vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-17 04:19:40 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'master' into concedo_experimental

Browse source 
# Conflicts:
#	Makefile
#	README.md
#	common/log.h
This commit is contained in:
Concedo 2023-09-02 11:24:28 +08:00
commit eed651494e
26 changed files with 1143 additions and 658 deletions
Download patch file Download diff file Expand all files Collapse all files

5
examples/baby-llama/baby-llama.cpp
View file

@ -1617,15 +1617,10 @@ int main(int argc, char ** argv) {
float error_before_opt = ggml_get_f32_1d(e, 0);
struct ggml_opt_params opt_params_adam = ggml_opt_default_params(GGML_OPT_ADAM);
struct ggml_opt_params opt_params_lbfgs = ggml_opt_default_params(GGML_OPT_LBFGS);
opt_params_adam.print_forward_graph = false;
opt_params_adam.print_backward_graph = false;
opt_params_lbfgs.print_forward_graph = false;
opt_params_lbfgs.print_backward_graph = false;
opt_params_adam.adam.n_iter = 16;
opt_params_lbfgs.lbfgs.n_iter = 16;
// ggml_opt(ctx0, opt_params_adam, e);
ggml_opt(ctx0, opt_params_lbfgs, e);
//
ggml_build_forward_expand(&gf, e);

8
examples/beam-search/beam-search.cpp
View file

@ -22,7 +22,9 @@
#include <unistd.h>
#elif defined (_WIN32)
#define WIN32_LEAN_AND_MEAN
#define NOMINMAX
#ifndef NOMINMAX
# define NOMINMAX
#endif
#include <windows.h>
#include <signal.h>
#endif
@ -73,7 +75,7 @@ void beam_search_callback(void * callback_data_ptr, llama_beams_state beams_stat
assert(0u < beams_state.n_beams);
const llama_token * tokens = beams_state.beam_views[0].tokens;
std::copy(tokens, tokens + n, callback_data.response.end() - n);
printf("%lu", n);
printf("%zu", n);
}
fflush(stdout);
#if 1 // DEBUG: print current beams for this iteration
@ -145,7 +147,7 @@ int main(int argc, char ** argv)
if (tokens_list.size() > max_tokens_list_size)
{
fprintf( stderr , "%s: error: prompt too long (%lu tokens, max %lu)\n" ,
fprintf( stderr , "%s: error: prompt too long (%zu tokens, max %zu)\n" ,
__func__ , tokens_list.size() , max_tokens_list_size );
return 1;
}

