diff --git a/Makefile b/Makefile index 6942fdbcc..781a64583 100644 --- a/Makefile +++ b/Makefile @@ -1,6 +1,6 @@ default: koboldcpp koboldcpp_failsafe koboldcpp_openblas koboldcpp_openblas_noavx2 koboldcpp_clblast simple: koboldcpp koboldcpp_failsafe -tools: quantize_gpt2 quantize_gptj quantize_llama quantize_neox +tools: quantize_gpt2 quantize_gptj quantize_llama quantize_neox quantize_mpt dev: koboldcpp_openblas dev2: koboldcpp_clblast @@ -281,7 +281,7 @@ gpttype_adapter_clblast.o: gpttype_adapter.cpp $(CXX) $(CXXFLAGS) $(CLBLAST_FLAGS) -c $< -o $@ clean: - rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize_neox quantize-stats perplexity embedding benchmark-matmult save-load-state main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe quantize_neox.exe koboldcpp.dll koboldcpp_openblas.dll koboldcpp_failsafe.dll koboldcpp_openblas_noavx2.dll koboldcpp_clblast.dll koboldcpp.so koboldcpp_openblas.so koboldcpp_failsafe.so koboldcpp_openblas_noavx2.so koboldcpp_clblast.so gptj.exe gpt2.exe + rm -vf *.o main quantize_llama quantize_gpt2 quantize_gptj quantize_neox quantize_mpt quantize-stats perplexity embedding benchmark-matmult save-load-state main.exe quantize_llama.exe quantize_gptj.exe quantize_gpt2.exe quantize_neox.exe quantize_mpt.exe koboldcpp.dll koboldcpp_openblas.dll koboldcpp_failsafe.dll koboldcpp_openblas_noavx2.dll koboldcpp_clblast.dll koboldcpp.so koboldcpp_openblas.so koboldcpp_failsafe.so koboldcpp_openblas_noavx2.so koboldcpp_clblast.so main: examples/main/main.cpp build-info.h ggml.o llama.o common.o $(OBJS) $(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS) @@ -308,6 +308,8 @@ quantize_gpt2: ggml.o llama.o otherarch/tools/gpt2_quantize.cpp otherarch/tools/ $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) quantize_neox: ggml.o llama.o otherarch/tools/neox_quantize.cpp otherarch/tools/common-ggml.cpp $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) +quantize_mpt: ggml.o llama.o otherarch/tools/mpt_quantize.cpp otherarch/tools/common-ggml.cpp + $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) quantize-stats: examples/quantize-stats/quantize-stats.cpp ggml.o llama.o $(OBJS) $(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS) perplexity: examples/perplexity/perplexity.cpp ggml.o llama.o common.o $(OBJS) diff --git a/expose.cpp b/expose.cpp index c2ae30767..fffa978d0 100644 --- a/expose.cpp +++ b/expose.cpp @@ -175,6 +175,19 @@ extern "C" return true; } } + else if(file_format==FileFormat::MPT_1) + { + printf("\n---\nIdentified as MPT 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); diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp index 483a1f957..db48b2cad 100644 --- a/gpttype_adapter.cpp +++ b/gpttype_adapter.cpp @@ -26,6 +26,7 @@ #include "rwkv_v3.cpp" #include "neox_v2.cpp" #include "neox_v3.cpp" +#include "mpt_v3.cpp" //return val: 0=fail, 1=(original ggml, alpaca), 2=(ggmf), 3=(ggjt) @@ -44,6 +45,8 @@ static gpt2_model gpt2_ctx_v3; static gpt_neox_v2_model neox_ctx_v2; static gpt_neox_model neox_ctx_v3; +static mpt_model mpt_ctx_v3; + static rwkv_v2_context * rwkv_ctx_v2; static rwkv_context * rwkv_ctx_v3; static llama_v2_context_params llama_ctx_params_v2; @@ -298,7 +301,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in params.n_ctx = inputs.max_context_length; neox_ctx_v2.hparams.n_ctx = gptj_ctx_v1.hparams.n_ctx = gptj_ctx_v2.hparams.n_ctx = gpt2_ctx_v1.hparams.n_ctx = gpt2_ctx_v2.hparams.n_ctx - = neox_ctx_v3.hparams.n_ctx = gptj_ctx_v3.hparams.n_ctx = gptj_ctx_v3.hparams.n_ctx = params.n_ctx; + = neox_ctx_v3.hparams.n_ctx = gptj_ctx_v3.hparams.n_ctx = gptj_ctx_v3.hparams.n_ctx = mpt_ctx_v3.hparams.n_ctx = params.n_ctx; printf("System Info: %s\n", llama_print_system_info()); SetQuantsUnshuffled(false); @@ -682,6 +685,19 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in } } + else if(file_format==FileFormat::MPT_1) + { + bool res = mpt_model_load(params.model, mpt_ctx_v3, vocab); + if(res==false) + { + fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str()); + return ModelLoadResult::FAIL; + } + + // determine the required inference memory per token: + mpt_eval(mpt_ctx_v3, params.n_threads, 0, { 0, 1, 2, 3 }, logits, false, mem_per_token); + return ModelLoadResult::SUCCESS; + } else { printf("\nUnknown Model, cannot load.