--- /dev/null
+#include "ggml/ggml.h"
+
+#include "utils.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <iostream>
+#include <unistd.h>
+
+// default hparams (StableLM 3B)
+struct stablelm_hparams {
+ int32_t n_vocab = 50257;
+ int32_t n_ctx = 4096;
+ int32_t n_embd = 4096;
+ int32_t n_head = 32;
+ int32_t n_layer = 16;
+ int32_t n_rot = 32; // rotary_pct * (n_embd / n_head)
+ int32_t ftype = 1;
+};
+
+struct stablelm_layer {
+ // pre normalization
+ struct ggml_tensor * ln_1_g;
+ struct ggml_tensor * ln_1_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;
+
+ // post normalization
+ struct ggml_tensor * ln_2_g;
+ struct ggml_tensor * ln_2_b;
+
+ // ff
+ struct ggml_tensor * c_mlp_fc_w;
+ struct ggml_tensor * c_mlp_fc_b;
+
+ struct ggml_tensor * c_mlp_proj_w;
+ struct ggml_tensor * c_mlp_proj_b;
+};
+
+struct stablelm_model {
+ stablelm_hparams hparams;
+
+ // normalization
+ struct ggml_tensor * ln_f_g;
+ struct ggml_tensor * ln_f_b;
+
+ struct ggml_tensor * wte; // position embedding
+
+ struct ggml_tensor * lmh_g; // language model head
+ //struct ggml_tensor * lmh_b; // language model bias
+
+ std::vector<stablelm_layer> layers;
+
+ // key + value memory
+ struct ggml_tensor * memory_k;
+ struct ggml_tensor * memory_v;
+
+ //
+ struct ggml_context * ctx;
+ std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool stablelm_model_load(const std::string & fname, stablelm_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.n_vocab, sizeof(hparams.n_vocab));
+ fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
+ fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
+ fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
+ fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+ fin.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
+ fin.read((char *) &hparams.ftype, sizeof(hparams.ftype));
+
+ printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+ printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx);
+ printf("%s: n_embd = %d\n", __func__, hparams.n_embd);
+ printf("%s: n_head = %d\n", __func__, hparams.n_head);
+ printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+ printf("%s: n_rot = %d\n", __func__, hparams.n_rot);
+ printf("%s: ftype = %d\n", __func__, hparams.ftype);
+ }
+
+ // load vocab
+ {
+ const int32_t n_vocab = model.hparams.n_vocab;
+
+ 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.ftype) {
+ 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;
+ case 5: wtype = GGML_TYPE_Q4_2; break;
+ case 6: wtype = GGML_TYPE_Q4_3; break;
+ default:
+ {
+ fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
+ __func__, fname.c_str(), model.hparams.ftype);
+ return false;
+ }
+ }
+
+ const ggml_type wtype2 = GGML_TYPE_F32;
+
+ auto & ctx = model.ctx;
+
+ size_t ctx_size = 0;
+
+ {
+ const auto & hparams = model.hparams;
+
+ const int n_embd = hparams.n_embd;
+ const int n_layer = hparams.n_layer;
+ const int n_ctx = hparams.n_ctx;
+ const int n_vocab = hparams.n_vocab;
+
+ ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+ ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
+
+ ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
+
+ ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // lmh_g
+ //ctx_size += n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
+
+ ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+ ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
+
+ ctx_size += n_layer*(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*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_proj_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*(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 + 16*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,
+ .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 int n_embd = hparams.n_embd;
+ const int n_layer = hparams.n_layer;
+ const int n_ctx = hparams.n_ctx;
+ const int n_vocab = hparams.n_vocab;
+
+ model.layers.resize(n_layer);
+
+ model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
+
+ model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+ model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+ model.lmh_g = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
