commit ff2fdecce1e480348086df6962ef320104c84550 Author: Adam Treat Date: Sat Apr 8 23:28:39 2023 -0400 Initial commit. diff --git a/.gitignore b/.gitignore new file mode 100644 index 00000000..01e00f3a --- /dev/null +++ b/.gitignore @@ -0,0 +1 @@ +CMakeLists.txt.user diff --git a/.gitmodules b/.gitmodules new file mode 100644 index 00000000..1a30094e --- /dev/null +++ b/.gitmodules @@ -0,0 +1,3 @@ +[submodule "ggml"] + path = ggml + url = https://github.com/manyoso/ggml.git diff --git a/CMakeLists.txt b/CMakeLists.txt new file mode 100644 index 00000000..5f40839c --- /dev/null +++ b/CMakeLists.txt @@ -0,0 +1,40 @@ +cmake_minimum_required(VERSION 3.16) + +project(gpt4all-chat VERSION 0.1 LANGUAGES CXX) + +set(CMAKE_AUTOMOC ON) +set(CMAKE_AUTORCC ON) +set(CMAKE_CXX_STANDARD_REQUIRED ON) + +find_package(Qt6 6.2 COMPONENTS Quick REQUIRED) + +set(GGML_BUILD_EXAMPLES ON CACHE BOOL "ggml: build examples" FORCE) +add_subdirectory(ggml) + +qt_add_executable(chat + main.cpp + gptj.h gptj.cpp + llm.h llm.cpp +) + +qt_add_qml_module(chat + URI gpt4all-chat + VERSION 1.0 + QML_FILES main.qml + RESOURCES icons/send_message.svg icons/stop_generating.svg icons/regenerate.svg +) + +set_target_properties(chat PROPERTIES + MACOSX_BUNDLE_GUI_IDENTIFIER my.example.com + MACOSX_BUNDLE_BUNDLE_VERSION ${PROJECT_VERSION} + MACOSX_BUNDLE_SHORT_VERSION_STRING ${PROJECT_VERSION_MAJOR}.${PROJECT_VERSION_MINOR} + MACOSX_BUNDLE TRUE + WIN32_EXECUTABLE TRUE +) + +target_compile_definitions(chat + PRIVATE $<$,$>:QT_QML_DEBUG>) +target_link_libraries(chat + PRIVATE Qt6::Quick Qt6::Svg) +target_link_libraries(chat + PRIVATE ggml ggml_utils) diff --git a/LICENSE b/LICENSE new file mode 100644 index 00000000..70d54194 --- /dev/null +++ b/LICENSE @@ -0,0 +1,21 @@ +MIT License + +Copyright (c) 2023 Adam Treat + +Permission is hereby granted, free of charge, to any person obtaining a copy +of this software and associated documentation files (the "Software"), to deal +in the Software without restriction, including without limitation the rights +to use, copy, modify, merge, publish, distribute, sublicense, and/or sell +copies of the Software, and to permit persons to whom the Software is +furnished to do so, subject to the following conditions: + +The above copyright notice and this permission notice shall be included in all +copies or substantial portions of the Software. + +THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +SOFTWARE. diff --git a/README.md b/README.md new file mode 100644 index 00000000..4a3e6d27 --- /dev/null +++ b/README.md @@ -0,0 +1 @@ +# gpt4all-chat diff --git a/ggml b/ggml new file mode 160000 index 00000000..c9f702ac --- /dev/null +++ b/ggml @@ -0,0 +1 @@ +Subproject commit c9f702ac573a2be4a1b9926979084941f95d0e33 diff --git a/gptj.cpp b/gptj.cpp new file mode 100644 index 00000000..0858e944 --- /dev/null +++ b/gptj.cpp @@ -0,0 +1,781 @@ +#include "gptj.h" +#include "ggml/ggml.h" + +#include "utils.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +// default hparams (GPT-J 6B) +struct gptj_hparams { + int32_t n_vocab = 50400; + int32_t n_ctx = 2048; + int32_t n_embd = 4096; + int32_t n_head = 16; + int32_t n_layer = 28; + int32_t n_rot = 64; + int32_t f16 = 1; +}; + +struct gptj_layer { + // normalization + struct ggml_tensor * ln_1_g; + struct ggml_tensor * ln_1_b; + + // attention + struct ggml_tensor * c_attn_q_proj_w; + struct ggml_tensor * c_attn_k_proj_w; + struct ggml_tensor * c_attn_v_proj_w; + + struct ggml_tensor * c_attn_proj_w; + + // ff + struct ggml_tensor * c_mlp_fc_w; + struct ggml_tensor * c_mlp_fc_b; + + struct ggml_tensor * c_mlp_proj_w; + struct ggml_tensor * c_mlp_proj_b; +}; + +struct gptj_model { + gptj_hparams hparams; + + // normalization + struct ggml_tensor * ln_f_g; + struct ggml_tensor * ln_f_b; + + struct ggml_tensor * wte; // position embedding + + struct ggml_tensor * lmh_g; // language model head + struct ggml_tensor * lmh_b; // language model bias + + std::vector layers; + + // key + value memory + struct ggml_tensor * memory_k; + struct ggml_tensor * memory_v; + + // + struct ggml_context * ctx; + std::map tensors; +}; + +// load the model's weights from a stream +bool gptj_model_load(const std::string &fname, std::istream &fin, gptj_model & model, gpt_vocab & vocab) { + printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str()); + + // verify magic + { + uint32_t magic; + fin.read((char *) &magic, sizeof(magic)); + if (magic != 0x67676d6c) { + fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str()); + return false; + } + } + + // load hparams + { + auto & hparams = model.hparams; + + fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); + fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx)); + fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd)); + fin.read((char *) &hparams.n_head, sizeof(hparams.n_head)); + fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer)); + fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot)); + fin.read((char *) &hparams.f16, sizeof(hparams.f16)); + + printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab); + printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx); + printf("%s: n_embd = %d\n", __func__, hparams.n_embd); + printf("%s: n_head = %d\n", __func__, hparams.n_head); + printf("%s: n_layer = %d\n", __func__, hparams.n_layer); + printf("%s: n_rot = %d\n", __func__, hparams.n_rot); + printf("%s: f16 = %d\n", __func__, hparams.f16); + } + + // load vocab + { + int32_t n_vocab = 0; + fin.read((char *) &n_vocab, sizeof(n_vocab)); + + if (n_vocab != model.hparams.n_vocab) { + fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n", + __func__, fname.c_str(), n_vocab, model.hparams.n_vocab); + return false; + } + + std::string word; + for (int i = 0; i < n_vocab; i++) { + uint32_t len; + fin.read((char *) &len, sizeof(len)); + + word.resize(len); + fin.read((char *) word.data(), len); + + vocab.token_to_id[word] = i; + vocab.id_to_token[i] = word; + } + } + + // for the big tensors, we have the option to store the data in 16-bit floats or quantized + // in order to save memory and also to speed up the computation + ggml_type wtype = GGML_TYPE_COUNT; + switch (model.hparams.f16) { + case 0: wtype = GGML_TYPE_F32; break; + case 1: wtype = GGML_TYPE_F16; break; + case 2: wtype = GGML_TYPE_Q4_0; break; + case 3: wtype = GGML_TYPE_Q4_1; break; + default: + { + fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n", + __func__, fname.c_str(), model.hparams.f16); + return false; + } + } + + const ggml_type wtype2 = GGML_TYPE_F32; + + auto & ctx = model.ctx; + + size_t ctx_size = 0; + + { + const auto & hparams = model.hparams; + + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_ctx = hparams.n_ctx; + const int n_vocab = hparams.n_vocab; + + ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g + ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b + + ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte + + ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // lmh_g + ctx_size += n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b + + ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g + ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b + + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w + + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w + + ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_fc_w + ctx_size += n_layer*( 4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b + + ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype)); // c_mlp_proj_w + ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b + + ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k + ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v + + ctx_size += (5 + 10*n_layer)*256; // object overhead + + printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0)); + } + + // create the ggml context + { + struct ggml_init_params params = { + .mem_size = ctx_size, + .mem_buffer = NULL, + }; + + model.ctx = ggml_init(params); + if (!model.ctx) { + fprintf(stderr, "%s: ggml_init() failed\n", __func__); + return false; + } + } + + // prepare memory for the weights + { + const auto & hparams = model.hparams; + + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_ctx = hparams.n_ctx; + const int n_vocab = hparams.n_vocab; + + model.layers.resize(n_layer); + + model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + + model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + model.lmh_g = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + model.lmh_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab); + + // map by name + model.tensors["transformer.wte.weight"] = model.wte; + + model.tensors["transformer.ln_f.weight"] = model.ln_f_g; + model.tensors["transformer.ln_f.bias"] = model.ln_f_b; + + model.tensors["lm_head.weight"] = model.lmh_g; + model.tensors["lm_head.bias"] = model.lmh_b; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = model.layers[i]; + + layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + layer.c_attn_q_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.c_attn_k_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.c_attn_v_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + + layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + + layer.c_mlp_fc_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 4*n_embd); + layer.c_mlp_fc_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd); + + layer.c_mlp_proj_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd); + layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + // map by name + model.tensors["transformer.h." + std::to_string(i) + ".ln_1.weight"] = layer.ln_1_g; + model.tensors["transformer.h." + std::to_string(i) + ".ln_1.bias"] = layer.ln_1_b; + + model.tensors["transformer.h." + std::to_string(i) + ".attn.q_proj.weight"] = layer.c_attn_q_proj_w; + model.tensors["transformer.h." + std::to_string(i) + ".attn.k_proj.weight"] = layer.c_attn_k_proj_w; + model.tensors["transformer.h." + std::to_string(i) + ".attn.v_proj.weight"] = layer.c_attn_v_proj_w; + + model.tensors["transformer.h." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_proj_w; + + model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.weight"] = layer.c_mlp_fc_w; + model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_in.bias"] = layer.c_mlp_fc_b; + + model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.weight"] = layer.c_mlp_proj_w; + model.tensors["transformer.h." + std::to_string(i) + ".mlp.fc_out.bias"] = layer.c_mlp_proj_b; + } + } + + // key + value memory + { + const auto & hparams = model.hparams; + + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_ctx = hparams.n_ctx; + + const int n_mem = n_layer*n_ctx; + const int n_elements = n_embd*n_mem; + + model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements); + model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements); + + const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v); + + printf("%s: memory_size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem); + } + + // load weights + { + int n_tensors = 0; + size_t total_size = 0; + + printf("%s: ", __func__); + + while (true) { + int32_t n_dims; + int32_t length; + int32_t ftype; + + fin.read(reinterpret_cast(&n_dims), sizeof(n_dims)); + fin.read(reinterpret_cast(&length), sizeof(length)); + fin.read(reinterpret_cast(&ftype), sizeof(ftype)); + + if (fin.eof()) { + break; + } + + int32_t nelements = 1; + int32_t ne[2] = { 1, 1 }; + for (int i = 0; i < n_dims; ++i) { + fin.read(reinterpret_cast(&ne[i]), sizeof(ne[i])); + nelements *= ne[i]; + } + + std::string name(length, 0); + fin.read(&name[0], length); + + if (model.tensors.find(name.data()) == model.tensors.end()) { + fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data()); + return false; + } + + auto tensor = model.tensors[name.data()]; + if (ggml_nelements(tensor) != nelements) { + fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data()); + return false; + } + + if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) { + fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n", + __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]); + return false; + } + + if (0) { + static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", }; + printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor)); + } + + size_t bpe = 0; + + switch (ftype) { + case 0: bpe = ggml_type_size(GGML_TYPE_F32); break; + case 1: bpe = ggml_type_size(GGML_TYPE_F16); break; + case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break; + case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break; + default: + { + fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype); + return false; + } + }; + + if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) { + fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", + __func__, name.data(), ggml_nbytes(tensor), nelements*bpe); + return false; + } + + fin.read(reinterpret_cast(tensor->data), ggml_nbytes(tensor)); + + //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0); + total_size += ggml_nbytes(tensor); + if (++n_tensors % 8 == 0) { + printf("."); + fflush(stdout); + } + } + + printf(" done\n"); + + printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors); + } + + return