\n"); @@ -869,6 +885,10 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o { n_vocab = neox_ctx_v3.hparams.n_vocab; } + else if( file_format==FileFormat::MPT_1) + { + n_vocab = mpt_ctx_v3.hparams.n_vocab; + } else if(file_format == FileFormat::RWKV_1 || file_format==FileFormat::RWKV_2) { n_vocab = vocab.id_to_token.size(); //handled seperately @@ -1006,6 +1026,10 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o { evalres = gptj_eval(gptj_ctx_v3, params.n_threads, n_past, embd, logits, mem_per_token); } + else if(file_format==FileFormat::MPT_1) + { + evalres = mpt_eval(mpt_ctx_v3, params.n_threads, n_past, embd, logits, false, mem_per_token); + } else { printf("\nCannot find eval function\n"); @@ -1098,7 +1122,8 @@ generation_outputs gpttype_generate(const generation_inputs inputs, generation_o file_format == FileFormat::NEOX_4 || file_format == FileFormat::NEOX_5 || file_format == FileFormat::NEOX_6 || - file_format == FileFormat::NEOX_7) + file_format == FileFormat::NEOX_7 || + file_format == FileFormat::MPT_1) { eosID = 0; int topid = std::min_element(logits.begin(),logits.end())-logits.begin(); diff --git a/model_adapter.cpp b/model_adapter.cpp index 042745912..29d12809b 100644 --- a/model_adapter.cpp +++ b/model_adapter.cpp @@ -98,7 +98,11 @@ void print_tok_vec(std::vector &embd) //we need to read more to determine int32_t vocabsiz = 0; fin.read((char *) &vocabsiz, sizeof(int32_t)); - if(vocabsiz==50400) //know GPT-J vocab size + if(vocabsiz==4096) //actually the d_model for mpt + { + fileformat = FileFormat::MPT_1; + } + else if(vocabsiz==50400) //know GPT-J vocab size { fileformat = FileFormat::GPTJ_1; uint32_t temp; diff --git a/model_adapter.h b/model_adapter.h index ab6769910..523b5b828 100644 --- a/model_adapter.h +++ b/model_adapter.h @@ -43,6 +43,8 @@ enum FileFormat NEOX_5=404, //unshuffled redpajama NEOX_6=405, //using 16bit scalar NEOX_7=406, //using 16bit scalar redpajama + + MPT_1=500, //first supported mpt version }; enum ModelLoadResult diff --git a/otherarch/mpt_v3.cpp b/otherarch/mpt_v3.cpp new file mode 100644 index 000000000..3fe1a6262 --- /dev/null +++ b/otherarch/mpt_v3.cpp @@ -0,0 +1,516 @@ +#include "ggml.h" +#include "otherarch.h" + +#include "utils.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +#include "model_adapter.h" + + + +// load the model's weights from a file +bool mpt_model_load(const std::string & fname, mpt_model & model, gpt_vocab & vocab) { + printf("%s: loading model from '%s' - please wait ...\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 false; + } + + // 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.d_model, sizeof(hparams.d_model)); + fin.read((char *) &hparams.max_seq_len, sizeof(hparams.max_seq_len)); + fin.read((char *) &hparams.n_heads, sizeof(hparams.n_heads)); + fin.read((char *) &hparams.n_layers, sizeof(hparams.n_layers)); + fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); + fin.read((char *) &hparams.alibi_bias_max, sizeof(hparams.alibi_bias_max)); + fin.read((char *) &hparams.clip_qkv, sizeof(hparams.clip_qkv)); + fin.read((char *) &hparams.ftype, sizeof(hparams.ftype)); + + hparams.n_ctx = std::min(hparams.max_seq_len, hparams.n_ctx); + + const int32_t qntvr = hparams.ftype / GGML_QNT_VERSION_FACTOR; + + printf("%s: d_model = %d\n", __func__, hparams.d_model); + printf("%s: max_seq_len = %d\n", __func__, hparams.max_seq_len); + printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx); + printf("%s: n_heads = %d\n", __func__, hparams.n_heads); + printf("%s: n_layers = %d\n", __func__, hparams.n_layers); + printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab); + printf("%s: alibi_bias_max = %f\n", __func__, hparams.alibi_bias_max); + printf("%s: clip_qkv = %f\n", __func__, hparams.clip_qkv); + printf("%s: ftype = %d\n", __func__, hparams.ftype); + printf("%s: qntvr = %d\n", __func__, qntvr); + + hparams.ftype %= GGML_QNT_VERSION_FACTOR; + } + + // load vocab + { + const int32_t n_vocab = model.hparams.n_vocab; + + std::string word; + std::vector buf(128); + + for (int i = 0; i < n_vocab; i++) { + uint32_t len; + fin.read((char *) &len, sizeof(len)); + + buf.resize(len); + fin.read((char *) buf.data(), len); + word.assign(buf.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_ftype_to_ggml_type((ggml_ftype)(model.hparams.ftype)); + if (wtype == GGML_TYPE_COUNT) { + fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n", __func__, fname.c_str(), + model.hparams.ftype); + return false; + } + + auto & ctx = model.ctx; + + size_t ctx_size = 0; + + const auto & hparams = model.hparams; + const size_t n_ctx = hparams.n_ctx; + + { + const size_t n_embd = hparams.d_model; + const size_t n_layer = hparams.n_layers; + const size_t n_vocab = hparams.n_vocab; + + ctx_size += n_embd * n_vocab * ggml_type_sizef(wtype); // wte_weight + ctx_size += n_embd * ggml_type_sizef(GGML_TYPE_F32); // norm_f_weight + + ctx_size += n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32)); // ln_1_weight + ctx_size += n_layer * (3 * n_embd * n_embd * ggml_type_sizef(wtype)); // attn_Wqkv_weight + ctx_size += n_layer * (n_embd * n_embd * ggml_type_sizef(wtype)); // attn_out_proj_weight + ctx_size += n_layer * (n_embd * ggml_type_sizef(GGML_TYPE_F32)); // ln_2_weight + ctx_size += n_layer * (4 * n_embd * n_embd * ggml_type_sizef(wtype)); // mlp_mlp_up_weight + ctx_size += n_layer * (n_embd * n_embd * 4 * ggml_type_sizef(wtype)); // mlp_mlp_down_weight + + ctx_size += n_ctx * n_layer * n_embd * ggml_type_sizef(GGML_TYPE_F16); // memory_k + ctx_size += n_ctx * n_layer * n_embd * ggml_type_sizef(GGML_TYPE_F16); // memory_v + + ctx_size += (1 + 6 * n_layer) * 512; // 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, + .no_alloc = false, + }; + + 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 size_t n_embd = hparams.d_model; + const size_t n_layer = hparams.n_layers; + const size_t n_vocab = hparams.n_vocab; + + model.layers.resize(n_layer); + + model.wte_weight = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab); + model.norm_f_weight = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + + // map by name + model.tensors["transformer.wte.weight"] = model.wte_weight; + model.tensors["transformer.norm_f.weight"] = model.norm_f_weight; + + for (int i = 0; i < (int) n_layer; ++i) { + auto & layer = model.layers[i]; + + layer.norm_1_weight = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + layer.c_attn_wqkv_weight = ggml_new_tensor_2d(ctx, wtype, n_embd, 3 * n_embd); + layer.c_attn_out_proj_weight = ggml_new_tensor_2d(ctx, wtype, n_embd, n_embd); + layer.norm_2_weight = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd); + layer.ffn_up_proj = ggml_new_tensor_2d(ctx, wtype, n_embd, 4 * n_embd); + layer.ffn_down_proj = ggml_new_tensor_2d(ctx, wtype, 4 * n_embd, n_embd); + + // map by name + model.tensors["transformer.blocks." + std::to_string(i) + ".norm_1.weight"] = layer.norm_1_weight; + model.tensors["transformer.blocks." + std::to_string(i) + ".attn.Wqkv.weight"] = layer.c_attn_wqkv_weight; + model.tensors["transformer.blocks." + std::to_string(i) + ".attn.out_proj.weight"] = layer.c_attn_out_proj_weight; + model.tensors["transformer.blocks." + std::to_string(i) + ".norm_2.weight"] = layer.norm_2_weight; + model.tensors["transformer.blocks." + std::to_string(i) + ".ffn.up_proj.weight"] = layer.ffn_up_proj; + model.tensors["transformer.blocks." + std::to_string(i) + ".ffn.down_proj.weight"] = layer.ffn_down_proj; + } + } + + // key + value memory + { + const auto & hparams = model.hparams; + + const size_t n_embd = hparams.d_model; + const size_t n_layer = hparams.n_layers; + + const int64_t n_mem = n_layer * n_ctx; + const int64_t n_elements = n_embd * n_mem; + + model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, n_elements); + model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F16, 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 = %" PRId64 "\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 ttype; + + fin.read(reinterpret_cast(&n_dims), sizeof(n_dims)); + fin.read(reinterpret_cast(&length), sizeof(length)); + fin.read(reinterpret_cast(&ttype), sizeof(ttype)); + + 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 [%5d, " + "%5d], expected [%5d, %5d]\n", + __func__, name.data(), (int)tensor->ne[0], (int)tensor->ne[1], ne[0], ne[1]); + return false; + } + + // for debugging + if (0) { + printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], + ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor) / 1024.0 / 1024.0, ggml_nbytes(tensor)); + } + + const size_t bpe = ggml_type_size(ggml_type(ttype)); + + 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)); + + 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); + } + + fin.close(); + + return true; +} + +// 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 mpt_eval(const mpt_model & model, const int n_threads, const int n_past, + const std::vector & embd_inp, std::vector & embd_w, bool logits_all, size_t & mem_per_token) { + const int N = embd_inp.size(); + + const auto & hparams = model.hparams; + + const int n_embd = hparams.d_model; + const int n_layer = hparams.n_layers; + const int n_head = hparams.n_heads; + const int n_vocab = hparams.n_vocab; + const int n_ctx = hparams.n_ctx; + + static size_t buf_size = 256u * 1024 * 1024; + static void * buf = malloc(buf_size); + + // use 2 scratch buffers + // TODO: very hacky solution - reimplement in a more elegant way + static size_t scr0_size = 256u*1024*1024; + static void * scr0 = malloc(scr0_size); + + static size_t scr1_size = 256u*1024*1024; + static void * scr1 = malloc(scr1_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; + params.mem_size = buf_size; + params.mem_buffer = buf; + params.no_alloc = false; + + 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 * inpL = ggml_get_rows(ctx0, model.wte_weight, embd); + + for (int il = 0; il < n_layer; ++il) { + + struct ggml_tensor * cur; + + ggml_set_scratch(ctx0, { 0, scr0_size, scr0, }); + + // a = self.ln_1(x) + { + cur = ggml_norm(ctx0, inpL); + + cur = ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].norm_1_weight, cur), cur); + } + + // self-attention + // b, _, past_key_value = self.attn(a, past_key_value=past_key_value, + // attn_bias=attn_bias, attention_mask=attention_mask, + // is_causal=is_causal) + { + // compute QKV + cur = ggml_mul_mat(ctx0, model.layers[il].c_attn_wqkv_weight, cur); + + if (model.hparams.clip_qkv > 0.0f) { + cur = ggml_clamp(ctx0, cur, -model.hparams.clip_qkv, model.hparams.clip_qkv); + } + + 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 + { + 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); + // 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))); + + struct ggml_tensor * KQ_scaled_alibi = + ggml_alibi(ctx0, KQ_scaled, n_past, n_head, model.hparams.alibi_bias_max); + + // KQ_masked = mask_past(KQ_scaled) + struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled_alibi, 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() [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 + 