse callback_data. we just want to free the tensors + void reset() { + for (auto ptr : v_pos) free(ptr->data); + for (auto ptr : v_neg) free(ptr->data); + for (auto ptr : v_diff_filtered) free(ptr->data); + v_pos.clear(); + v_neg.clear(); + v_diff_filtered.clear(); + if (ctx_ggml) { + ggml_free(ctx_ggml); + } + ctx_ggml = nullptr; + } +}; + +/** + * process_ctx is used to store the ggml context for pre-post processing the diff vectors + * in short, input => v_diff and output => v_final + */ +struct train_context { + ggml_context * ctx_ggml; + int n_embd; + int n_layers; + + /* pair of prompts to be used for generating final vector */ + std::vector positive_entries; + std::vector negative_entries; + + // each element of the vector correspond to one layer + // NOTE: the last layer is discard. therefore, we will have (n_layers - 1) elements here + // NOTE (2): v_diff is transposed from v_diff_tmp + std::vector v_diff; // vector of matrices of size [m, n_embd] where m ~ n_tokens * n_completions (v_diff contains no zero-rows) + std::vector v_final; // vector of vectors of size [n_embd] to be written to file + + // to easily re-alloc when concat v_diff, we temporary store v_diff in a vector instead of a tensor + // v_diff_tmp will get converted unto v_diff later on + std::vector> v_diff_tmp; + + train_context(int n_embd_, int n_layers_) { + n_embd = n_embd_; + n_layers = n_layers_; + struct ggml_init_params params_ggml = { + /*.mem_size =*/ ggml_tensor_overhead() * (n_layers - 1) * 2u, + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + ctx_ggml = ggml_init(params_ggml); + for (int il = 0; il < n_layers - 1; il++) { + std::vector empty; + v_diff_tmp.push_back(empty); + auto t = ggml_new_tensor_1d(ctx_ggml, GGML_TYPE_F32, n_embd); + t->data = malloc(ggml_nbytes(t)); // TODO: get rid of malloc if possible + v_final.push_back(t); + } + } + + // add new rows into existing tensor in v_diff_tmp + void concat_diff_tmp(const std::vector & diff_filtered) { + GGML_ASSERT((int) diff_filtered.size() == n_layers - 1); + for (int il = 0; il < n_layers - 1; il++) { + auto t = diff_filtered[il]; + auto & diff_tmp = v_diff_tmp[il]; + size_t curr_size = diff_tmp.size(); + diff_tmp.resize(curr_size + ggml_nbytes(t)); + memcpy(diff_tmp.data() + curr_size, t->data, ggml_nbytes(t)); + } + } + + // build the v_diff tensors from v_diff_tmp (v_diff need to be transposed) + // TODO @ngxson : maybe add option NOT to transpose v_diff; will be useful for "mean" method + void build_v_diff() { + printf("build_v_diff\n"); + for (int il = 0; il < n_layers - 1; il++) { + auto & diff_tmp = v_diff_tmp[il]; + int n_elem = diff_tmp.size() / sizeof(float); + GGML_ASSERT(n_elem % n_embd == 0); + int n_rows = n_elem / n_embd; + struct ggml_tensor * diff = ggml_new_tensor_2d(ctx_ggml, GGML_TYPE_F32, n_rows, n_embd); + ggml_set_name(diff, (std::string("diff_") + std::to_string(il)).c_str()); + // copy data & transpose + diff->data = malloc(ggml_nbytes(diff)); // TODO: get rid of this malloc if possible + float * arr = (float *) diff_tmp.data(); + for (int ir = 0; ir < n_rows; ++ir) { + for (int ic = 0; ic < n_embd; ++ic) { + float f = arr[ir*n_embd + ic]; + ggml_set_f32_nd(diff, ir, ic, 0, 0, f); + } + } + v_diff.push_back(diff); + print_debug_tensor(diff); + // free memory of diff_tmp + diff_tmp.resize(0); + } + } + + ~train_context() { + for (auto ptr : v_final) free(ptr->data); + for (auto ptr : v_diff) free(ptr->data); + // no need to free v_diff_tmp, since we didn't use malloc + ggml_free(ctx_ggml); + } +}; + +struct tokenized_prompt { + std::vector tokens_pos; + std::vector tokens_neg; + size_t max_seq_len; + + tokenized_prompt(llama_context * ctx, std::string