+ //model.lmh_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_vocab);
+
+ // map by name
+ model.tensors["gpt_neox.embed_in.weight"] = model.wte;
+
+ model.tensors["gpt_neox.final_layer_norm.weight"] = model.ln_f_g;
+ model.tensors["gpt_neox.final_layer_norm.bias"] = model.ln_f_b;
+
+ model.tensors["embed_out.weight"] = model.lmh_g;
+ //model.tensors["lm_head.bias"] = model.lmh_b;
+
+ for (int i = 0; i < n_layer; ++i) {
+ auto & layer = model.layers[i];
+
+ layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+ layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+
+ layer.c_attn_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.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_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["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.weight"] = layer.ln_1_g;
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".input_layernorm.bias"] = layer.ln_1_b;
+
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.weight"] = layer.c_attn_attn_w;
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.query_key_value.bias"] = layer.c_attn_attn_b;
+
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.weight"] = layer.c_attn_proj_w;
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".attention.dense.bias"] = layer.c_attn_proj_b;
+
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.weight"] = layer.ln_2_g;
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".post_attention_layernorm.bias"] = layer.ln_2_b;
+
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.weight"] = layer.c_mlp_fc_w;
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_h_to_4h.bias"] = layer.c_mlp_fc_b;
+
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.weight"] = layer.c_mlp_proj_w;
+ model.tensors["gpt_neox.layers." + std::to_string(i) + ".mlp.dense_4h_to_h.bias"] = layer.c_mlp_proj_b;
+ }
+ }
+
+ // key + value memory
+ {
+ const auto & hparams = model.hparams;
+
+ const int n_embd = hparams.n_embd;
+ const int n_layer = hparams.n_layer;
+ const int n_ctx = hparams.n_ctx;
+
+ const int n_mem = n_layer*n_ctx;
+ const int n_elements = n_embd*n_mem;
+
+ model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_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 = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+ }
+
+ // load weights
+ {
+ int n_tensors = 0;
+ size_t total_size = 0;
+
+ printf("%s: ", __func__);
+
+ while (true) {
+ int32_t n_dims;
+ int32_t length;
+ int32_t ftype;
+
+ fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+ fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+ fin.read(reinterpret_cast<char *>(&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<char *>(&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;
+ }
+
+ if (0) {
+ static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", "q4_2", };
+ 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;
+ case 5: bpe = ggml_type_size(GGML_TYPE_Q4_2); assert(ne[0] % 64 == 0); break;
+ case 6: bpe = ggml_type_size(GGML_TYPE_Q4_3); assert(ne[0] % 64 == 0); break;
+ default:
+ {
+ fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+ return false;
+ }
+ };
+
+ if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+ fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+ __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+ return false;
+ }
+
+ fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+ //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+ total_size += ggml_nbytes(tensor);
+ if (++n_tensors % 8 == 0) {
+ printf(".");
+ fflush(stdout);
+ }
+ }
+
+ printf(" done\n");
+
+ printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+ }
+
+ 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 stablelm_eval(
+ const stablelm_model & model,
+ const int n_threads,
+ const int n_past,
+ const std::vector<gpt_vocab::id> & embd_inp,
+ std::vector<float> & embd_w,
+ size_t & mem_per_token) {
+ const int N = embd_inp.size();
+
+ const auto & hparams = model.hparams;
+
+ const int n_embd = hparams.n_embd;
+ const int n_layer = hparams.n_layer;