true; +} + +// load the model's weights from a file path +bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab & vocab) { + + auto fin = std::ifstream(fname, std::ios::binary); + if (!fin) { + fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str()); + return false; + } + + bool loaded = gptj_model_load(fname, fin, model, vocab); + fin.close(); + return loaded; +} + +// evaluate the transformer +// +// - model: the model +// - n_threads: number of threads to use +// - n_past: the context size so far +// - embd_inp: the embeddings of the tokens in the context +// - embd_w: the predicted logits for the next token +// +// The GPT-J model requires about 16MB of memory per input token. +// +bool gptj_eval( + const gptj_model & model, + const int n_threads, + const int n_past, + const std::vector & embd_inp, + std::vector & embd_w, + size_t & mem_per_token) { + const int N = embd_inp.size(); + + const auto & hparams = model.hparams; + + const int n_embd = hparams.n_embd; + const int n_layer = hparams.n_layer; + const int n_ctx = hparams.n_ctx; + const int n_head = hparams.n_head; + const int n_vocab = hparams.n_vocab; + const int n_rot = hparams.n_rot; + + const int d_key = n_embd/n_head; + + static size_t buf_size = 256u*1024*1024; + static void * buf = malloc(buf_size); + + if (mem_per_token > 0 && mem_per_token*N > buf_size) { + const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead + //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new); + + // reallocate + buf_size = buf_size_new; + buf = realloc(buf, buf_size); + if (buf == nullptr) { + fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size); + return false; + } + } + + struct ggml_init_params params = { + .mem_size = buf_size, + .mem_buffer = buf, + }; + + struct ggml_context * ctx0 = ggml_init(params); + struct ggml_cgraph gf = { .n_threads = n_threads }; + + struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd)); + + // wte + struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * cur; + + // norm + { + cur = ggml_norm(ctx0, inpL); + + // cur = ln_1_g*cur + ln_1_b + cur = ggml_add(ctx0, + ggml_mul(ctx0, + ggml_repeat(ctx0, model.layers[il].ln_1_g, cur), + cur), + ggml_repeat(ctx0, model.layers[il].ln_1_b, cur)); + } + + struct ggml_tensor * inpSA = cur; + + // self-attention + { + struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur); + struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur); + struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur); + + // store key and value to memory + if (N >= 1) { + struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past)); + struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past)); + + ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k)); + ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v)); + } + + // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3) + struct ggml_tensor * Q = + ggml_permute(ctx0, + ggml_rope(ctx0, + ggml_cpy(ctx0, + Qcur, + ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)), + n_past, n_rot, 0), + 0, 2, 1, 3); + + // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3) + struct ggml_tensor * K = + ggml_permute(ctx0, + ggml_rope(ctx0, + ggml_reshape_3d(ctx0, + ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd), + n_embd/n_head, n_head, n_past + N), + n_past, n_rot, 1), + 0, 2, 1, 3); + + // K * Q + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + + // KQ_scaled = KQ / sqrt(n_embd/n_head) + struct ggml_tensor * KQ_scaled = + ggml_scale(ctx0, + KQ, + ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head)) + ); + + // KQ_masked = mask_past(KQ_scaled) + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); + + // KQ = soft_max(KQ_masked) + struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked); + + // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous() + struct ggml_tensor * V_trans = + ggml_cpy(ctx0, + ggml_permute(ctx0, + ggml_reshape_3d(ctx0, + ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd), + n_embd/n_head, n_head, n_past + N), + 1, 2, 0, 3), + ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head)); + + // KQV = transpose(V) * KQ_soft_max + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max); + + // KQV_merged = KQV.permute(0, 2, 1, 3) + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + + // cur = KQV_merged.contiguous().view(n_embd, N) + cur = ggml_cpy(ctx0, + KQV_merged, + ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + + // projection (no bias) + cur = ggml_mul_mat(ctx0, + model.layers[il].c_attn_proj_w, + cur); + } + + struct ggml_tensor * inpFF = cur; + + // feed-forward network + // this is independent of the self-attention result, so it could be done in parallel to the self-attention + { + // note here we pass inpSA instead of cur + cur = ggml_mul_mat(ctx0, + model.layers[il].c_mlp_fc_w, + inpSA); + + cur = ggml_add(ctx0, + ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur), + cur); + + // GELU activation + cur = ggml_gelu(ctx0, cur); + + // projection + // cur = proj_w*cur + proj_b + cur = ggml_mul_mat(ctx0, + model.layers[il].c_mlp_proj_w, + cur); + + cur = ggml_add(ctx0, + ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur), + cur); + } + + // self-attention + FF + cur = ggml_add(ctx0, cur, inpFF); + + // input for next layer + inpL = ggml_add(ctx0, cur, inpL); + } + + // norm + { + inpL = ggml_norm(ctx0, inpL); + + // inpL = ln_f_g*inpL + ln_f_b + inpL = ggml_add(ctx0, + ggml_mul(ctx0, + ggml_repeat(ctx0, model.ln_f_g, inpL), + inpL), + ggml_repeat(ctx0, model.ln_f_b, inpL)); + } + + // lm_head + { + inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL); + + inpL = ggml_add(ctx0, + ggml_repeat(ctx0, model.lmh_b, inpL), + inpL); + } + + // logits -> probs + //inpL = ggml_soft_max(ctx0, inpL); + + // run the computation + ggml_build_forward_expand(&gf, inpL); + ggml_graph_compute (ctx0, &gf); + + //if (n_past%100 == 0) { + // ggml_graph_print (&gf); + // ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot"); + //} + + //embd_w.resize(n_vocab*N); + //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N); + + // return result for just the last token + embd_w.resize(n_vocab); + memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab); + + if (mem_per_token == 0) { + mem_per_token = ggml_used_mem(ctx0)/N; + } + //printf("used_mem = %zu\n", ggml_used_mem(ctx0)); + + ggml_free(ctx0); + + return true; +} + +struct GPTJPrivate { + const std::string modelPath; + bool modelLoaded; + gpt_vocab vocab; + gptj_model model; + int64_t t_main_start_us = 0; + int64_t t_load_us = 0; + int64_t n_threads = 0; + std::mt19937 rng; +}; + +GPTJ::GPTJ() + : d_ptr(new GPTJPrivate) { + + d_ptr->modelLoaded = false; +} + +bool GPTJ::loadModel(const std::string &modelPath, std::istream &fin) { + d_ptr->t_main_start_us = ggml_time_us(); + std::mt19937 rng(time(NULL)); + d_ptr->rng = rng; + + // load the model + { + const int64_t t_start_us = ggml_time_us(); + + if (!gptj_model_load(modelPath, fin, d_ptr->model, d_ptr->vocab)) { + std::cerr << "GPT-J ERROR: failed to load model from" << modelPath; + return false; + } + + d_ptr->t_load_us = ggml_time_us() - t_start_us; + } + + d_ptr->n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency()); + d_ptr->modelLoaded = true; + return true; +} + +GPTJ::~GPTJ() +{ + ggml_free(d_ptr->model.ctx); +} + +bool GPTJ::isModelLoaded() const +{ + return d_ptr->modelLoaded; +} + +void GPTJ::prompt(const std::string &prompt, std::function response, + int32_t n_predict, int32_t top_k, float top_p, float temp, + int32_t n_batch) { + + if (!isModelLoaded()) { + std::cerr << "GPT-J ERROR: prompt won't work with an unloaded model!\n"; + return; + } + + int n_past = 0; + + int64_t t_sample_us = 0; + int64_t t_predict_us = 0; + + std::vector logits; + + // tokenize the prompt + std::vector embd_inp = ::gpt_tokenize(d_ptr->vocab, prompt); + + n_predict = std::min(n_predict, d_ptr->model.hparams.n_ctx - (int) embd_inp.size()); + + std::vector embd; + + // determine the required inference memory per token: + size_t mem_per_token = 0; + gptj_eval(d_ptr->model, d_ptr->n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token); + + for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) { + // predict + if (embd.size() > 0) { + const int64_t t_start_us = ggml_time_us(); + + if (!gptj_eval(d_ptr->model, d_ptr->n_threads, n_past, embd, logits, mem_per_token)) { + std::cerr << "GPT-J ERROR: Failed to predict\n"; + return; + } + + t_predict_us += ggml_time_us() - t_start_us; + } + + n_past += embd.size(); + embd.clear(); + + if (i >= embd_inp.size()) { + // sample next token + + const int n_vocab = d_ptr->model.hparams.n_vocab; + + gpt_vocab::id id = 0; + + { + const int64_t