pos, std::string neg) { + const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx)); + tokens_pos = ::llama_tokenize(ctx, pos, add_bos); + tokens_neg = ::llama_tokenize(ctx, neg, add_bos); + max_seq_len = std::max(tokens_pos.size(), tokens_neg.size()); + padding_seq(ctx, tokens_pos, max_seq_len); + padding_seq(ctx, tokens_neg, max_seq_len); + } + + void padding_seq(llama_context * ctx, std::vector & tokens, size_t len) { + // TODO: customize padding token + std::vector pad_tokens = ::llama_tokenize(ctx, " ", false); + llama_token pad_tok = pad_tokens.back(); + while (tokens.size() < len) { + tokens.push_back(pad_tok); + } + } +}; + +////////////////////////////////////////////////// + +template +static std::string to_string(const T & val) { + std::stringstream ss; + ss << val; + return ss.str(); +} + +static std::vector ctrlvec_load_prompt_file(std::string path, bool skip_empty_lines) { + std::vector output; + std::ifstream file(path); + if (!file.is_open()) { + fprintf(stderr, "error: unable to open file: %s\n", path.c_str()); + exit(1); + } + std::string line; + while (std::getline(file, line)) { + bool is_skip = skip_empty_lines && line.empty(); + if (!is_skip) { + string_process_escapes(line); + output.push_back(line); + } + } + file.close(); + return output; +} + +////////////////////////////////////////////////// + +static bool cb_eval(struct ggml_tensor * t, bool ask, void * user_data) { + auto * cb_data = (callback_data *) user_data; + static const char * l_out_name = "l_out"; + const bool is_l_out = strncmp(t->name, l_out_name, strlen(l_out_name)) == 0; + + if (ask) { + return is_l_out; + } + + if (!is_l_out || t->ne[1] != cb_data->n_tokens) { + return true; + } + + // save the tensor to current context + cb_data->save_tensor_for_layer(t); + return true; +} + +static bool get_hidden_layers(llama_context * ctx, std::vector & tokens) { + llama_kv_cache_clear(ctx); + if (llama_decode(ctx, llama_batch_get_one(tokens.data(), tokens.size(), 0, 0))) { + fprintf(stderr, "%s : failed to eval\n", __func__); + return false; + } + return true; +} + +static void export_gguf(const std::vector & v_ctrl, const std::string fname, const std::string model_hint) { + struct gguf_context * ctx = gguf_init_empty(); + + const std::string arch = "controlvector"; + gguf_set_val_str(ctx, "general.architecture", arch.c_str()); + gguf_set_val_str(ctx, (arch + ".model_hint").c_str(), model_hint.c_str()); + gguf_set_val_i32(ctx, (arch + ".layer_count").c_str(), v_ctrl.size()); + + for (size_t i = 0; i < v_ctrl.size(); ++i) { + gguf_add_tensor(ctx, v_ctrl[i]); + print_debug_tensor(v_ctrl[i]); + printf("Added tensor: %s\n", v_ctrl[i]->name); + } + + printf("%s: writing file...\n", __func__); + gguf_write_to_file(ctx, fname.c_str(), false); + printf("%s: wrote file '%s'\n", __func__, fname.c_str()); + gguf_free(ctx); +} + +/** + * Load prompt files and completion file. + * Then format each pair of prompt + completion to make an entry. + */ +static int prepare_entries(gpt_params & params, train_context & ctx_train) { + // load prompts + std::vector positive_prompts = ctrlvec_load_prompt_file(params.cvector_positive_file, true); + std::vector negative_prompts = ctrlvec_load_prompt_file(params.cvector_negative_file, true); + if (positive_prompts.size() != negative_prompts.size()) { + fprintf(stderr, "number of positive and negative prompts must be equal\n"); + return 1; + } + if (positive_prompts.empty()) { + fprintf(stderr, "must provide at least one prompt pair\n"); + return 1; + } + + // create templated prompts + std::vector completions = ctrlvec_load_prompt_file(params.cvector_completions_file, false); + auto format_template = [](std::string