diff --git a/Makefile b/Makefile index 3ee7ac9db..7da2624ef 100644 --- a/Makefile +++ b/Makefile @@ -121,7 +121,7 @@ BLAS_BUILD = ifeq ($(OS),Windows_NT) BLAS_BUILD = $(CXX) $(CXXFLAGS) ggml_blas.o ggml_v1.o expose.o common.o llama_adapter.o gpttype_adapter.o libopenblas.lib -shared -o koboldcpp_blas.dll $(LDFLAGS) else - BLAS_BUILD = @echo 'Your OS is $(OS) and does not appear to be Windows. If you want to use openblas, please link it manually with LLAMA_OPENBLAS=1' + BLAS_BUILD = @echo 'Your OS $(OS) does not appear to be Windows. If you want to use openblas, please install it seperately, then link it manually with LLAMA_OPENBLAS=1' endif # @@ -196,8 +196,8 @@ perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o embedding: examples/embedding/embedding.cpp ggml.o llama.o common.o $(CXX) $(CXXFLAGS) examples/embedding/embedding.cpp ggml.o llama.o common.o -o embedding $(LDFLAGS) -gpt2: ggml_v1.o - $(CXX) $(CXXFLAGS) otherarch/gpt2_v1.cpp otherarch/utils.cpp ggml_v1.o -o gpt2 $(LDFLAGS) +gpt2: ggml.o + $(CXX) $(CXXFLAGS) otherarch/gpt2_v2.cpp otherarch/utils.cpp ggml.o -o gpt2 $(LDFLAGS) # # Tests # diff --git a/expose.cpp b/expose.cpp index 3daabe475..604031159 100644 --- a/expose.cpp +++ b/expose.cpp @@ -31,19 +31,19 @@ extern "C" std::string model = inputs.model_filename; file_format = check_file_format(model.c_str()); - if(file_format==FileFormat::GPTJ1 || file_format==FileFormat::GPTJ2 || file_format==FileFormat::GPTJ3) + if(file_format==FileFormat::GPTJ_1 || file_format==FileFormat::GPTJ_2 || file_format==FileFormat::GPTJ_3) { printf("\n---\nIdentified as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format); ModelLoadResult lr = gpttype_load_model(inputs, file_format); if (lr == ModelLoadResult::RETRY_LOAD) { - file_format = FileFormat::GPTJ2; + file_format = FileFormat::GPTJ_2; printf("\n---\nRetrying as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format); lr = gpttype_load_model(inputs, file_format); } if (lr == ModelLoadResult::RETRY_LOAD) { - file_format = FileFormat::GPTJ3; + file_format = FileFormat::GPTJ_3; printf("\n---\nRetrying as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format); lr = gpttype_load_model(inputs, file_format); } @@ -57,10 +57,16 @@ extern "C" return true; } } - else if(file_format==FileFormat::GPT2) + else if(file_format==FileFormat::GPT2_1||file_format==FileFormat::GPT2_2) { printf("\n---\nIdentified as GPT-2 model: (ver %d)\nAttempting to Load...\n---\n", file_format); ModelLoadResult lr = gpttype_load_model(inputs, file_format); + if (lr == ModelLoadResult::RETRY_LOAD) + { + file_format = FileFormat::GPT2_2; + printf("\n---\nRetrying as GPT-2 model: (ver %d)\nAttempting to Load...\n---\n", file_format); + lr = gpttype_load_model(inputs, file_format); + } if (lr == ModelLoadResult::FAIL || lr == ModelLoadResult::RETRY_LOAD) { return false; @@ -79,7 +85,8 @@ extern "C" generation_outputs generate(const generation_inputs inputs, generation_outputs &output) { - if (file_format == FileFormat::GPTJ1 || file_format == FileFormat::GPTJ2 || file_format==FileFormat::GPTJ3 || file_format==FileFormat::GPT2) + if (file_format == FileFormat::GPTJ_1 || file_format == FileFormat::GPTJ_2 || file_format==FileFormat::GPTJ_3 + || file_format==FileFormat::GPT2_1 || file_format==FileFormat::GPT2_2 ) { return gpttype_generate(inputs, output); } diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp index 9cf0a7a40..5be8145c1 100644 --- a/gpttype_adapter.cpp +++ b/gpttype_adapter.cpp @@ -16,13 +16,15 @@ #include "otherarch/gptj_v1.cpp" #include "otherarch/gptj_v2.cpp" #include "otherarch/gpt2_v1.cpp" +#include "otherarch/gpt2_v2.cpp" //return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt) static FileFormat file_format = FileFormat::BADFORMAT; static gpt_vocab vocab; static gptj_model_v1 model_v1; static gptj_model model_v2; -static gpt2_model model_gpt2; +static gpt2_v1_model model_gpt2_v1; +static gpt2_model model_gpt2_v2; static gpt_params params; static int n_past = 0; static int n_threads = 4; @@ -42,19 +44,41 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in n_batch = params.n_batch = inputs.batch_size; modelname = params.model = inputs.model_filename; - if (file_format == FileFormat::GPT2) + if (file_format == FileFormat::GPT2_1) { - ModelLoadResult