+ const int n_ctx = hparams.n_ctx;
+ const int n_head = hparams.n_head;
+ const int n_vocab = hparams.n_vocab;
+ const int n_rot = hparams.n_rot;
+
+ 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,
+ .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));
+
+ // wte
+ struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.wte, embd);
+
+ for (int il = 0; il < n_layer; ++il) {
+ struct ggml_tensor * cur;
+
+ // self-attention
+ {
+ {
+ cur = ggml_norm(ctx0, inpL);
+
+ 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));
+ }
+
+ // compute QKV
+ {
+ 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);
+ }
+
+ struct ggml_tensor * Qcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 0*sizeof(float)*n_embd/n_head));
+ struct ggml_tensor * Kcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 1*sizeof(float)*n_embd/n_head));
+ struct ggml_tensor * Vcur = ggml_cont(ctx0, ggml_view_3d(ctx0, cur, n_embd/n_head, n_head, N, cur->nb[1]/n_head, cur->nb[1], 2*sizeof(float)*n_embd/n_head));
+
+ // using mode = 2 for GPTNeoX mode
+ Qcur = ggml_rope(ctx0, Qcur, n_past, n_rot, 2);
+ Kcur = ggml_rope(ctx0, Kcur, n_past, n_rot, 2);
+
+ // store key and value to memory
+ {
+ Vcur = ggml_transpose(ctx0, ggml_reshape_2d(ctx0, Vcur, n_embd, N));
+
+ 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_2d(ctx0, model.memory_v, N, n_embd,
+ ( n_ctx)*ggml_element_size(model.memory_v),
+ (il*n_ctx)*ggml_element_size(model.memory_v)*n_embd + n_past*ggml_element_size(model.memory_v));
+
+ ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+ ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+ }
+
+ // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+ struct ggml_tensor * Q =
+ ggml_permute(ctx0,
+ Qcur,
+ 0, 2, 1, 3);
+
+ // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+ struct ggml_tensor * K =
+ ggml_permute(ctx0,
+ ggml_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))
+ );
+
+ // KQ_masked = mask_past(KQ_scaled)
+ struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+
+ // KQ = soft_max(KQ_masked)
+ struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+ // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+ struct ggml_tensor * V =
+ ggml_view_3d(ctx0, model.memory_v,
+ n_past + N, n_embd/n_head, n_head,
+ n_ctx*ggml_element_size(model.memory_v),
+ n_ctx*ggml_element_size(model.memory_v)*n_embd/n_head,
+ il*n_ctx*ggml_element_size(model.memory_v)*n_embd);
+
+ // KQV = transpose(V) * KQ_soft_max
+ struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, 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_proj_w,
+ cur);
+
+ cur = ggml_add(ctx0, ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur), cur);
+ }
+ }
+
+ struct ggml_tensor * inpFF = cur;
+
+ // feed-forward network
+ // this is independent of the self-attention result, so it could be done in parallel to the self-attention
+ {
+ // post attention layer norm
+ // note here we pass inpL instead of cur
+ {
+ cur = ggml_norm(ctx0, inpL);
+
+ 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));
+ }
+
+ 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
+ cur = ggml_gelu(ctx0, cur);
+
+ // projection
+ // cur = proj_w*cur + proj_b
+ cur = ggml_mul_mat(ctx0,
+ model.layers[il].c_mlp_proj_w,
+ cur);
+
+ cur = ggml_add(ctx0,
+ ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
+ cur);
+ }
+
+ // layer input + FF
+ cur = ggml_add(ctx0, cur, inpFF);
+
+ // input for next layer
+ inpL = ggml_add(ctx0, cur, inpL);
+ }
+
+ // norm
+ {
+ inpL = ggml_norm(ctx0, inpL);
+
+ // inpL = ln_f_g*inpL + ln_f_b
+ inpL = ggml_add(ctx0,
+ ggml_mul(ctx0,
+ ggml_repeat(ctx0, model.ln_f_g, inpL),
+ inpL),
+ ggml_repeat(ctx0, model.ln_f_b, inpL));
+ }
+
+ // lm_head
+ {
+ inpL = ggml_mul_mat(ctx0, model.lmh_g, inpL);
+
+ //inpL = ggml_add(ctx0,
+ // ggml_repeat(ctx0, model.lmh_b, inpL),
+ // inpL);
+ }
+
+ // logits -> probs
+ //inpL = ggml_soft_max(ctx0, inpL);
+
+ // run the computation
+ ggml_build_forward_expand(&gf, inpL);
+ ggml_graph_compute (ctx0, &gf);
+
+ //if (n_past%100 == 0) {
+ // ggml_graph_print (&gf);
+ // ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+ //}