persona, std::string suffix) { + // entry in positive/negative.txt must already be formatted i.e. "[INST] Act as if you're extremely happy. [/INST]" + return persona + " " + suffix; + }; + for (size_t i = 0; i < positive_prompts.size(); ++i) { + for (int j = 0; j < std::min((int) completions.size(), params.n_completions); ++j) { + // TODO replicate the truncations done by the python implementation + ctx_train.positive_entries.push_back(format_template(positive_prompts[i], completions[j])); + ctx_train.negative_entries.push_back(format_template(negative_prompts[i], completions[j])); + } + } + return 0; +} + +int main(int argc, char ** argv) { + gpt_params params; + + if (!gpt_params_parse(argc, argv, params)) { + print_usage(argc, argv, params); + return 1; + } + + if (params.n_pca_iterations % params.n_pca_batch != 0) { + fprintf(stderr, "PCA iterations must by multiply of PCA batch size\n"); + return 1; + } + + + callback_data cb_data; + + // pass the callback to the backend scheduler + // it will be executed for each node during the graph computation + params.cb_eval = cb_eval; + params.cb_eval_user_data = &cb_data; + params.warmup = false; + + print_build_info(); + llama_backend_init(); + llama_numa_init(params.numa); + + // load the model to get hparams + llama_model * model; + llama_context * ctx; + std::tie(model, ctx) = llama_init_from_gpt_params(params); + + // int n_ctx = llama_n_ctx(ctx); + int n_layers = llama_n_layer(model); + int n_embd = llama_n_embd(model); + // get model hint param (a.k.a model arch name) + char model_hint[128]; + llama_model_meta_val_str(model, "general.architecture", model_hint, 128); + + // init train_context + train_context ctx_train(n_embd, n_layers); + + // load and prepare entries for training + prepare_entries(params, ctx_train); + + // we have to pretokenize everything because otherwise we don't know how much overhead to allocate ctx_diffs_wrapped + std::vector tokenized_prompts; + size_t n_total_tokens = 0; + for (size_t i = 0; i < ctx_train.positive_entries.size(); ++i) { + tokenized_prompt t(ctx, ctx_train.positive_entries[i], ctx_train.negative_entries[i]); + n_total_tokens += 2 * t.max_seq_len; + tokenized_prompts.push_back(std::move(t)); + } + + std::cout << "n_total_tokens: " << n_total_tokens << std::endl; + + for(size_t i = 0; i < ctx_train.positive_entries.size(); ++i) { + bool success = false; + tokenized_prompt t = tokenized_prompts[i]; + cb_data.n_layers = n_layers; + cb_data.n_tokens = t.max_seq_len; + + printf("Evaluating prompt[%d/%d]: \"%s\" - \"%s\" (%d tokens)\n", + (int) i+1, (int) ctx_train.positive_entries.size(), + tokens_to_str(ctx, t.tokens_pos.cbegin(), t.tokens_pos.cend()).c_str(), + tokens_to_str(ctx, t.tokens_neg.cbegin(), t.tokens_neg.cend()).c_str(), + (int) t.max_seq_len); + + cb_data.is_eval_pos = true; + success = get_hidden_layers(ctx, t.tokens_pos); + if (!success) break; + + cb_data.is_eval_pos = false; + success = get_hidden_layers(ctx, t.tokens_neg); + if (!success) break; + + // calculate diff and remove all zero rows + auto v_diff_filtered = cb_data.calc_diff(); + + // save & concat the filtered v_diff to ctx_train + ctx_train.concat_diff_tmp(v_diff_filtered); + + // reset for next iteration + cb_data.reset(); + } + + // done with the model, we can now free it to make gain some memory + printf("Done evaluate prompts, unload model...