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 + { cur = ggml_mul_mat(ctx0, model.layers[il].c_attn_out_proj_weight, cur); } + } + + inpL = ggml_add(ctx0, inpL, cur); + + ggml_set_scratch(ctx0, { 0, scr1_size, scr1, }); + + // m = self.ln_2(x) + { + cur = ggml_norm(ctx0, inpL); + + cur = ggml_mul(ctx0, ggml_repeat(ctx0, model.layers[il].norm_2_weight, cur), cur); + } + + // n = self.mlp(m) + { + + cur = ggml_mul_mat(ctx0, model.layers[il].ffn_up_proj, cur); + + // GELU activation + cur = ggml_gelu(ctx0, cur); + + // projection + // cur = proj_w*cur + proj_b + cur = ggml_mul_mat(ctx0, model.layers[il].ffn_down_proj, cur); + } + + // x = x + n + inpL = ggml_add(ctx0, inpL, cur); + } + + ggml_set_scratch(ctx0, { 0, scr0_size, scr0, }); + + // norm + { + inpL = ggml_norm(ctx0, inpL); + // inpL = ln_f_g*inpL + inpL = ggml_mul(ctx0, ggml_repeat(ctx0, model.norm_f_weight, inpL), inpL); + } + + ggml_set_scratch(ctx0, { 0, 0, nullptr, }); + + // output embedding weight tied to input embedding + inpL = ggml_mul_mat(ctx0, model.wte_weight, inpL); + + // logits -> probs + // inpL = ggml_soft_max(ctx0, inpL); + + // run the computation + ggml_build_forward_expand(&gf, inpL); + ggml_graph_compute(ctx0, &gf); + + // std::cout << "Qcur" << std::endl; + // print_tensor(Qcur); + + // if (n_past%100 == 0) { + // ggml_graph_print(&gf); + // ggml_graph_dump_dot(&gf, NULL, "mpt-model.dot"); + // } + + if (logits_all) { + // return result for all tokens + embd_w.resize(n_vocab *N); + memcpy(embd_w.data(), (float *)ggml_get_data(inpL) , sizeof(float) * n_vocab * N); + } else { + // 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; +} diff --git a/otherarch/otherarch.h b/otherarch/otherarch.h index 83e558270..469315066 100644 --- a/otherarch/otherarch.h +++ b/otherarch/otherarch.h @@ -407,3 +407,50 @@ struct gpt_neox_model { }; +// no defaults for now +struct mpt_hparams { + int32_t d_model = 0; + int32_t max_seq_len = 0; + int32_t n_heads = 0; + int32_t n_layers = 0; + int32_t n_vocab = 0; + float alibi_bias_max = 0; + float clip_qkv = 0; + int32_t ftype = 0; + int32_t n_ctx = 0; + +}; + +struct mpt_layer { + // pre normalization + struct ggml_tensor * norm_1_weight; + + // attention + struct ggml_tensor * c_attn_wqkv_weight; + struct ggml_tensor * c_attn_out_proj_weight; + + // post normalization + struct ggml_tensor * norm_2_weight; + + // ff + struct ggml_tensor * ffn_up_proj; + struct ggml_tensor * ffn_down_proj; +}; + +struct mpt_model { + mpt_hparams hparams; + + struct ggml_tensor * wte_weight; // position embedding + struct ggml_tensor * norm_f_weight; // 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; +}; + + diff --git a/otherarch/tools/convert_hf_mpt.py b/otherarch/tools/convert_hf_mpt.py new file mode 100644 index 000000000..fd21dfb62 --- /dev/null +++ b/otherarch/tools/convert_hf_mpt.py @@ -0,0 +1,158 @@ +import sys +import struct +import json +import numpy as np +from transformers import AutoModelForCausalLM, AutoTokenizer +import sentencepiece.sentencepiece_model_pb2 as model + +# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py +def bytes_to_unicode(): + + """ + Returns list of utf-8 byte and a corresponding list of unicode strings. + The reversible bpe codes work on unicode strings. + This means you need a large # of unicode characters in your vocab if you want to avoid UNKs. + When you're at something like a 10B token dataset you end up needing around 5K for decent coverage. + This is a signficant percentage of your normal, say, 32K bpe vocab. + To avoid that, we want lookup tables between utf-8 bytes and unicode strings. + And avoids mapping to whitespace/control characters the bpe code barfs on. + """ + bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1)) + cs = bs[:] + n = 0 + for b in range(2**8): + if b not in bs: + bs.append(b) + cs.append(2**8+n) + n += 1 + + cs = [chr(n) for n in cs] + + return dict(zip(bs, cs)) + +if len(sys.argv) < 3: + print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n") + print(" ftype == 0 -> float32") + print(" ftype == 1 -> float16") + sys.exit(1) + + +# output in the same directory as the model +dir_model = sys.argv[1] +fname_out = sys.argv[1] + "/ggml-model.bin" + + +with open(dir_model + "/config.json", "r", encoding="utf-8") as f: + hparams = json.load(f) + +# possible data types +# ftype == 0 -> float32 +# ftype == 1 -> float16 +# +# map from ftype to string +ftype_str = ["f32", "f16"] + +ftype = 1 +if len(sys.argv) > 2: + ftype = int(sys.argv[2]) + if ftype < 0 or ftype > 1: + print("Invalid ftype: " + str(ftype)) + sys.exit(1) + fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin" + + +tokenizer = AutoTokenizer.from_pretrained(dir_model, trust_remote_code=True) +model = AutoModelForCausalLM.from_pretrained( + dir_model, low_cpu_mem_usage=True, trust_remote_code=True +) +# print (model) + +# print(tokenizer.encode('I believe the meaning of life is')) + +list_vars = model.state_dict() +for name in list_vars.keys(): + print(name, list_vars[name].shape, list_vars[name].dtype) + +fout = open(fname_out, "wb") + +print(hparams) + +fout.write(struct.pack("i", 0x67676D6C)) # magic: ggml in hex +fout.write(struct.pack("i", hparams["d_model"])) +fout.write(struct.pack("i", hparams["max_seq_len"])) +fout.write(struct.pack("i", hparams["n_heads"])) +fout.write(struct.pack("i", hparams["n_layers"])) +fout.write(struct.pack("i", hparams["vocab_size"])) +fout.write(struct.pack("f", hparams["attn_config"]["alibi_bias_max"])) +fout.write(struct.pack("f", hparams["attn_config"]["clip_qkv"] or 0.0)) +fout.write(struct.pack("i", ftype)) + +vocab_size = hparams["vocab_size"] + +encoder = tokenizer.vocab +# Add added_tokens (special tokens) to the encoder +encoder.update(tokenizer.get_added_vocab()) + +byte_encoder = bytes_to_unicode() +byte_decoder = {v:k for k, v in byte_encoder.items()} + +counter = 0 +# sort by value +for key in sorted(encoder, key=encoder.get): + # workaround for key error when c not found + text="" + for c in key: + if c not in byte_decoder: + text += c + else: + text += chr(byte_decoder[c] ) + text = bytearray( text, encoding="utf-8" ) + fout.write(struct.pack("i", len(text))) + fout.write(text) + counter += 1 + +# Repeat last token until vocab_size +while counter < vocab_size: + fout.write(struct.pack("i", len(text))) + fout.write(text) + counter += 1 + +# assert counter == config.vocab_size + +for name in list_vars.keys(): + data = list_vars[name].squeeze().numpy() + print("Processing variable: " + name + " with shape: ", data.shape) + + n_dims = len(data.shape) + + # ftype == 0 -> float32, ftype == 1 -> float16 + ftype_cur = 0 + if ftype != 0: + if name[-7:] == ".weight" and n_dims == 2: + print(" Converting to float16") + data = data.astype(np.float16) + ftype_cur = 1 + else: + print(" Converting to float32") + data = data.astype(np.float32) + ftype_cur = 0 + else: + if data.dtype != np.float32: + print(" Converting to float32") + data = data.astype(np.float32) + ftype_cur = 0 + + # header + str = name.encode("utf-8") + fout.write(struct.pack("iii", n_dims, len(str), ftype_cur)) + for i in range(n_dims): + fout.write(struct.pack("i", data.shape[n_dims - 1 - i])) + fout.write(str) + + # data + data.tofile(fout) + +fout.close() + +print("Done. Output file: " + fname_out) +print("") \ No newline at end of file diff --git a/otherarch/tools/mpt_quantize.cpp b/otherarch/tools/mpt_quantize.cpp new file mode 100644 index 000000000..d3d22aee3 --- /dev/null +++ b/otherarch/tools/mpt_quantize.cpp @@ -0,0 +1,184 @@ +#include "utils.h" +#include "common-ggml.h" + +#include +#include +#include +#include +#include +#include +#include +#include +#include + +struct mpt_hparams { + int32_t