res = gpt2_model_load(params.model, model_gpt2, vocab, file_format); + ModelLoadResult res = legacy_gpt2_model_load(params.model, model_gpt2_v1, vocab, file_format); if(res==ModelLoadResult::FAIL) { fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); return res; + } + else if(res==ModelLoadResult::RETRY_LOAD) + { + printf("\nTensor Transposition Detected! Retrying GPT-2 model loading..."); + return res; } // determine the required inference memory per token: - gpt2_eval(model_gpt2, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token, file_format); + legacy_gpt2_eval(model_gpt2_v1, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token, file_format); return ModelLoadResult::SUCCESS; } - else if (file_format == FileFormat::GPTJ1 || file_format == FileFormat::GPTJ2) + else if (file_format == FileFormat::GPT2_2) + { + ModelLoadResult res = gpt2_model_load(params.model, model_gpt2_v2, vocab, file_format); + if(res==ModelLoadResult::FAIL) + { + fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); + return res; + } + else if(res==ModelLoadResult::RETRY_LOAD) + { + printf("\nTensor Transposition Detected! Retrying GPT-2 model loading..."); + return res; + } + // determine the required inference memory per token: + gpt2_eval(model_gpt2_v2, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token, file_format); + return ModelLoadResult::SUCCESS; + } + else if (file_format == FileFormat::GPTJ_1 || file_format == FileFormat::GPTJ_2) { ModelLoadResult res = legacy_gptj_model_load(params.model, model_v1, vocab, file_format); if(res==ModelLoadResult::FAIL) @@ -125,17 +149,21 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o //truncate to front of the prompt if its too long int32_t nctx = 512; - if(file_format == FileFormat::GPTJ1||file_format == FileFormat::GPTJ2) + if(file_format == FileFormat::GPTJ_1||file_format == FileFormat::GPTJ_2) { nctx = model_v1.hparams.n_ctx; } - else if(file_format==FileFormat::GPTJ3) + else if(file_format==FileFormat::GPTJ_3) { nctx = model_v2.hparams.n_ctx; } - else if(file_format==FileFormat::GPT2) + else if(file_format==FileFormat::GPT2_1) { - nctx = model_gpt2.hparams.n_ctx; + nctx = model_gpt2_v1.hparams.n_ctx; + } + else if(file_format==FileFormat::GPT2_2) + { + nctx = model_gpt2_v2.hparams.n_ctx; } if (embd_inp.size() + params.n_predict > nctx) @@ -198,17 +226,21 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o double time1 = 0, time2 = 0; unsigned int embd_inp_size = embd_inp.size(); int32_t n_vocab = 0; - if(file_format == FileFormat::GPTJ1||file_format == FileFormat::GPTJ2) + if(file_format == FileFormat::GPTJ_1||file_format == FileFormat::GPTJ_2) { n_vocab = model_v1.hparams.n_vocab; } - else if(file_format == FileFormat::GPTJ3) + else if(file_format == FileFormat::GPTJ_3) { n_vocab = model_v2.hparams.n_vocab; } - else if(file_format == FileFormat::GPT2) + else if(file_format == FileFormat::GPT2_1) { - n_vocab = model_gpt2.hparams.n_vocab; + n_vocab = model_gpt2_v1.hparams.n_vocab; + } + else if(file_format == FileFormat::GPT2_2) + { + n_vocab = model_gpt2_v2.hparams.n_vocab; } else { @@ -236,11 +268,15 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o bool evalres = false; //print_tok_vec(logits); - if(file_format==FileFormat::GPT2) + if(file_format==FileFormat::GPT2_1) { - evalres = gpt2_eval(model_gpt2, params.n_threads, n_past, embd, logits, mem_per_token, file_format); + evalres = legacy_gpt2_eval(model_gpt2_v1, params.n_threads, n_past, embd, logits, mem_per_token, file_format); } - else if(file_format==FileFormat::GPTJ1 || file_format==FileFormat::GPTJ2) + else if(file_format==FileFormat::GPT2_2) + { + evalres = gpt2_eval(model_gpt2_v2, params.n_threads, n_past, embd, logits, mem_per_token, file_format); + } + else if(file_format==FileFormat::GPTJ_1 || file_format==FileFormat::GPTJ_2) { evalres = legacy_gptj_eval(model_v1, params.n_threads, n_past, embd, logits, mem_per_token, file_format); } diff --git a/llama.cpp b/llama.cpp index 581a8399d..75cab4894 100644 --- a/llama.cpp +++ b/llama.cpp @@ -761,6 +761,8 @@ static bool llama_eval_internal( auto & kv_self = model.kv_self; + printf("\ns:%d\n",llama_get_kv_cache_size(&lctx)); + LLAMA_ASSERT(!!kv_self.ctx); const int n_embd = hparams.n_embd; diff --git a/model_adapter.cpp