diff --git a/Makefile b/Makefile index 6e989e6d0..e29b7704f 100644 --- a/Makefile +++ b/Makefile @@ -119,7 +119,7 @@ endif BLAS_BUILD = ifeq ($(OS),Windows_NT) - BLAS_BUILD = $(CXX) $(CXXFLAGS) ggml_blas.o ggml_v1.o expose.o common.o llama_adapter.o gptj_adapter.o libopenblas.lib -shared -o koboldcpp_blas.dll $(LDFLAGS) + 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' endif @@ -166,8 +166,8 @@ expose.o: expose.cpp expose.h llama_adapter.o: $(CXX) $(CXXFLAGS) -c llama_adapter.cpp -o llama_adapter.o -gptj_adapter.o: - $(CXX) $(CXXFLAGS) -c gptj_adapter.cpp -o gptj_adapter.o +gpttype_adapter.o: + $(CXX) $(CXXFLAGS) -c gpttype_adapter.cpp -o gpttype_adapter.o clean: rm -vf *.o main quantize perplexity embedding main.exe quantize.exe koboldcpp.dll koboldcpp_blas.dll gptj.exe @@ -178,10 +178,10 @@ main: examples/main/main.cpp ggml.o llama.o common.o @echo '==== Run ./main -h for help. ====' @echo -llamalib: ggml.o ggml_v1.o expose.o common.o llama_adapter.o gptj_adapter.o - $(CXX) $(CXXFLAGS) ggml.o ggml_v1.o expose.o common.o llama_adapter.o gptj_adapter.o -shared -o koboldcpp.dll $(LDFLAGS) +llamalib: ggml.o ggml_v1.o expose.o common.o llama_adapter.o gpttype_adapter.o + $(CXX) $(CXXFLAGS) ggml.o ggml_v1.o expose.o common.o llama_adapter.o gpttype_adapter.o -shared -o koboldcpp.dll $(LDFLAGS) -llamalib_blas: ggml_blas.o ggml_v1.o expose.o common.o llama_adapter.o gptj_adapter.o +llamalib_blas: ggml_blas.o ggml_v1.o expose.o common.o llama_adapter.o gpttype_adapter.o $(BLAS_BUILD) quantize: examples/quantize/quantize.cpp ggml.o llama.o @@ -193,8 +193,8 @@ 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) -gptj: ggml_v1.o - $(CXX) $(CXXFLAGS) otherarch/gptj_v1_main.cpp otherarch/utils.cpp ggml_v1.o -o gptj $(LDFLAGS) +gpt2: ggml_v1.o + $(CXX) $(CXXFLAGS) otherarch/gpt2_v1.cpp otherarch/utils.cpp ggml_v1.o -o gpt2 $(LDFLAGS) # # Tests # diff --git a/expose.cpp b/expose.cpp index 5dec31cee..3daabe475 100644 --- a/expose.cpp +++ b/expose.cpp @@ -34,18 +34,18 @@ extern "C" if(file_format==FileFormat::GPTJ1 || file_format==FileFormat::GPTJ2 || file_format==FileFormat::GPTJ3) { printf("\n---\nIdentified as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format); - ModelLoadResult lr = gptj_load_model(inputs, file_format); + ModelLoadResult lr = gpttype_load_model(inputs, file_format); if (lr == ModelLoadResult::RETRY_LOAD) { file_format = FileFormat::GPTJ2; printf("\n---\nRetrying as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format); - lr = gptj_load_model(inputs, file_format); + lr = gpttype_load_model(inputs, file_format); } if (lr == ModelLoadResult::RETRY_LOAD) { file_format = FileFormat::GPTJ3; printf("\n---\nRetrying as GPT-J model: (ver %d)\nAttempting to Load...\n---\n", file_format); - lr = gptj_load_model(inputs, file_format); + lr = gpttype_load_model(inputs, file_format); } if (lr == ModelLoadResult::FAIL || lr == ModelLoadResult::RETRY_LOAD) @@ -57,6 +57,19 @@ extern "C" return true; } } + else if(file_format==FileFormat::GPT2) + { + 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::FAIL || lr == ModelLoadResult::RETRY_LOAD) + { + return false; + } + else + { + return true; + } + } else { printf("\n---\nIdentified as LLAMA model: (ver %d)\nAttempting to Load...\n---\n", file_format); @@ -66,13 +79,13 @@ 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) + if (file_format == FileFormat::GPTJ1 || file_format == FileFormat::GPTJ2 || file_format==FileFormat::GPTJ3 || file_format==FileFormat::GPT2) { - return gptj_generate(inputs, output); + return gpttype_generate(inputs, output); } else { return llama_generate(inputs, output); - } + } } } \ No newline at end of file diff --git a/gptj_adapter.cpp b/gpttype_adapter.cpp similarity index 79% rename from gptj_adapter.cpp rename to gpttype_adapter.cpp index e9b1ac300..9cf0a7a40 100644 --- a/gptj_adapter.cpp +++ b/gpttype_adapter.cpp @@ -15,12 +15,14 @@ #include "otherarch/utils.cpp" #include "otherarch/gptj_v1.cpp" #include "otherarch/gptj_v2.cpp" +#include "otherarch/gpt2_v1.