d_model = 0; + int32_t max_seq_len = 0; + int32_t n_heads = 0; + int32_t n_layers = 0; + int32_t n_vocab = 0; + float alibi_bias_max = 0; + float clip_qkv = 0; + int32_t ftype = 0; +}; + +// quantize a model +bool mpt_model_quantize(const std::string & fname_inp, + const std::string & fname_out, ggml_ftype ftype) { + + printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str()); + + auto finp = std::ifstream(fname_inp, std::ios::binary); + if (!finp) { + fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, + fname_inp.c_str()); + return false; + } + + auto fout = std::ofstream(fname_out, std::ios::binary); + if (!fout) { + fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, + fname_out.c_str()); + return false; + } + + // verify magic + { + uint32_t magic; + finp.read((char *)&magic, sizeof(magic)); + if (magic != 0x67676d6c) { + fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", + __func__, fname_inp.c_str()); + return false; + } + + fout.write((char *)&magic, sizeof(magic)); + } + + mpt_hparams hparams; + + // load hparams + { + finp.read((char *) &hparams.d_model, sizeof(hparams.d_model)); + finp.read((char *) &hparams.max_seq_len, sizeof(hparams.max_seq_len)); + finp.read((char *) &hparams.n_heads, sizeof(hparams.n_heads)); + finp.read((char *) &hparams.n_layers, sizeof(hparams.n_layers)); + finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); + finp.read((char *) &hparams.alibi_bias_max, sizeof(hparams.alibi_bias_max)); + finp.read((char *) &hparams.clip_qkv, sizeof(hparams.clip_qkv)); + finp.read((char *) &hparams.ftype, sizeof(hparams.ftype)); + + const int32_t qntvr_src = hparams.ftype / GGML_QNT_VERSION_FACTOR; + const int32_t ftype_dst = GGML_QNT_VERSION * GGML_QNT_VERSION_FACTOR + ftype; + + printf("%s: d_model = %d\n", __func__, hparams.d_model); + printf("%s: max_seq_len = %d\n", __func__, hparams.max_seq_len); + printf("%s: n_heads = %d\n", __func__, hparams.n_heads); + printf("%s: n_layers = %d\n", __func__, hparams.n_layers); + printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab); + printf("%s: alibi_bias_max = %f\n", __func__, hparams.alibi_bias_max); + printf("%s: clip_qkv = %f\n", __func__, hparams.clip_qkv); + printf("%s: ftype (src) = %d\n", __func__, hparams.ftype); + printf("%s: qntvr (src) = %d\n", __func__, qntvr_src); + printf("%s: ftype (dst) = %d\n", __func__, ftype_dst); + printf("%s: qntvr (dst) = %d\n", __func__, GGML_QNT_VERSION); + + fout.write((char *) &hparams.d_model, sizeof(hparams.d_model)); + fout.write((char *) &hparams.max_seq_len, sizeof(hparams.max_seq_len)); + fout.write((char *) &hparams.n_heads, sizeof(hparams.n_heads)); + fout.write((char *) &hparams.n_layers, sizeof(hparams.n_layers)); + fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab)); + fout.write((char *) &hparams.alibi_bias_max, sizeof(hparams.alibi_bias_max)); + fout.write((char *) &hparams.clip_qkv, sizeof(hparams.clip_qkv)); + fout.write((char *) &ftype_dst, sizeof(ftype_dst)); + } + + // load vocab + { + const int32_t n_vocab = hparams.n_vocab; + + std::string word; + for (int i = 0; i < n_vocab; i++) { + uint32_t len; + finp.read((char *)&len, sizeof(len)); + fout.write((char *)&len, sizeof(len)); + + word.resize(len); + finp.read((char *)word.data(), len); + fout.write((char *)word.data(), len); + } + } + + printf("%s: quantizing tensors\n", __func__); + + // regexes of tensor names to be quantized + const std::vector to_quant = { + ".