\n"); + llama_free(ctx); + llama_free_model(model); + + // prepare ctx_train for PCA + ctx_train.build_v_diff(); + + // run PCA + PCA::pca_params pca_params; + pca_params.n_threads = params.n_threads; + pca_params.n_batch = params.n_pca_batch; + pca_params.n_iterations = params.n_pca_iterations; + PCA::run_pca(pca_params, ctx_train.v_diff, ctx_train.v_final); + + // write output vectors to gguf + export_gguf(ctx_train.v_final, params.cvector_outfile, model_hint); + + llama_backend_free(); + + return 0; +} diff --git a/examples/cvector-generator/negative.txt b/examples/cvector-generator/negative.txt new file mode 100644 index 000000000..2ac3387f1 --- /dev/null +++ b/examples/cvector-generator/negative.txt @@ -0,0 +1 @@ +[INST] Act like a person who is extremely sad. [/INST] \ No newline at end of file diff --git a/examples/cvector-generator/pca.hpp b/examples/cvector-generator/pca.hpp new file mode 100644 index 000000000..8b95cec37 --- /dev/null +++ b/examples/cvector-generator/pca.hpp @@ -0,0 +1,322 @@ +#include "common.h" +#include "llama.h" +#include "ggml.h" + +#ifdef GGML_USE_CUDA +#include "ggml-cuda.h" +#endif + +#ifdef GGML_USE_METAL +#include "ggml-metal.h" +#endif + +#include +#include +#include +#include +#include +#include +#include +#include + +#define DEBUG_POS 5 + +static void print_debug_tensor(struct ggml_tensor * t, bool with_data = true) { + printf("%s: %s (%s): [%d, %d]\n", __func__, t->name, ggml_type_name(t->type), (int) t->ne[0], (int) t->ne[1]); + if (!with_data) return; + printf("%s: %s[0] = [", __func__, t->name); + for (size_t i = 0; i <= DEBUG_POS; i++) { + printf(" %f,", ggml_get_f32_nd(t, i, 0, 0, 0)); + } + printf(" ... ]\n"); +} + +namespace PCA { + +// input params for PCA computations +struct pca_params { + int n_threads = 1; + int n_batch = 20; // number of iterations do to in one batch. larger the batch, more memory is used + int n_iterations = 1000; + float tolerance = 1e-7; + + // for debugging + int i_layer = 0; + int n_layers = 0; +}; + +// result from each iteration +struct pca_result { + struct ggml_tensor * calculated_square = NULL; + std::vector eigenvectors; + std::vector distances; +}; + +struct pca_model { + ggml_backend_t backend = NULL; + ggml_backend_buffer_t buffer; + struct ggml_context * ctx; // context to compute graph on target device + struct ggml_context * ctx_host; // host context to store results + + // tensors on target device + struct ggml_tensor * dev_input; + struct ggml_tensor * dev_square; + struct ggml_tensor * dev_eigenvector; + + pca_model(struct ggml_tensor * t_input) { +// TODO: enable GPU support when support for GGML_OP_SQRT is added +// #ifdef GGML_USE_CUDA +// fprintf(stderr, "%s: using CUDA backend\n", __func__); +// backend = ggml_backend_cuda_init(0); // init device 0 +// if (!backend) { +// fprintf(stderr, "%s: ggml_backend_cuda_init() failed\n", __func__); +// } +// #endif + +// #ifdef GGML_USE_METAL +// fprintf(stderr, "%s: using Metal backend\n", __func__); +// backend = ggml_backend_metal_init(); +// if (!backend) { +// fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__); +// } +// #endif + + // if there aren't GPU Backends fallback to CPU backend + if (!backend) { + backend = ggml_backend_cpu_init(); + } + + const int num_tensors = 4; + struct ggml_init_params params { + /*.mem_size =*/ ggml_tensor_overhead() * num_tensors, + /*.mem_buffer =*/ NULL, + /*.no_alloc =*/ true, + }; + ctx = ggml_init(params); + + auto n_samples = t_input->ne[0]; + auto n_embd = t_input->ne[1]; + + dev_input = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_samples, n_embd); + dev_square = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_embd); + dev_eigenvector = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + ggml_set_name(dev_input, "dev_input"); + ggml_set_name(dev_square, "dev_square"); + ggml_set_name(dev_eigenvector, "dev_eigenvector"); + buffer = ggml_backend_alloc_ctx_tensors(ctx, backend); + ggml_backend_tensor_set(dev_input, t_input->data, 0, ggml_nbytes(t_input)); + + // initialize