b/model_adapter.cpp index 6f4cb7316..159d40e5c 100644 --- a/model_adapter.cpp +++ b/model_adapter.cpp @@ -85,11 +85,11 @@ void print_tok_vec(std::vector &embd) fin.read((char *) &vocabsiz, sizeof(int32_t)); if(vocabsiz==50400) //know GPT-J vocab size { - fileformat = FileFormat::GPTJ1; + fileformat = FileFormat::GPTJ_1; } if(vocabsiz==50257) { - fileformat = FileFormat::GPT2; + fileformat = FileFormat::GPT2_1; } } else if(magic == 0x67676d66) //v2 format ggmf diff --git a/model_adapter.h b/model_adapter.h index fb6976703..5cfe74057 100644 --- a/model_adapter.h +++ b/model_adapter.h @@ -20,11 +20,12 @@ enum FileFormat GGHF=2, // 2=(llama ggmf) GGJT=3, // 3=(llama ggjt) - GPTJ1=100, //the very first super old GPTJ format - GPTJ2=101, //pygmalion, uses old ggml lib - GPTJ3=102, //uses new ggml lib + GPTJ_1=100, //the very first super old GPTJ format + GPTJ_2=101, //pygmalion, uses old ggml lib + GPTJ_3=102, //uses new ggml lib - GPT2=200, + GPT2_1=200, + GPT2_2=201 }; enum ModelLoadResult diff --git a/otherarch/gpt2_v1.cpp b/otherarch/gpt2_v1.cpp index f40e1feae..6a784a85d 100644 --- a/otherarch/gpt2_v1.cpp +++ b/otherarch/gpt2_v1.cpp @@ -17,7 +17,7 @@ // load the model's weights from a file -ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab, FileFormat file_format) { +ModelLoadResult legacy_gpt2_model_load(const std::string & fname, gpt2_v1_model & model, gpt_vocab & vocab, FileFormat file_format) { printf("%s: loading model from '%s'\n", __func__, fname.c_str()); auto fin = std::ifstream(fname, std::ios::binary); @@ -267,9 +267,19 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g } 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 ModelLoadResult::FAIL; + //test for transposition and retry older loader + if(tensor->ne[0]==ne[1] && tensor->ne[1]==ne[0] && should_transpose_layer(name)) + { + printf("\nFound a transposed tensor. This could be an older or newer model. Retrying load..."); + ggml_v1_free(ctx); + return ModelLoadResult::RETRY_LOAD; + } + else + { + 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 ModelLoadResult::FAIL; + } } const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_v1_fp16_t); @@ -302,8 +312,8 @@ ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, g // - embd_inp: the embeddings of the tokens in the context // - embd_w: the predicted logits for the next token // -bool gpt2_eval( - const gpt2_model & model, +bool legacy_gpt2_eval( + const gpt2_v1_model & model, const int n_threads, const int n_past, const std::vector & embd_inp, @@ -641,13 +651,13 @@ bool gpt2_eval( // int64_t t_load_us = 0; // gpt_vocab vocab; -// gpt2_model model; +// gpt2_v1_model model; // // load the model // { // const int64_t t_start_us = ggml_v1_time_us(); -// if (!gpt2_model_load(params.model, model, vocab, FileFormat::GPT2)) { +// if (!legacy_gpt2_model_load(params.model, model, vocab, FileFormat::GPT2_1)) { // fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); // return 1; // } @@ -676,14 +686,14 @@ bool gpt2_eval( // // determine the required inference memory per token: // size_t mem_per_token = 0; -// gpt2_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token, FileFormat::GPT2); +// legacy_gpt2_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token, FileFormat::GPT2_1); // for (int i = embd.size(); i < embd_inp.size() + params.n_predict; i++) { // // predict // if (embd.size() > 0) { // const int64_t t_start_us = ggml_v1_time_us(); -// if (!gpt2_eval(model, params.n_threads, n_past, embd, logits, mem_per_token, FileFormat::GPT2)) { +// if (!legacy_gpt2_eval(model, params.n_threads, n_past, embd, logits, mem_per_token, FileFormat::GPT2_1)) { // printf("Failed to predict\n"); // return 1; // } diff --git a/otherarch/gpt2_v2.cpp b/otherarch/gpt2_v2.cpp new file mode 100644 index 000000000..f54aac56f --- /dev/null +++ b/otherarch/gpt2_v2.cpp @@ -0,0 +1,805 @@ +#include "ggml.h" +#include "otherarch.h" + +#include "utils.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "model_adapter.h" + + + +// load the model's weights from a file +ModelLoadResult gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab, FileFormat file_format) { + printf("%s: loading model from '%s'\n", __func__, fname.c_str()); + + auto fin = std::ifstream(fname, std::ios::binary); + if (!fin) { + fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str()); + return ModelLoadResult::FAIL; + } + + // 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 ModelLoadResult::FAIL; + } + } + + // 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.