+
+ //embd_w.resize(n_vocab*N);
+ //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+ // return result for just the last token
+ embd_w.resize(n_vocab);
+ memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+ if (mem_per_token == 0) {
+ mem_per_token = ggml_used_mem(ctx0)/N;
+ }
+ //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+ ggml_free(ctx0);
+
+ return true;
+}
+
+int main(int argc, char ** argv) {
+ const int64_t t_main_start_us = ggml_time_us();
+
+ gpt_params params;
+ params.model = "models/stablelm-base-alpha-3b/ggml-model-f16.bin";
+
+ if (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 = gpt_random_prompt(rng);
+ }
+ }
+
+ int64_t t_load_us = 0;
+
+ gpt_vocab vocab;
+ stablelm_model model;
+
+ // load the model
+ {
+ const int64_t t_start_us = ggml_time_us();
+
+ if (!stablelm_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<float> logits;
+
+ // tokenize the prompt
+ std::vector<gpt_vocab::id> 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());
+ for (int i = 0; i < embd_inp.size(); i++) {
+ printf("%s: token[%d] = %6d, %s\n", __func__, i, embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str());
+ }
+ printf("\n");
+
+ std::vector<gpt_vocab::id> embd;
+
+ // determine the required inference memory per token:
+ size_t mem_per_token = 0;
+ stablelm_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 (!stablelm_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() == 0) {
+ 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;
+}
--- /dev/null
+#include "ggml/ggml.h"
+
+#include "utils.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// default hparams (StableLM 3B)
+struct stablelm_hparams {
+ int32_t n_vocab = 50257;
+ int32_t n_ctx = 4096;
+ int32_t n_embd = 4096;
+ int32_t n_head = 32;
+ int32_t n_layer = 16;
+ int32_t n_rot = 32; // 0.25 * (n_embd / n_head)
+ int32_t ftype = 1;
+};
+
+// quantize a model
+bool stablelm_model_quantize(const std::string & fname_inp, const std::string & fname_out, int itype) {
+ ggml_type type = GGML_TYPE_Q4_1;
+
+ switch (itype) {
+ case 2: type = GGML_TYPE_Q4_0; break;
+ case 3: type = GGML_TYPE_Q4_1; break;
+ case 5: type = GGML_TYPE_Q4_2; break;
+ case 6: type = GGML_TYPE_Q4_3; break;
+ default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1;
+ };
+
+ if (!ggml_is_quantized(type)) {
+ fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type);
+ return false;
+ }
+
+ gpt_vocab vocab;
+
+ 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));
+ }
+
+ stablelm_hparams hparams;
+
+ // load hparams
+ {
+ finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+ finp.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
+ finp.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
+ finp.read((char *) &hparams.n_head, sizeof(hparams.n_head));
+ finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+ finp.read((char *) &hparams.n_rot, sizeof(hparams.n_rot));
+ finp.read((char *) &hparams.ftype, sizeof(hparams.ftype));
+
+ 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: ftype = %d\n", __func__, hparams.ftype);
+
+ fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+ fout.write((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
+ fout.write((char *) &hparams.n_embd, sizeof(hparams.n_embd));
+ fout.write((char *) &hparams.n_head, sizeof(hparams.n_head));
+ fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+ fout.write((char *) &hparams.n_rot, sizeof(hparams.n_rot));
+ fout.write((char *) &itype, sizeof(hparams.ftype));
+ }
+
+ // 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);
+
+ vocab.token_to_id[word] = i;
+ vocab.id_to_token[i] = word;
+ }
+ }
+
+ // load weights
+ {
+ size_t total_size_org = 0;
+ size_t total_size_new = 0;
+
+ std::vector<float> work;
+
+ std::vector<uint8_t> data_u8;
+ std::vector<ggml_fp16_t> data_f16;
+ std::vector<float> data_f32;
+
+ std::vector<int64_t> hist_all(1 << 4, 0);
+
+ while (true) {
+ int32_t n_dims;
+ int32_t length;
+ int32_t ftype;
+
+ finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+ finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+ finp.read(reinterpret_cast<char *>(&ftype), sizeof(ftype));