t_start_sample_us = ggml_time_us(); + + id = gpt_sample_top_k_top_p(d_ptr->vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, d_ptr->rng); + + t_sample_us += ggml_time_us() - t_start_sample_us; + } + + // add it to the context + embd.push_back(id); + } else { + // if here, it means we are still processing the input prompt + for (int k = i; k < embd_inp.size(); k++) { + embd.push_back(embd_inp[k]); + if (embd.size() > n_batch) { + break; + } + } + i += embd.size() - 1; + } + + // display text + for (auto id : embd) { + if (!response(d_ptr->vocab.id_to_token[id])) + goto stop_generating; + } + + // end of text token + if (embd.back() == 50256) { + break; + } + } + +stop_generating: +#if 1 + // report timing + { + const int64_t t_main_end_us = ggml_time_us(); + + std::cout << "GPT-J INFO: mem per token = " << mem_per_token << " bytes\n"; + std::cout << "GPT-J INFO: load time = " << d_ptr->t_load_us/1000.0f << " ms\n"; + std::cout << "GPT-J INFO: sample time = " << t_sample_us/1000.0f << " ms\n"; + std::cout << "GPT-J INFO: predict time = " << t_predict_us/1000.0f << " ms / " << t_predict_us/1000.0f/n_past << " ms per token\n"; + std::cout << "GPT-J INFO: total time = " << (t_main_end_us - d_ptr->t_main_start_us)/1000.0f << " ms\n"; + fflush(stdout); + fflush(stderr); + } +#endif + + return; +} diff --git a/gptj.h b/gptj.h new file mode 100644 index 00000000..3a698180 --- /dev/null +++ b/gptj.h @@ -0,0 +1,24 @@ +#ifndef GPTJ_H +#define GPTJ_H + +#include +#include +#include + +class GPTJPrivate; +class GPTJ { +public: + GPTJ(); + ~GPTJ(); + + bool loadModel(const std::string &modelPath, std::istream &fin); + bool isModelLoaded() const; + void prompt(const std::string &prompt, std::function response, + int32_t n_predict = 200, int32_t top_k = 40, float top_p = 0.9f, float temp = 0.9f, + int32_t n_batch = 9); + +private: + GPTJPrivate *d_ptr; +}; + +#endif // GPTJ_H \ No newline at end of file diff --git a/icons/regenerate.svg b/icons/regenerate.svg new file mode 100644 index 00000000..016e6a52 --- /dev/null +++ b/icons/regenerate.svg @@ -0,0 +1 @@ + diff --git a/icons/send_message.svg b/icons/send_message.svg new file mode 100644 index 00000000..d8650b66 --- /dev/null +++ b/icons/send_message.svg @@ -0,0 +1 @@ + diff --git a/icons/stop_generating.svg b/icons/stop_generating.svg new file mode 100644 index 00000000..c627ac0e --- /dev/null +++ b/icons/stop_generating.svg @@ -0,0 +1 @@ + diff --git a/llm.cpp b/llm.cpp new file mode 100644 index 00000000..6e2ca906 --- /dev/null +++ b/llm.cpp @@ -0,0 +1,132 @@ +#include "llm.h" + +#include +#include +#include +#include + +class MyLLM: public LLM { }; +Q_GLOBAL_STATIC(MyLLM, llmInstance) +LLM *LLM::globalInstance() +{ + return llmInstance(); +} + +GPTJObject::GPTJObject() + : QObject{nullptr} + , m_gptj(new GPTJ) +{ + moveToThread(&m_llmThread); + connect(&m_llmThread, &QThread::started, this, &GPTJObject::loadModel); + m_llmThread.setObjectName("llm thread"); + m_llmThread.start(); +} + +bool GPTJObject::loadModel() +{ + if (isModelLoaded()) + return true; + + QString modelName("ggml-model-q4_0.bin"); + QFile file(QCoreApplication::applicationDirPath() + QDir::separator() + modelName); + if (file.open(QIODevice::ReadOnly)) { + + QByteArray data = file.readAll(); + std::istringstream iss(data.toStdString()); + + m_gptj->loadModel(modelName.toStdString(), iss); + emit isModelLoadedChanged(); + } + + return m_gptj; +} + +bool GPTJObject::isModelLoaded() const +{ + return m_gptj->isModelLoaded(); +} + +void GPTJObject::resetResponse() +{ + m_response = std::string(); +} + +QString GPTJObject::response() const +{ + return QString::fromStdString(m_response); +} + +bool GPTJObject::handleResponse(const std::string &response) +{ +#if 0 + printf("%s", response.c_str()); + fflush(stdout); +#endif + m_response.append(response); + emit responseChanged(); + return !m_stopGenerating; +} + +bool GPTJObject::prompt(const QString &prompt) +{ + if (!isModelLoaded()) + return false; + + m_stopGenerating = false; + auto func = std::bind(&GPTJObject::handleResponse, this, std::placeholders::_1); + emit responseStarted(); + m_gptj->prompt(prompt.toStdString(), func); + emit responseStopped(); + return true; +} + +LLM::LLM() + : QObject{nullptr} + , m_gptj(new GPTJObject) + , m_responseInProgress(false) +{ + connect(m_gptj, &GPTJObject::isModelLoadedChanged, this, &LLM::isModelLoadedChanged, Qt::QueuedConnection); + connect(m_gptj, &GPTJObject::responseChanged, this, &LLM::responseChanged, Qt::QueuedConnection); + connect(m_gptj, &GPTJObject::responseStarted, this, &LLM::responseStarted, Qt::QueuedConnection); + connect(m_gptj, &GPTJObject::responseStopped, this, &LLM::responseStopped, Qt::QueuedConnection); + + connect(this, &LLM::promptRequested, m_gptj, &GPTJObject::prompt, Qt::QueuedConnection); + connect(this, &LLM::resetResponseRequested, m_gptj, &GPTJObject::resetResponse, Qt::BlockingQueuedConnection); +} + +bool LLM::isModelLoaded() const +{ + return m_gptj->isModelLoaded(); +} + +void LLM::prompt(const QString &prompt) +{ + emit promptRequested(prompt); +} + +void LLM::resetResponse() +{ + emit resetResponseRequested(); // blocking queued connection +} + +void LLM::stopGenerating() +{ + m_gptj->stopGenerating(); +} + +QString LLM::response() const +{ + return m_gptj->response(); +} + +void LLM::responseStarted() +{ + m_responseInProgress = true; + emit responseInProgressChanged(); +} + +void LLM::responseStopped() +{ + m_responseInProgress = false; + emit responseInProgressChanged(); +} diff --git a/llm.h b/llm.h new file mode 100644 index 00000000..285aa009 --- /dev/null +++ b/llm.h @@ -0,0 +1,84 @@ +#ifndef LLM_H +#define LLM_H + +#include +#include +#include "gptj.h" + +class GPTJObject : public QObject +{ + Q_OBJECT + Q_PROPERTY(bool isModelLoaded READ isModelLoaded NOTIFY isModelLoadedChanged) + Q_PROPERTY(QString response READ response NOTIFY responseChanged) + +public: + + GPTJObject(); + + bool loadModel(); + bool isModelLoaded() const; + void resetResponse(); + void stopGenerating() { m_stopGenerating = true; } + + QString response() const; + +public Q_SLOTS: + bool prompt(const QString &prompt); + +Q_SIGNALS: + void isModelLoadedChanged(); + void responseChanged(); + void responseStarted(); + void responseStopped(); + +private: + bool handleResponse(const std::string &response); + +private: + GPTJ *m_gptj; + std::stringstream m_debug; + std::string m_response; + QThread m_llmThread; + std::atomic m_stopGenerating; +}; + +class LLM : public QObject +{ + Q_OBJECT + Q_PROPERTY(bool isModelLoaded READ isModelLoaded NOTIFY isModelLoadedChanged) + Q_PROPERTY(QString response READ response NOTIFY responseChanged) + Q_PROPERTY(bool responseInProgress READ responseInProgress NOTIFY responseInProgressChanged) +public: + + static LLM *globalInstance(); + + Q_INVOKABLE bool isModelLoaded() const; + Q_INVOKABLE void prompt(const QString &prompt); + Q_INVOKABLE void resetResponse(); + Q_INVOKABLE void stopGenerating(); + + QString response() const; + bool responseInProgress() const { return m_responseInProgress; } + +Q_SIGNALS: + void isModelLoadedChanged(); + void responseChanged(); + void responseInProgressChanged(); + void promptRequested(const QString &prompt); + void resetResponseRequested(); + +private Q_SLOTS: + void responseStarted(); + void responseStopped(); + +private: + GPTJObject *m_gptj; + bool m_responseInProgress; + +private: + explicit LLM(); + ~LLM() {} + friend class MyLLM; +}; + +#endif // LLM_H diff --git a/main.cpp b/main.cpp new file mode 100644 index 00000000..04b25c53 --- /dev/null +++ b/main.cpp @@ -0,0 +1,31 @@ +#include +#include +#include + +#include + +#include "llm.h" + +int main(int argc, char *argv[]) +{ + QGuiApplication app(argc, argv); + QQmlApplicationEngine engine; + qmlRegisterSingletonInstance("llm", 1, 0, "LLM", LLM::globalInstance()); + const QUrl url(u"qrc:/gpt4all-chat/main.qml"_qs); + + QObject::connect(&engine, &QQmlApplicationEngine::objectCreated, + &app, [url](QObject *obj, const QUrl &objUrl) { + if (!obj && url == objUrl) + QCoreApplication::exit(-1); + }, Qt::QueuedConnection); + engine.load(url); + +#if 1 + QDirIterator it("qrc:", QDirIterator::Subdirectories); + while (it.hasNext()) { + qDebug() << it.next(); + } +#endif + + return app.exec(); +} diff --git a/main.qml b/main.qml new file mode 100644 index 00000000..2e9038ca --- /dev/null +++ b/main.qml @@ -0,0 +1,233 @@ +import QtQuick +import QtQuick.Controls +import llm + +Window { + id: window + width: 1280 + height: 720 + visible: true + title: qsTr("GPT4All Chat") + + Rectangle { + id: conversationList + width: 300 + anchors.left: parent.left + anchors.top: parent.top + anchors.bottom: parent.bottom + color: "#202123" + + Button { + id: newChat + text: qsTr("New chat") + anchors.top: parent.top + anchors.left: parent.left + anchors.right: parent.right + anchors.margins: 15 + padding: 15 + background: Rectangle { + opacity: .5 + border.color: "#7d7d8e" + border.width: 1 + radius: 10 + color: "#343541" + } + } + } + + Rectangle { + id: conversation + color: "#343541" + anchors.left: conversationList.right + anchors.right: parent.right + anchors.bottom: parent.bottom + anchors.top: parent.top + + ScrollView { + id: scrollView + anchors.left: parent.left + anchors.right: parent.right + anchors.top: parent.top + anchors.bottom: textInput.top + anchors.bottomMargin: 30 + ScrollBar.vertical.policy: ScrollBar.AlwaysOn + + ListModel { + id: chatModel + } + + Rectangle { + anchors.fill: parent + color: "#444654" + + ListView { + id: listView + anchors.fill: parent + header: TextField { + id: modelName + width: parent.width + color: "#d1d5db" + padding: 20 + font.pixelSize: 24 + text: "Model: GPT-J-6B-4bit" + background: Rectangle { + color: "#444654" + } + focus: false + horizontalAlignment: TextInput.AlignHCenter + } + + model: chatModel + delegate: TextArea { + text: currentResponse ? LLM.response : value + width: parent.width + color: "#d1d5db" + wrapMode: Text.WordWrap + focus: false + padding: 20 + font.pixelSize: 24 + cursorVisible: currentResponse ? LLM.responseInProgress : false + cursorPosition: text.length + background: Rectangle { + color: name === qsTr("Response: ") ? "#444654" : "#343541" + } + + leftPadding: 100 + + Rectangle { + anchors.left: parent.left + anchors.top: parent.top + anchors.leftMargin: 20 + anchors.topMargin: 20 + width: 30 + height: 30 + radius: 5 + color: name === qsTr("Response: ") ? "#10a37f" : "#ec86bf" + + Text { + anchors.centerIn: parent + text: name === qsTr("Response: ") ? "R" : "P" + color: "white" + } + } + } + + property bool shouldAutoScroll: true + property bool isAutoScrolling: false + + Connections { + target: LLM + function onResponseChanged() { + if (listView.shouldAutoScroll) { + listView.isAutoScrolling = true + listView.positionViewAtEnd() + listView.isAutoScrolling = false + } + } + } + + onContentYChanged: { + if (!isAutoScrolling) + shouldAutoScroll = atYEnd + } + + Component.onCompleted: { + shouldAutoScroll = true + positionViewAtEnd() + } + + footer: Item { + id: bottomPadding + width: parent.width + height: 60 + } + } + } + } + + Button { + Image { + anchors.verticalCenter: parent.verticalCenter + anchors.left: parent.left + anchors.leftMargin: 15 + source: LLM.responseInProgress ? "qrc:/gpt4all-chat/icons/stop_generating.svg" : "qrc:/gpt4all-chat/icons/regenerate.svg" + } + text: LLM.responseInProgress ? qsTr(" Stop generating") : qsTr(" Regenerate response") + onClicked: { + if (LLM.responseInProgress) + LLM.stopGenerating() + else { + LLM.resetResponse() + if (chatModel.count) { + var listElement = chatModel.get(chatModel.count - 1) + if (listElement.name === qsTr("Response: ")) { + listElement.currentResponse = true + listElement.value = LLM.response + LLM.prompt(listElement.prompt) + } + } + } + } + anchors.bottom: textInput.top + anchors.horizontalCenter: textInput.horizontalCenter + anchors.bottomMargin: 40 + padding: 15 + background: Rectangle { + opacity: .5 + border.color: "#7d7d8e" + border.width: 1 + radius: 10 + color: "#343541" + } + } + + TextField { + id: textInput + anchors.left: parent.left + anchors.right: parent.right + anchors.bottom: parent.bottom + anchors.margins: 30 + color: "#dadadc" + padding: 20 + font.pixelSize: 24 + placeholderText: qsTr("Send a message...") + placeholderTextColor: "#7d7d8e" + background: Rectangle { + color: "#40414f" + radius: 10 + } + onAccepted: { + LLM.stopGenerating() + + if (chatModel.count) { + var listElement = chatModel.get(chatModel.count - 1) + listElement.currentResponse = false + listElement.value = LLM.response + } + chatModel.append({"name": qsTr("Prompt: "), "currentResponse": false, "value": textInput.text}) + chatModel.append({"name": qsTr("Response: "), "currentResponse": true, "value": "", "prompt": textInput.text}) + + LLM.resetResponse() + LLM.prompt(textInput.text) + textInput.text = "" + } + + Button { + anchors.right: textInput.right + anchors.verticalCenter: textInput.verticalCenter + anchors.rightMargin: 15 + width: 30 + height: 30 + + background: Image { + anchors.centerIn: parent + source: "qrc:/gpt4all-chat/icons/send_message.svg" + } + + onClicked: { + textInput.onAccepted() + } + } + } + } +}