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: 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 ModelLoadResult::FAIL; + } + + 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 ModelLoadResult::FAIL; + } + } + + 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_vocab*n_embd*ggml_type_sizef(wtype); // wte + ctx_size += n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe + ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype); // lm_head + + 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*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g + ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b + + ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_attn_w + ctx_size += n_layer*( 3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b + + ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w + ctx_size += n_layer*( n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_b + + 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 += (6 + 12*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 ModelLoadResult::FAIL; + } + } + + // 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.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.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx); + model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + + // map by name + model.tensors["model/ln_f/g"] = model.ln_f_g; + model.tensors["model/ln_f/b"] = model.ln_f_b; + + model.tensors["model/wte"] = model.wte; + model.tensors["model/wpe"] = model.wpe; + model.tensors["model/lm_head"] = model.lm_head; + + 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.ln_2_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + layer.ln_2_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + layer.c_attn_attn_w = ggml_new_tensor_2d(ctx, wtype, n_embd, 3*n_embd); + layer.c_attn_attn_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd); + + layer.c_attn_proj_w = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.c_attn_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 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["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g; + model.tensors["model/h" + std::to_string(i) + "/ln_1/b"] = layer.ln_1_b; + + model.tensors["model/h" + std::to_string(i) + "/ln_2/g"] = layer.ln_2_g; + model.tensors["model/h" + std::to_string(i) + "/ln_2/b"] = layer.ln_2_b; + + model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w; + model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b; + + model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w; + model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b; + + model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"] = layer.c_mlp_fc_w; + model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"] = layer.c_mlp_fc_b; + + model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w; + model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"] = 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 + { + size_t total_size = 0; + + bool has_lm_head = false; + + 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 ModelLoadResult::FAIL; + } + + 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 ModelLoadResult::FAIL; + } + + 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 ModelLoadResult::FAIL; + } + + 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 ModelLoadResult::FAIL; + } + }; + + 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 ModelLoadResult::FAIL; + } + + fin.read(reinterpret_cast(tensor->data), ggml_nbytes(tensor)); + + // GPT-2 models share the WTE tensor as the LM head + if (name == "model/wte" && has_lm_head == false) { + memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor)); + } + + if (name == "model/lm_head") { + has_lm_head = true; + } + + total_size += ggml_nbytes(tensor); + } + + printf("%s: model size = %8.2f MB\n", __func__, total_size/1024.0/1024.0); + } + + fin.close(); + + return ModelLoadResult::SUCCESS; +} + +// 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 +// +bool gpt2_eval( + const gpt2_model & model, + const int n_threads, + const int n_past, + const std::vector & embd_inp, + std::vector & embd_w, + size_t & mem_per_token, + FileFormat file_format) { + 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; + + 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)); + + struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N); + for (int i = 0; i < N; ++i) { + ((int32_t *) position->data)[i] = n_past + i; + } + + // wte + wpe + struct ggml_tensor * inpL = + ggml_add(ctx0, + ggml_get_rows(ctx0, model.wte, embd), + ggml_get_rows(ctx0, model.wpe, position)); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_tensor * cur; + + // norm + { + // [ 768, N] + cur = ggml_norm(ctx0, inpL); + + // cur = ln_1_g*cur + ln_1_b + // [ 768, N] + 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)); + } + + // attn + // [2304, 768] - model.layers[il].c_attn_attn_w + // [2304, 1] - model.layers[il].c_attn_attn_b + // [ 768, N] - cur (in) + // [2304, N] - cur (out) + // + // cur = attn_w*cur + attn_b + // [2304, N] + { + cur = ggml_mul_mat(ctx0, + model.layers[il].c_attn_attn_w, + cur); + + cur = ggml_add(ctx0, + ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur), + cur); + } + + // self-attention + { + struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd); + struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd); + struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd); + + // 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) + // [64, N, 12] + struct ggml_tensor * Q = + ggml_permute(ctx0, + ggml_cpy(ctx0, + Qcur, + ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)), + 0, 2, 1, 3); + + // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3) + // [64, n_past + N, 12] + struct ggml_tensor * K = + ggml_permute(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), + 0, 2, 1, 3); + + // GG: flash attention + //struct ggml_tensor * V = + // 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, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head)); + + //struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true); + + // K * Q + // [n_past + N, N, 12] + struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q); + + // KQ_scaled = KQ / sqrt(n_embd/n_head) + // [n_past + N, N, 12] + 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) + // [n_past + N, N, 12] + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past); + + // KQ = soft_max(KQ_masked) + // [n_past + N, N, 12] + 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() + // [n_past + N, 64, 12] + 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 + // [64, N, 12] + struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max); + + // KQV_merged = KQV.permute(0, 2, 1, 3) + // [64, 12, N] + struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3); + + // cur = KQV_merged.contiguous().view(n_embd, N) + // [768, N] + cur = ggml_cpy(ctx0, + KQV_merged, + ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N)); + } + + // projection + // [ 768, 768] - model.layers[il].c_attn_proj_w + // [ 768, 1] - model.layers[il].c_attn_proj_b + // [ 768, N] - cur (in) + // [ 768, N] - cur (out) + // + // cur = proj_w*cur + proj_b + // [768, N] + { + cur = ggml_mul_mat(ctx0, + model.layers[il].c_attn_proj_w, + cur); + + cur = ggml_add(ctx0, + ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur), + cur); + } + + // add the input + cur = ggml_add(ctx0, cur, inpL); + + struct ggml_tensor * inpFF = cur; + + // feed-forward network + { + // norm + { + cur = ggml_norm(ctx0, inpFF); + + // cur = ln_2_g*cur + ln_2_b + // [ 768, N] + cur = ggml_add(ctx0, + ggml_mul(ctx0, + ggml_repeat(ctx0, model.layers[il].ln_2_g, cur), + cur), + ggml_repeat(ctx0, model.layers[il].ln_2_b, cur)); + } + + // fully connected + // [3072, 768] - model.layers[il].c_mlp_fc_w + // [3072, 1] - model.layers[il].c_mlp_fc_b + // [ 768, N] - cur (in) + // [3072, N] - cur (out) + // + // cur = fc_w*cur + fc_b + // [3072, N] + cur = ggml_mul_mat(ctx0, + model.layers[il].c_mlp_fc_w, + cur); + + cur = ggml_add(ctx0, + ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur), + cur); + + // GELU activation + // [3072, N] + cur = ggml_gelu(ctx0, cur); + + // projection + // [ 768, 3072] - model.layers[il].c_mlp_proj_w + // [ 768, 1] - model.layers[il].c_mlp_proj_b + // [3072, N] - cur (in) + // [ 768, N] - cur (out) + // + // cur = proj_w*cur + proj_b + // [768, N] + 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); + } + + // input for next layer + inpL = ggml_add(ctx0, cur, inpFF); + } + + // norm + { + // [ 768, N] + inpL = ggml_norm(ctx0, inpL); + + // inpL = ln_f_g*inpL + ln_f_b + // [ 768, N] + inpL = ggml_add(ctx0, + ggml_mul(ctx0, + ggml_repeat(ctx0, model.ln_f_g, inpL), + inpL), + ggml_repeat(ctx0, model.ln_f_b, inpL)); + } + + // inpL = WTE * inpL + // [ 768, 50257] - model.lm_head + // [ 768, N] - inpL + inpL = ggml_mul_mat(ctx0, model.lm_head, 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 just for 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; +} + +// int main(int argc, char ** argv) { +// ggml_time_init(); +// const int64_t t_main_start_us = ggml_time_us(); + +// gpt_params params; +// params.model = "models/gpt-2-117M/ggml-model.bin"; + +// if (utils_gpt_params_parse(argc, argv, params) == false) { +// return 1; +// } + +// if (params.seed < 0) { +// params.seed = time(NULL); +// } + +// printf("%s: seed = %d\n", __func__, params.seed); + +// std::mt19937 rng(params.seed); +// if (params.prompt.empty()) { +// if( !isatty(STDIN_FILENO) ){ +// std::string line; +// while( std::getline(std::cin, line) ){ +// params.prompt = params.prompt + "\n" + line; +// } +// } else { +// params.prompt = utils_gpt_random_prompt(rng); +// } +// } + +// int64_t t_load_us = 0; + +// gpt_vocab vocab; +// gpt2_model model; + +// // load the model +// { +// const int64_t t_start_us = ggml_time_us(); + +// if (!gpt2_model_load(params.model, model, vocab)) { +// fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); +// return 1; +// } + +// t_load_us = ggml_time_us() - t_start_us; +// } + +// 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(vocab, params.prompt); + +// params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size()); + +// printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str()); +// printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size()); +// for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) { +// printf("%d ", embd_inp[i]); +// } +// printf("\n\n"); + +// // submit the input prompt token-by-token +// // this reduces the memory usage during inference, at the cost of a bit of speed at the beginning +// std::vector embd; + +// // determine the required inference memory per token: +// size_t mem_per_token = 0; +// gpt2_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token); + +// for (int i = embd.size(); i < embd_inp.size() + params.n_predict; i++) { +// // predict +// if (embd.size() > 0) { +// const int64_t t_start_us = ggml_time_us(); + +// if (!gpt2_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) { +// printf("Failed to predict\n"); +// return 1; +// } + +// 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 top_k = params.top_k; +// const float top_p = params.top_p; +// const float temp = params.temp; + +// const int n_vocab = 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(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, 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() >= params.n_batch) { +// break; +// } +// } +// i += embd.size() - 1; +// } + +// // display text +// for (auto id : embd) { +// printf("%s", vocab.id_to_token[id].c_str()); +// } +// fflush(stdout); + +// // end of text token +// if (embd.back() == 50256) { +// break; +// } +// } + +// // report timing +// { +// const int64_t t_main_end_us = ggml_time_us(); + +// printf("\n\n"); +// printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token); +// printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f); +// printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f); +// printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past); +// printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f); +// } + +// ggml_free(model.ctx); + +// return 0; +// } \ No newline at end of file diff --git a/otherarch/gptj_v1.cpp b/otherarch/gptj_v1.cpp index e88f5917a..6d3530c69 100644 --- a/otherarch/gptj_v1.cpp +++ b/otherarch/gptj_v1.cpp @@ -20,7 +20,7 @@ ModelLoadResult legacy_gptj_model_load(const std::string & fname, gptj_model_v1 & model, gpt_vocab & vocab, FileFormat file_format) { printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str()); - bool super_old_format = (file_format==FileFormat::GPTJ1); + bool super_old_format = (file_format==FileFormat::GPTJ_1); auto fin = std::ifstream(fname, std::ios::binary); if (!fin) { @@ -372,7 +372,7 @@ bool legacy_gptj_eval( size_t & mem_per_token, FileFormat file_format) { - bool super_old_format = (file_format==FileFormat::GPTJ1); + bool super_old_format = (file_format==FileFormat::GPTJ_1); const int N = embd_inp.size(); const auto & hparams = model.hparams; diff --git a/otherarch/gptj_v1_main.cpp b/otherarch/gptj_v1_main.cpp index dd7f98591..6cc152cb7 100644 --- a/otherarch/gptj_v1_main.cpp +++ b/otherarch/gptj_v1_main.cpp @@ -33,7 +33,7 @@ int main(int argc, char ** argv) { gpt_vocab vocab; gptj_model_v1 model; - FileFormat file_format = FileFormat::GPTJ2; + FileFormat file_format = FileFormat::GPTJ_2; // load the model { diff --git a/otherarch/otherarch.h b/otherarch/otherarch.h index 1b734e5a5..032b75984 100644 --- a/otherarch/otherarch.h +++ b/otherarch/otherarch.h @@ -123,7 +123,7 @@ struct gpt2_hparams { int32_t f16 = 1; }; -struct gpt2_layer { +struct gpt2_v1_layer { // normalization struct ggml_v1_tensor * ln_1_g; struct ggml_v1_tensor * ln_1_b; @@ -146,7 +146,7 @@ struct gpt2_layer { struct ggml_v1_tensor * c_mlp_proj_b; }; -struct gpt2_model { +struct gpt2_v1_model { gpt2_hparams hparams; // normalization @@ -156,7 +156,7 @@ struct gpt2_model { struct ggml_v1_tensor * wte; // position embedding struct ggml_v1_tensor * wpe; // token embedding - std::vector layers; + std::vector layers; // key + value memory struct ggml_v1_tensor * memory_k; @@ -167,7 +167,53 @@ struct gpt2_model { std::map tensors; }; -ModelLoadResult legacy_gptj_model_load(const std::string &fname, gptj_model_v1 &model, gpt_vocab &vocab, FileFormat file_format); -bool legacy_gptj_eval(const gptj_model_v1 &model, const int n_threads, const int n_past, const std::vector &embd_inp, std::vector &embd_w, size_t &mem_per_token, FileFormat file_format); -ModelLoadResult gptj_model_load(const std::string &fname, gptj_model &model, gpt_vocab &vocab); -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); + +struct gpt2_layer { + // normalization + struct ggml_tensor * ln_1_g; + struct ggml_tensor * ln_1_b; + + struct ggml_tensor * ln_2_g; + struct ggml_tensor * ln_2_b; + + // attention + struct ggml_tensor * c_attn_attn_w; + struct ggml_tensor * c_attn_attn_b; + + struct ggml_tensor * c_attn_proj_w; + struct ggml_tensor * c_attn_proj_b; + + // mlp + 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 gpt2_model { + gpt2_hparams hparams; + + // normalization + struct ggml_tensor * ln_f_g; + struct ggml_tensor * ln_f_b; + + struct ggml_tensor * wte; // position embedding + struct ggml_tensor * wpe; // token embedding + struct ggml_tensor * lm_head; // language model head + + std::vector layers; + + // key + value memory + struct ggml_tensor * memory_k; + struct ggml_tensor * memory_v; + + // + struct ggml_context * ctx; + std::map tensors; +}; + +// ModelLoadResult legacy_gptj_model_load(const std::string &fname, gptj_model_v1 &model, gpt_vocab &vocab, FileFormat file_format); +// bool legacy_gptj_eval(const gptj_model_v1 &model, const int n_threads, const int n_past, const std::vector &embd_inp, std::vector &embd_w, size_t &mem_per_token, FileFormat file_format); +// ModelLoadResult gptj_model_load(const std::string &fname, gptj_model &model, gpt_vocab &vocab); +// 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); diff --git a/otherarch/utils.cpp b/otherarch/utils.cpp index 1cbdb72b1..3a6095407 100644 --- a/otherarch/utils.cpp +++ b/otherarch/utils.cpp @@ -433,7 +433,11 @@ bool should_transpose_layer(std::string name) name.find(".attn.out_proj.weight")!=std::string::npos || name.find(".attn.q_proj.weight")!=std::string::npos || name.find(".attn.k_proj.weight")!=std::string::npos || - name.find(".attn.v_proj.weight")!=std::string::npos) + name.find(".attn.v_proj.weight")!=std::string::npos || + name.find("/attn/c_attn/w")!=std::string::npos || + name.find("/attn/c_proj/w")!=std::string::npos || + name.find("/mlp/c_fc/w")!=std::string::npos || + name.find("/mlp/c_proj/w")!=std::string::npos) { return true; }