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 gpt_params params; static int n_past = 0; static int n_threads = 4; @@ -31,7 +33,7 @@ static std::vector current_context_tokens; static size_t mem_per_token = 0; static std::vector logits; -ModelLoadResult gptj_load_model(const load_model_inputs inputs, FileFormat in_file_format) +ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in_file_format) { ggml_time_init(); @@ -40,8 +42,20 @@ ModelLoadResult gptj_load_model(const load_model_inputs inputs, FileFormat in_fi n_batch = params.n_batch = inputs.batch_size; modelname = params.model = inputs.model_filename; - if (file_format == FileFormat::GPTJ1 || file_format == FileFormat::GPTJ2) - { + if (file_format == FileFormat::GPT2) + { + ModelLoadResult res = gpt2_model_load(params.model, model_gpt2, vocab, file_format); + if(res==ModelLoadResult::FAIL) + { + fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); + 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); + return ModelLoadResult::SUCCESS; + } + else if (file_format == FileFormat::GPTJ1 || file_format == FileFormat::GPTJ2) + { ModelLoadResult res = legacy_gptj_model_load(params.model, model_v1, vocab, file_format); if(res==ModelLoadResult::FAIL) { @@ -80,7 +94,7 @@ ModelLoadResult gptj_load_model(const load_model_inputs inputs, FileFormat in_fi -generation_outputs gptj_generate(const generation_inputs inputs, generation_outputs &output) +generation_outputs gpttype_generate(const generation_inputs inputs, generation_outputs &output) { params.prompt = inputs.prompt; params.seed = inputs.seed; @@ -110,7 +124,20 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp std::vector embd_inp = ::gpt_tokenize(vocab, params.prompt); //truncate to front of the prompt if its too long - auto nctx = ( (file_format == FileFormat::GPTJ1||file_format == FileFormat::GPTJ2)? model_v1.hparams.n_ctx:model_v2.hparams.n_ctx); + int32_t nctx = 512; + if(file_format == FileFormat::GPTJ1||file_format == FileFormat::GPTJ2) + { + nctx = model_v1.hparams.n_ctx; + } + else if(file_format==FileFormat::GPTJ3) + { + nctx = model_v2.hparams.n_ctx; + } + else if(file_format==FileFormat::GPT2) + { + nctx = model_gpt2.hparams.n_ctx; + } + if (embd_inp.size() + params.n_predict > nctx) { int offset = embd_inp.size() - nctx + params.n_predict; @@ -170,10 +197,25 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp timer_start(); double time1 = 0, time2 = 0; unsigned int embd_inp_size = embd_inp.size(); - const int n_vocab = ((file_format == FileFormat::GPTJ1||file_format == FileFormat::GPTJ2)? model_v1.hparams.n_vocab:model_v2.hparams.n_vocab); + int32_t n_vocab = 0; + if(file_format == FileFormat::GPTJ1||file_format == FileFormat::GPTJ2) + { + n_vocab = model_v1.hparams.n_vocab; + } + else if(file_format == FileFormat::GPTJ3) + { + n_vocab = model_v2.hparams.n_vocab; + } + else if(file_format == FileFormat::GPT2) + { + n_vocab = model_gpt2.hparams.n_vocab; + } + else + { + printf("Bad format!"); + } printf("\n"); - while (remaining_tokens > 0) { gpt_vocab::id id = 0; @@ -192,10 +234,17 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp } bool evalres = false; - if(file_format==FileFormat::GPTJ1 || file_format==FileFormat::GPTJ2) + + //print_tok_vec(logits); + if(file_format==FileFormat::GPT2) + { + evalres = gpt2_eval(model_gpt2, params.n_threads, n_past, embd, logits, mem_per_token, file_format); + } + else if(file_format==FileFormat::GPTJ1 || file_format==FileFormat::GPTJ2) { evalres = legacy_gptj_eval(model_v1, params.n_threads, n_past, embd, logits, mem_per_token, file_format); - }else + } + else { evalres = gptj_eval(model_v2, params.n_threads, n_past, embd, logits, mem_per_token); } @@ -229,11 +278,13 @@ generation_outputs gptj_generate(const generation_inputs inputs, generation_outp } { - // set the logit of the eos token (2) to zero to avoid sampling it - logits[50256] = 0; - //set logits of