*weight", + }; + + if (!ggml_common_quantize_0(finp, fout, ftype, to_quant, {})) { + fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, + fname_inp.c_str()); + return false; + } + + finp.close(); + fout.close(); + + return true; +} + +// usage: +// ./mpt-quantize models/mpt/ggml-model.bin +// models/mpt/ggml-model-quant.bin type +// +int main(int argc, char ** argv) { + if (argc != 4) { + fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", + argv[0]); + ggml_print_ftypes(stderr); + return 1; + } + + // needed to initialize f16 tables + { + struct ggml_init_params params = {0, NULL, false}; + struct ggml_context * ctx = ggml_init(params); + ggml_free(ctx); + } + + const std::string fname_inp = argv[1]; + const std::string fname_out = argv[2]; + + const ggml_ftype ftype = ggml_parse_ftype(argv[3]); + + const int64_t t_main_start_us = ggml_time_us(); + + int64_t t_quantize_us = 0; + + // load the model + { + const int64_t t_start_us = ggml_time_us(); + + if (!mpt_model_quantize(fname_inp, fname_out, ggml_ftype(ftype))) { + fprintf(stderr, "%s: failed to quantize model from '%s'\n", + __func__, fname_inp.c_str()); + return 1; + } + + t_quantize_us = ggml_time_us() - t_start_us; + } + + // report timing + { + const int64_t t_main_end_us = ggml_time_us(); + + printf("\n"); + printf("%s: quantize time = %8.2f ms\n", __func__, + t_quantize_us / 1000.0f); + printf("%s: total time = %8.2f ms\n", __func__, + (t_main_end_us - t_main_start_us) / 1000.0f); + } + + return 0; +} \ No newline at end of file diff --git a/otherarch/utils.cpp b/otherarch/utils.cpp index 2226c22d9..e56a2048d 100644 --- a/otherarch/utils.cpp +++ b/otherarch/utils.cpp @@ -1,7 +1,12 @@ #include "utils.h" +#include +#include #include #include +#include +#include +#include @@ -109,24 +114,16 @@ void gpt_vocab::add_special_token(const std::string & token) { special_tokens.push_back(token); } -static void append_utf8(char32_t ch, std::string & out) { - if (ch <= 0x7F) { - out.push_back(static_cast(ch)); - } else if (ch <= 0x7FF) { - out.push_back(static_cast(0xC0 | ((ch >> 6) & 0x1F))); - out.push_back(static_cast(0x80 | (ch & 0x3F))); - } else if (ch <= 0xFFFF) { - out.push_back(static_cast(0xE0 | ((ch >> 12) & 0x0F))); - out.push_back(static_cast(0x80 | ((ch >> 6) & 0x3F))); - out.push_back(static_cast(0x80 | (ch & 0x3F))); - } else if (ch <= 0x10FFFF) { - out.push_back(static_cast(0xF0 | ((ch >> 18) & 0x07))); - out.push_back(static_cast(0x80 | ((ch >> 12) & 0x3F))); - out.push_back(static_cast(0x80 | ((ch >> 6) & 0x3F))); - out.push_back(static_cast(0x80 | (ch & 0x3F))); - } else { - printf("Invalid Unicode code point\n"); - } + +std::string convert_to_utf8(const std::wstring & input) { + std::wstring_convert> converter; + return converter.to_bytes(input); +} + + +std::wstring convert_to_wstring(const std::string & input) { + std::wstring_convert> converter; + return converter.from_bytes(input); } std::vector gpt_tokenize(const gpt_vocab & vocab, const std::string & text) { @@ -162,40 +159,27 @@ std::vector gpt_tokenize(const gpt_vocab & vocab, const std::stri } } - // find the longest tokens that form the words: + // find the longest token that forms each word in words: std::vector tokens; for (const auto & word : words) { - if (word.size() == 0) continue; - - int i = 0; - int n = word.size(); - while (i < n) { - int j = n; - while (j > i) { - auto it = vocab.token_to_id.find(word.substr(i, j-i)); - if (it != vocab.token_to_id.end()) { + for (int i = 0; i < word.size(); ){ + for (int j = word.size() - 1; j >= i; j--){ + auto cand = word.substr(i, j-i+1); + auto it = vocab.token_to_id.find(cand); + if (it != vocab.token_to_id.end()){ // word.substr(i, j-i+1) in vocab tokens.push_back(it->second); - i = j; - j = n; - continue; + i = j + 1; + break; } - --j; - } - if (i == n) { - break; - } - if (j == i) { - auto sub = word.substr(i, 1); - if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) { - tokens.push_back(vocab.token_to_id.at(sub)); - } else { - fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data()); + else if (j == i){ // word.substr(i, 1) has no matching + fprintf(stderr, "%s: unknown token '%s'\n", __func__, word.substr(i, 1).data()); + i++; } - ++i; } } } + return tokens; }