+
+ if (finp.eof()) {
+ break;
+ }
+
+ int32_t nelements = 1;
+ int32_t ne[2] = { 1, 1 };
+ for (int i = 0; i < n_dims; ++i) {
+ finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+ nelements *= ne[i];
+ }
+
+ std::string name(length, 0);
+ finp.read (&name[0], length);
+
+ {
+ static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", "q4_2" };
+ printf("%64s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
+ }
+
+ // regexes of tensor names to be quantized
+ const std::vector<std::string> k_names = {
+ ".*weight",
+ };
+
+ bool quantize = false;
+ for (const auto & s : k_names) {
+ if (std::regex_match(name, std::regex(s))) {
+ quantize = true;
+ break;
+ }
+ }
+
+ // quantize only 2D tensors
+ quantize &= (n_dims == 2);
+
+ if (quantize) {
+ if (ftype != 0 && ftype != 1) {
+ fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype);
+ return false;
+ }
+
+ if (ftype == 1) {
+ data_f16.resize(nelements);
+ finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+ data_f32.resize(nelements);
+ for (int i = 0; i < nelements; ++i) {
+ data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+ }
+ } else {
+ data_f32.resize(nelements);
+ finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+ }
+
+ ftype = itype;
+ } else {
+ const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+ data_u8.resize(nelements*bpe);
+ finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+ }
+
+ fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+ fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+ fout.write(reinterpret_cast<char *>(&ftype), sizeof(ftype));
+ for (int i = 0; i < n_dims; ++i) {
+ fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+ }
+ fout.write(&name[0], length);
+
+ if (quantize) {
+ printf("quantizing .. ");
+ work.resize(nelements); // for quantization
+
+ size_t cur_size = 0;
+ std::vector<int64_t> hist_cur(1 << 4, 0);
+
+ switch (type) {
+ case GGML_TYPE_Q4_0:
+ {
+ cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+ } break;
+ case GGML_TYPE_Q4_1:
+ {
+ cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+ } break;
+ case GGML_TYPE_Q4_2:
+ {
+ cur_size = ggml_quantize_q4_2(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+ } break;
+ case GGML_TYPE_Q4_3:
+ {
+ cur_size = ggml_quantize_q4_3(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
+ } break;
+ default:
+ {
+ fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type);
+ return false;
+ }
+ }
+
+ fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+ total_size_new += cur_size;
+
+ printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+ for (int i = 0; i < hist_cur.size(); ++i) {
+ hist_all[i] += hist_cur[i];
+ }
+
+ for (int i = 0; i < hist_cur.size(); ++i) {
+ printf("%5.3f ", hist_cur[i] / (float)nelements);
+ }
+ printf("\n");
+ } else {
+ printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+ fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+ total_size_new += data_u8.size();
+ }
+
+ total_size_org += nelements * sizeof(float);
+ }
+
+ printf("%s: model size = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+ printf("%s: quant size = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+
+ {
+ int64_t sum_all = 0;
+ for (int i = 0; i < hist_all.size(); ++i) {
+ sum_all += hist_all[i];
+ }
+
+ printf("%s: hist: ", __func__);
+ for (int i = 0; i < hist_all.size(); ++i) {
+ printf("%5.3f ", hist_all[i] / (float)sum_all);
+ }
+ printf("\n");
+ }
+ }
+
+ finp.close();
+ fout.close();
+
+ return true;
+}
+
+// usage:
+// ./stablelm2-quantize models/stablelm2-117M/ggml-model.bin models/stablelm2-117M/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]);
+ fprintf(stderr, " type = 2 - q4_0\n");
+ fprintf(stderr, " type = 3 - q4_1\n");
+ fprintf(stderr, " type = 5 - q4_2\n");
+ fprintf(stderr, " type = 6 - q4_3\n");
+ 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 int itype = atoi(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 (!stablelm_model_quantize(fname_inp, fname_out, itype)) {
+ 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;
+}
overview / pkg / ggml / sources / ggml / commitdiff