opening square bracket to zero. + // set the logit of the eos token (2) to zero to avoid sampling it + logits[50256] = (logits[50256]<0?logits[50256]:0); + + //gpt2 uses negative logits, so we cant zero it - id = gptj_sample_top_p_top_k(vocab, logits.data() + (logits.size() - n_vocab),last_n_tokens,repeat_penalty, top_k, top_p, temp, rng); + id = gptj_sample_top_p_top_k(vocab, logits.data() + (logits.size() - n_vocab), last_n_tokens, repeat_penalty, top_k, top_p, temp, rng); + last_n_tokens.erase(last_n_tokens.begin()); last_n_tokens.push_back(id); diff --git a/koboldcpp.dll b/koboldcpp.dll index d34727ee5..b16f8a476 100644 Binary files a/koboldcpp.dll and b/koboldcpp.dll differ diff --git a/koboldcpp_blas.dll b/koboldcpp_blas.dll index f3d5a614f..0393adf10 100644 Binary files a/koboldcpp_blas.dll and b/koboldcpp_blas.dll differ diff --git a/model_adapter.cpp b/model_adapter.cpp index a6bf611a4..6f4cb7316 100644 --- a/model_adapter.cpp +++ b/model_adapter.cpp @@ -40,7 +40,27 @@ void print_tok_vec(std::vector &embd) first = false; std::cout << i; } - std::cout << "]"; + std::cout << "]\n"; +} +void print_tok_vec(std::vector &embd) +{ + std::cout << "["; + bool first = true; + int n = 0; + for (auto i : embd) + { + if (!first) + { + std::cout << ','; + } + first = false; + std::cout << i; + if(++n>20) + { + break; + } + } + std::cout << "]\n"; } //return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt) @@ -67,6 +87,10 @@ void print_tok_vec(std::vector &embd) { fileformat = FileFormat::GPTJ1; } + if(vocabsiz==50257) + { + fileformat = FileFormat::GPT2; + } } else if(magic == 0x67676d66) //v2 format ggmf { diff --git a/model_adapter.h b/model_adapter.h index 0814d1d29..fb6976703 100644 --- a/model_adapter.h +++ b/model_adapter.h @@ -36,11 +36,12 @@ enum ModelLoadResult bool llama_load_model(const load_model_inputs inputs, FileFormat file_format); generation_outputs llama_generate(const generation_inputs inputs, generation_outputs &output); -ModelLoadResult gptj_load_model(const load_model_inputs inputs, FileFormat in_file_format); -generation_outputs gptj_generate(const generation_inputs inputs, generation_outputs &output); +ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in_file_format); +generation_outputs gpttype_generate(const generation_inputs inputs, generation_outputs &output); void timer_start(); double timer_check(); void print_tok_vec(std::vector &embd); +void print_tok_vec(std::vector &embd); FileFormat check_file_format(const std::string & fname); \ No newline at end of file diff --git a/otherarch/gpt2_v1.cpp b/otherarch/gpt2_v1.cpp new file mode 100644 index 000000000..fd11d17a2 --- /dev/null +++ b/otherarch/gpt2_v1.cpp @@ -0,0 +1,753 @@ +#include "ggml_v1.h" +#include "otherarch.h" + +#include "utils.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include +#include + + + +// 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 + // in order to save memory and also to speed up the computation + const ggml_v1_type wtype = model.hparams.f16 ? GGML_V1_TYPE_F16 : GGML_V1_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_v1_type_size(GGML_V1_TYPE_F32); // ln_f_g + ctx_size += n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // ln_f_b + + ctx_size += n_vocab*n_embd*ggml_v1_type_size(wtype); // wte + ctx_size += n_ctx*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // wpe + + ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_1_g + ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_1_b + + ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_2_g + ctx_size += n_layer*(n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // ln_2_b + + ctx_size += n_layer*(3*n_embd*n_embd*ggml_v1_type_size(wtype)); // c_attn_attn_w + ctx_size += n_layer*( 3*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_attn_attn_b + + ctx_size += n_layer*(n_embd*n_embd*ggml_v1_type_size(wtype)); // c_attn_proj_w + ctx_size += n_layer*( n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_attn_proj_b + + ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_size(wtype)); // c_mlp_fc_w + ctx_size += n_layer*( 4*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_mlp_fc_b + + ctx_size += n_layer*(4*n_embd*n_embd*ggml_v1_type_size(wtype)); // c_mlp_proj_w + ctx_size += n_layer*( n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32)); // c_mlp_proj_b + + ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(GGML_V1_TYPE_F32); // memory_k + ctx_size += n_ctx*n_layer*n_embd*ggml_v1_type_size(GGML_V1_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_v1_init_params params = { + .mem_size = ctx_size, + .mem_buffer = NULL, + }; + + model.ctx = ggml_v1_init(params); + if (!model.ctx) { + fprintf(stderr, "%s: ggml_v1_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_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd); + model.ln_f_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd); + + model.wte = ggml_v1_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + model.wpe = ggml_v1_new_tensor_2d(ctx, GGML_V1_TYPE_F32, n_embd, n_ctx); + + // 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; + + for (int i = 0; i < n_layer; ++i) { + auto & layer = model.layers[i]; + + layer.ln_1_g = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd); + layer.ln_1_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd); + + layer.ln_2_g = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd); + layer.ln_2_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd); + + layer.c_attn_attn_w = ggml_v1_new_tensor_2d(ctx, wtype, 3*n_embd, n_embd); + layer.c_attn_attn_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, 3*n_embd); + + layer.c_attn_proj_w = ggml_v1_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.c_attn_proj_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_embd); + + layer.c_mlp_fc_w = ggml_v1_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd); + layer.c_mlp_fc_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, 4*n_embd); + + layer.c_mlp_proj_w_trans = ggml_v1_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd); + layer.c_mlp_proj_b = ggml_v1_new_tensor_1d(ctx, GGML_V1_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_trans; + 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_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_elements); + model.memory_v = ggml_v1_new_tensor_1d(ctx, GGML_V1_TYPE_F32, n_elements); + + const size_t memory_size = ggml_v1_nbytes(model.memory_k) + ggml_v1_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; + + 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_v1_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; + } + + const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_v1_fp16_t); + + if (nelements*bpe != ggml_v1_nbytes(tensor)) { + fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n", + __func__, name.data(), ggml_v1_nbytes(tensor), nelements*bpe); + return ModelLoadResult::FAIL; + } + + fin.read(reinterpret_cast(tensor->data), ggml_v1_nbytes(tensor)); + + //printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_v1_nbytes(tensor)/1024.0/1024.0); + total_size += ggml_v1_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_v1_init_params params = { + .mem_size = buf_size, + .mem_buffer = buf, + }; + + struct ggml_v1_context * ctx0 = ggml_v1_init(params); + struct ggml_v1_cgraph gf = { .n_threads = n_threads }; + + struct ggml_v1_tensor * embd = ggml_v1_new_tensor_1d(ctx0, GGML_V1_TYPE_I32, N); + memcpy(embd->data, embd_inp.data(), N*ggml_v1_element_size(embd)); + + struct ggml_v1_tensor * position = ggml_v1_new_tensor_1d(ctx0, GGML_V1_TYPE_I32, N); + for (int i = 0; i < N; ++i) { + ((int32_t *) position->data)[i] = n_past + i; + } + + // wte + wpe + struct ggml_v1_tensor * inpL = + ggml_v1_add(ctx0, + ggml_v1_get_rows(ctx0, model.wte, embd), + ggml_v1_get_rows(ctx0, model.wpe, position)); + + for (int il = 0; il < n_layer; ++il) { + struct ggml_v1_tensor * cur; + + // norm + { + // [ 768, N] + cur = ggml_v1_norm(ctx0, inpL); + + // cur = ln_1_g*cur + ln_1_b + // [ 768, N] + cur = ggml_v1_add(ctx0, + ggml_v1_mul(ctx0, + ggml_v1_repeat(ctx0, model.layers[il].ln_1_g, cur), + cur), + ggml_v1_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_v1_mul_mat(ctx0, + ggml_v1_transpose(ctx0, model.layers[il].c_attn_attn_w), + cur); + + cur = ggml_v1_add(ctx0, + ggml_v1_repeat(ctx0, model.layers[il].c_attn_attn_b, cur), + cur); + } + + // self-attention + { + struct ggml_v1_tensor * Qcur = ggml_v1_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd); + struct ggml_v1_tensor * Kcur = ggml_v1_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd); + struct ggml_v1_tensor * Vcur = ggml_v1_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_v1_tensor * k = ggml_v1_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_v1_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past)); + struct ggml_v1_tensor * v = ggml_v1_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_v1_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past)); + + ggml_v1_build_forward_expand(&gf, ggml_v1_cpy(ctx0, Kcur, k)); + ggml_v1_build_forward_expand(&gf, ggml_v1_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_v1_tensor * Q = + ggml_v1_permute(ctx0, + ggml_v1_cpy(ctx0, + Qcur, + ggml_v1_new_tensor_3d(ctx0, GGML_V1_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_v1_tensor * K = + ggml_v1_permute(ctx0, + ggml_v1_reshape_3d(ctx0, + ggml_v1_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_v1_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_v1_tensor * V = + // ggml_v1_cpy(ctx0, + // ggml_v1_permute(ctx0, + // ggml_v1_reshape_3d(ctx0, + // ggml_v1_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_v1_element_size(model.memory_v)*n_embd), + // n_embd/n_head, n_head, n_past + N), + // 1, 2, 0, 3), + // ggml_v1_new_tensor_3d(ctx0, GGML_V1_TYPE_F32, n_past + N, n_embd/n_head, n_head)); + + //struct ggml_v1_tensor * KQV = ggml_v1_flash_attn(ctx0, Q, K, V, true); + + // K * Q + // [n_past + N, N, 12] + struct ggml_v1_tensor * KQ = ggml_v1_mul_mat(ctx0, K, Q); + + // KQ_scaled = KQ / sqrt(n_embd/n_head) + // [n_past + N, N, 12] + struct ggml_v1_tensor * KQ_scaled = + ggml_v1_scale(ctx0, + KQ, + ggml_v1_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head)) + ); + + // KQ_masked = mask_past(KQ_scaled) + // [n_past + N, N, 12] + struct ggml_v1_tensor * KQ_masked = ggml_v1_diag_mask_inf(ctx0, KQ_scaled, n_past); + + // KQ = soft_max(KQ_masked) + // [n_past + N, N, 12] + struct ggml_v1_tensor * KQ_soft_max = ggml_v1_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_v1_tensor * V_trans = + ggml_v1_permute(ctx0, + ggml_v1_reshape_3d(ctx0, + ggml_v1_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_v1_element_size(model.memory_v)*n_embd), + n_embd/n_head, n_head, n_past + N), + 1, 2, 0, 3); + + // KQV = transpose(V) * KQ_soft_max + // [64, N, 12] + struct ggml_v1_tensor * KQV = ggml_v1_mul_mat(ctx0, V_trans, KQ_soft_max); + + // KQV_merged = KQV.permute(0, 2, 1, 3) + // [64, 12, N] + struct ggml_v1_tensor * KQV_merged = ggml_v1_permute(ctx0, KQV, 0, 2, 1, 3); + + // cur = KQV_merged.contiguous().view(n_embd, N) + // [768, N] + cur = ggml_v1_cpy(ctx0, + KQV_merged, + ggml_v1_new_tensor_2d(ctx0, GGML_V1_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_v1_mul_mat(ctx0, + ggml_v1_transpose(ctx0, model.layers[il].c_attn_proj_w), + cur); + + cur = ggml_v1_add(ctx0, + ggml_v1_repeat(ctx0, model.layers[il].c_attn_proj_b, cur), + cur); + } + + // add the input + cur = ggml_v1_add(ctx0, cur, inpL); + + struct ggml_v1_tensor * inpFF = cur; + + // feed-forward network + { + // norm + { + cur = ggml_v1_norm(ctx0, inpFF); + + // cur = ln_2_g*cur + ln_2_b + // [ 768, N] + cur = ggml_v1_add(ctx0, + ggml_v1_mul(ctx0, + ggml_v1_repeat(ctx0, model.layers[il].ln_2_g, cur), + cur), + ggml_v1_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_v1_mul_mat(ctx0, + ggml_v1_transpose(ctx0, model.layers[il].c_mlp_fc_w), + cur); + + cur = ggml_v1_add(ctx0, + ggml_v1_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur), + cur); + + // GELU activation + // [3072, N] + cur = ggml_v1_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_v1_mul_mat(ctx0, + model.layers[il].c_mlp_proj_w_trans, + cur); + + cur = ggml_v1_add(ctx0, + ggml_v1_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur), + cur); + } + + // input for next layer + inpL = ggml_v1_add(ctx0, cur, inpFF); + } + + // norm + { + // [ 768, N] + inpL = ggml_v1_norm(ctx0, inpL); + + // inpL = ln_f_g*inpL + ln_f_b + // [ 768, N] + inpL = ggml_v1_add(ctx0, + ggml_v1_mul(ctx0, + ggml_v1_repeat(ctx0, model.ln_f_g, inpL), + inpL), + ggml_v1_repeat(ctx0, model.ln_f_b, inpL)); + } + + // inpL = WTE * inpL + // [ 768, 50257] - model.wte + // [ 768, N] - inpL + inpL = ggml_v1_mul_mat(ctx0, model.wte, inpL); + + // logits -> probs + //inpL = ggml_v1_soft_max(ctx0, inpL); + + // run the computation + ggml_v1_build_forward_expand(&gf, inpL); + ggml_v1_graph_compute (ctx0, &gf); + + //if (n_past%100 == 0) { + // ggml_v1_graph_print (&gf); + // ggml_v1_graph_dump_dot(&gf, NULL, "gpt-2.dot"); + //} + + //embd_w.resize(n_vocab*N); + //memcpy(embd_w.data(), ggml_v1_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_v1_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab); + + if (mem_per_token == 0) { + mem_per_token = ggml_v1_used_mem(ctx0)/N; + } + //printf("used_mem = %zu\n", ggml_v1_used_mem(ctx0)); + + ggml_v1_free(ctx0); + + return true; +} + +// int main(int argc, char ** argv) { +// ggml_v1_time_init(); +// const int64_t t_main_start_us = ggml_v1_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_v1_time_us(); + +// if (!gpt2_model_load(params.model, model, vocab, FileFormat::GPT2)) { +// fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); +// return 1; +// } + +// t_load_us = ggml_v1_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: number of tokens in prompt = %zu\n", __func__, embd_inp.size()); +// printf("\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, FileFormat::GPT2); + +// 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)) { +// printf("Failed to predict\n"); +// return 1; +// } + +// t_predict_us += ggml_v1_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_v1_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_v1_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_v1_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_v1_free(model.ctx); + +// return 0; +// } \ No newline at end of file diff --git a/otherarch/otherarch.h b/otherarch/otherarch.h index d07257862..1b734e5a5 100644 --- a/otherarch/otherarch.h +++ b/otherarch/otherarch.h @@ -113,6 +113,60 @@ struct gptj_model { std::map tensors; }; +// default hparams (GPT-2 117M) +struct gpt2_hparams { + int32_t n_vocab = 50257; + int32_t n_ctx = 1024; + int32_t n_embd = 768; + int32_t n_head = 12; + int32_t n_layer = 12; + int32_t f16 = 1; +}; + +struct gpt2_layer { + // normalization + struct ggml_v1_tensor * ln_1_g; + struct ggml_v1_tensor * ln_1_b; + + struct ggml_v1_tensor * ln_2_g; + struct ggml_v1_tensor * ln_2_b; + + // attention + struct ggml_v1_tensor * c_attn_attn_w; + struct ggml_v1_tensor * c_attn_attn_b; + + struct ggml_v1_tensor * c_attn_proj_w; + struct ggml_v1_tensor * c_attn_proj_b; + + // mlp + struct ggml_v1_tensor * c_mlp_fc_w; + struct ggml_v1_tensor * c_mlp_fc_b; + + struct ggml_v1_tensor * c_mlp_proj_w_trans; // transposed for efficiency + struct ggml_v1_tensor * c_mlp_proj_b; +}; + +struct gpt2_model { + gpt2_hparams hparams; + + // normalization + struct ggml_v1_tensor * ln_f_g; + struct ggml_v1_tensor * ln_f_b; + + struct ggml_v1_tensor * wte; // position embedding + struct ggml_v1_tensor * wpe; // token embedding + + std::vector layers; + + // key + value memory + struct ggml_v1_tensor * memory_k; + struct ggml_v1_tensor * memory_v; + + // + struct ggml_v1_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);