eigenvector to random normalized vector + { + std::vector random_vec(ggml_nelements(dev_eigenvector), 0.0); + std::default_random_engine generator(static_cast(std::time(0))); + std::uniform_real_distribution distribution(0.0, 1.0); + float sum_sqr = 0.0; // for normalizing random_vec + for (size_t i = 0; i < random_vec.size(); ++i) { + float f = distribution(generator); + sum_sqr += f * f; + random_vec[i] = f; + } + // normalize it + float random_vec_norm = std::sqrt(sum_sqr); + for (size_t i = 0; i < random_vec.size(); ++i) { + random_vec[i] /= random_vec_norm; + } + ggml_backend_tensor_set(dev_eigenvector, random_vec.data(), 0, ggml_nbytes(dev_eigenvector)); + } + } + + ~pca_model() { + ggml_free(ctx); + ggml_backend_buffer_free(buffer); + ggml_backend_free(backend); + } +}; + +static struct ggml_cgraph * build_graph_piter( + const struct pca_params & params, + const pca_model & model, + bool calc_square = false) { + GGML_ASSERT(params.n_batch > 0); + // TODO: buf_size must be able to scale with params.n_batch + static size_t buf_size = ggml_tensor_overhead()*GGML_DEFAULT_GRAPH_SIZE + ggml_graph_overhead(); + static std::vector buf(buf_size); + + struct ggml_init_params params0 = { + /*.mem_size =*/ buf_size, + /*.mem_buffer =*/ buf.data(), + /*.no_alloc =*/ true, // the tensors will be allocated later by ggml_allocr_alloc_graph() + }; + // create a temporally context to build the graph + struct ggml_context * ctx0 = ggml_init(params0); + struct ggml_cgraph * gf = ggml_new_graph(ctx0); + + // turn v_diff_original into square matrix if needed + struct ggml_tensor * tmp_square; + if (calc_square) { + tmp_square = ggml_mul_mat(ctx0, model.dev_input, model.dev_input); + ggml_set_name(tmp_square, "tmp_square"); + } + + struct ggml_tensor * b_tensor; + struct ggml_tensor * distance; + struct ggml_tensor * old_eigen = model.dev_eigenvector; + struct ggml_tensor * input_square = calc_square ? tmp_square : model.dev_square; + + for (int i = 0; i < params.n_batch; ++i) { + // b_tensor = square * eigenvector^T + b_tensor = ggml_mul_mat(ctx0, input_square, old_eigen); + ggml_set_name(b_tensor, "b_tensor"); + + // normalize + b_tensor = ggml_div_inplace(ctx0, + b_tensor, + ggml_sqrt_inplace(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, b_tensor))) + ); + ggml_format_name(b_tensor, "b_tensor_norm_%d", i); + + // calculate distance(new eigenvector - old eigenvector) + // we don't use ggml_sub because it may not be implemented on GPU backend + struct ggml_tensor * new_sub_old = ggml_add(ctx0, old_eigen, ggml_scale(ctx0, b_tensor, -1)); + distance = ggml_sqrt_inplace(ctx0, + ggml_sum_rows(ctx0, ggml_sqr_inplace(ctx0, new_sub_old))); + ggml_format_name(distance, "distance_%d", i); + + old_eigen = b_tensor; + + // build operations nodes + ggml_build_forward_expand(gf, distance); + } + + // delete the temporally context used to build the graph + ggml_free(ctx0); + return gf; +} + +static ggml_status compute_piter( + const struct pca_params & params, + const pca_model & model, + struct ggml_cgraph * gf, + ggml_gallocr_t allocr, + struct pca_result & result) { + // allocate tensors + ggml_gallocr_alloc_graph(allocr, gf); + + if (ggml_backend_is_cpu(model.backend)) { + ggml_backend_cpu_set_n_threads(model.backend, params.n_threads); + } + +// TODO: enable GPU support when support for GGML_OP_SQRT is added +//#ifdef GGML_USE_METAL +// if (ggml_backend_is_metal(model.backend)) { +// ggml_backend_metal_set_n_cb(model.backend, params.n_threads); +// } +//#endif + + ggml_status res = ggml_backend_graph_compute(model.backend, gf); + if (res == GGML_STATUS_SUCCESS) { + auto extract_i = [](std::string prefix, std::string str) -> int { + int i = -1; + if (str.rfind(prefix, 0) == 0) { + sscanf(str.c_str(), (prefix + "%d").c_str(), &i); + } + return i; + }; + result.calculated_square = NULL; + result.eigenvectors.clear(); + result.distances.clear(); + result.eigenvectors.resize(params.n_batch); + result.distances.resize(params.n_batch); + // get output nodes + for (int i = 0; i < gf->n_nodes; ++i) { + auto node = gf->nodes[i]; + int iter = -1; + // find b_tensor (without copying data from device) + if ((iter = extract_i("b_tensor_norm_", node->name)) > -1) { + result.eigenvectors[iter] = node; + } + // find distances, then copy data from device + if ((iter = extract_i("distance_", node->name)) > -1) { + float d; + ggml_backend_tensor_get(node, &d, 0, sizeof(float)); + result.distances[iter] = d; + // std::cout << node->name << " = " << d << "\n"; + } + // find tmp_square if it exists (without copying data from device) + if (std::string(node->name) == "tmp_square") { + result.calculated_square = node; + } + } + } + return res; +} + +static void power_iteration( + const struct pca_params & params, + struct ggml_tensor * input, // shape of input: [n_samples, n_embd] + struct ggml_tensor * output) { + //printf("in power iteration\n"); + struct pca_model model(input); + + ggml_gallocr_t allocr = ggml_gallocr_new(ggml_backend_get_default_buffer_type(model.backend)); + struct pca_result result; + struct ggml_tensor * last_eigenvector = NULL; + + int n_iters = params.n_iterations / params.n_batch; // more batch, fewer iterations + for (int iter = 0; iter < n_iters; ++iter) { + bool calc_square = (iter == 0); // only need to calculate square for first iteration + struct ggml_cgraph * gf = build_graph_piter(params, model, calc_square); + // ggml_graph_dump_dot(gf, nullptr, "/tmp/_cgraph.dot"); + compute_piter(params, model, gf, allocr, result); + + for (size_t k = 0; k < result.distances.size(); ++k) { + last_eigenvector = result.eigenvectors[k]; + if (result.distances[k] < params.tolerance) { + break; // done + } + } + + if (calc_square) { + // copy and store the square matrix if needed + GGML_ASSERT(result.calculated_square != NULL); + ggml_backend_tensor_copy(result.calculated_square, model.dev_square); + } + + { + // copy last eigen vector and store as input for next iteration + GGML_ASSERT(last_eigenvector != NULL); + ggml_backend_tensor_copy(last_eigenvector, model.dev_eigenvector); + } + + printf("%s: layer %d/%d, iteration: %d / total: %d (batch = %d) ...\n", + __func__, params.i_layer+1, params.n_layers, iter, n_iters, params.n_batch); + } + + // get output tensor + GGML_ASSERT(last_eigenvector); + ggml_backend_tensor_get(last_eigenvector, output->data, 0, ggml_nbytes(last_eigenvector)); + //print_debug_tensor(output); + ggml_gallocr_free(allocr); +} + +static void run_pca( + struct pca_params & params, + const std::vector & v_input, // shape of v_input[0]: [n_samples, n_embd] + const std::vector & v_output) { + printf("%s: Running PCA...\n", __func__); + for (size_t il = 0; il < v_input.size(); ++il) { + + // prepare output vector + struct ggml_tensor * ctrl_out = v_output[il]; + ggml_format_name(ctrl_out, "direction.%ld", il+1); + + // run power_iteration + params.i_layer = il; + params.n_layers = v_input.size(); + power_iteration(params, v_input[il], ctrl_out); + printf("%s: Done layer %d / %d\n", __func__, (int) il+1, (int) v_input.size()); + } +} + +} diff --git a/examples/cvector-generator/positive.txt b/examples/cvector-generator/positive.txt new file mode 100644 index 000000000..f28e9aa1a --- /dev/null +++ b/examples/cvector-generator/positive.txt @@ -0,0 +1 @@ +[INST] Act like a person who is extremely happy. [/INST] \ No newline at end of file