71
examples/convert-llama2c-to-ggml/convert-llama2c-to-ggml.cpp
View file

@ -75,7 +75,7 @@ typedef struct {
int seq_len; // max sequence length
} Config;
typedef struct {
struct TransformerWeights {
// token embedding table
float* token_embedding_table; // (vocab_size, dim)
// weights for rmsnorms
@ -97,7 +97,22 @@ typedef struct {
// float* freq_cis_imag; // (seq_len, dim/2)
// (optional) classifier weights for the logits, on the last layer
float* wcls;
} TransformerWeights;
~TransformerWeights() {
delete[] token_embedding_table;
delete[] rms_att_weight;
delete[] rms_ffn_weight;
delete[] wq;
delete[] wk;
delete[] wv;
delete[] wo;
delete[] w1;
delete[] w2;
delete[] w3;
delete[] rms_final_weight;
delete[] wcls;
}
};
void malloc_weights(TransformerWeights* w, Config* p, bool shared_weights) {
// we calloc instead of malloc to keep valgrind happy
@ -173,21 +188,6 @@ int checkpoint_init_weights(TransformerWeights *w, Config* p, FILE* f, bool shar
return 0;
}
void free_weights(TransformerWeights* w) {
delete w->token_embedding_table;
delete w->rms_att_weight;
delete w->rms_ffn_weight;
delete w->wq;
delete w->wk;
delete w->wv;
delete w->wo;
delete w->w1;
delete w->w2;
delete w->w3;
delete w->rms_final_weight;
if (w->wcls) delete w->wcls;
}
void print_sample_weights(TransformerWeights *w){
printf("----- Quick print of first of the weight vales of all the variables\n");
printf("%f\n", w->token_embedding_table[0]);
@ -596,6 +596,10 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
// assume llama2.c vocabulary
printf("Assuming llama2.c vocabulary since %s is not a gguf file\n", filename);
llama_file file(filename, "rb");
if (!file.fp) {
fprintf(stderr, "error: %s: %s\n", strerror(errno), filename);
exit(1);
}
const int n_vocab = config->vocab_size;
/* uint32_t max_token_length = */ file.read_u32(); // unused
vocab->id_to_token.resize(n_vocab);
@ -633,7 +637,7 @@ void load_vocab(const char *filename, Config *config, struct llama_vocab *vocab)
}
}
void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * karpathy_weights){
void convert_weights_ak_to_gg(struct ggml_tensor * gg_weights, const float * karpathy_weights) {
int ct;
switch (gg_weights->n_dims){
case 1:
@ -670,13 +674,13 @@ void stuff_karpathy_weights_into_gg(struct ggml_tensor * gg_weights, float * kar
}
void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * model, TransformerWeights* w, const char * filename) {
// stuff AK weights into GG weights one by one.
// convert AK weights into GG weights one by one.
// w->token_embedding_table -> model->tok_embeddings
// float* -> struct ggml_tensor
stuff_karpathy_weights_into_gg(model->tok_embeddings, w->token_embedding_table);
stuff_karpathy_weights_into_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
convert_weights_ak_to_gg(model->tok_embeddings, w->token_embedding_table);
convert_weights_ak_to_gg(model->output, w->wcls ? w->wcls : w->token_embedding_table);
stuff_karpathy_weights_into_gg(model->norm, w->rms_final_weight);
convert_weights_ak_to_gg(model->norm, w->rms_final_weight);
//print_row(model->norm, 0);
// for rms-att-weight
@ -686,18 +690,18 @@ void save_as_llama_model(struct llama_vocab * vocab, struct my_llama_model * mod
for (uint32_t i = 0; i < model->hparams.n_layer; ++i){
auto & layer = model->layers[i];
// 1d
stuff_karpathy_weights_into_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
stuff_karpathy_weights_into_gg(layer.ffn_norm , &w->rms_ffn_weight[i*row_length]);
convert_weights_ak_to_gg(layer.attention_norm, &w->rms_att_weight[i*row_length]);
convert_weights_ak_to_gg(layer.ffn_norm , &w->rms_ffn_weight[i*row_length]);
// from 3d matrix layer x dim x dim to 2d matrix dim x dim
stuff_karpathy_weights_into_gg(layer.wq , &w->wq[i*row_length*row_length]);
stuff_karpathy_weights_into_gg(layer.wk , &w->wk[i*row_length*row_length]);
stuff_karpathy_weights_into_gg(layer.wv , &w->wv[i*row_length*row_length]);
stuff_karpathy_weights_into_gg(layer.wo , &w->wo[i*row_length*row_length]);
convert_weights_ak_to_gg(layer.wq , &w->wq[i*row_length*row_length]);
convert_weights_ak_to_gg(layer.wk , &w->wk[i*row_length*row_length]);
convert_weights_ak_to_gg(layer.wv , &w->wv[i*row_length*row_length]);
convert_weights_ak_to_gg(layer.wo , &w->wo[i*row_length*row_length]);
stuff_karpathy_weights_into_gg(layer.w1 , &w->w1[i*row_length*n_ff]);
stuff_karpathy_weights_into_gg(layer.w2 , &w->w2[i*n_ff*row_length]);
stuff_karpathy_weights_into_gg(layer.w3 , &w->w3[i*row_length*n_ff]);
convert_weights_ak_to_gg(layer.w1 , &w->w1[i*row_length*n_ff]);
convert_weights_ak_to_gg(layer.w2 , &w->w2[i*n_ff*row_length]);
convert_weights_ak_to_gg(layer.w3 , &w->w3[i*row_length*n_ff]);
}
struct gguf_context * ctx = gguf_init_empty();
@ -898,7 +902,7 @@ bool params_parse(int argc, char ** argv, struct train_params * params) {
}
std::string basename(const std::string &path) {
size_t pos = path.find_last_of("/");
size_t pos = path.find_last_of("/\\");
if (pos == std::string::npos) {
return path;
}
@ -911,7 +915,7 @@ int main(int argc, char ** argv) {
return 1;
}
Config config;
TransformerWeights weights;
TransformerWeights weights = {};
{
FILE *file = fopen(params.fn_llama2c_model, "rb");
if (!file) { printf("Unable to open the checkpoint file %s!\n", params.fn_llama2c_model); return 1; }
@ -953,6 +957,5 @@ int main(int argc, char ** argv) {
printf("Saving llama.c model file %s in ggml format at %s\n", params.fn_llama2c_model, params.fn_llama2c_output_model);
ggml_free(model.ctx);
free_weights(&weights);
return 0;
}

4
examples/main/README.md
View file

@ -34,7 +34,7 @@ For an interactive experience, try this command:
#### Unix-based systems (Linux, macOS, etc.):
```bash
./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " \
./main -m models/7B/ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -p \
'User: Hi
AI: Hello. I am an AI chatbot. Would you like to talk?
User: Sure!
@ -45,7 +45,7 @@ User:'
#### Windows:
```powershell
main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -e --prompt "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:"
main.exe -m models\7B\ggml-model.bin -n -1 --color -r "User:" --in-prefix " " -i -e -p "User: Hi\nAI: Hello. I am an AI chatbot. Would you like to talk?\nUser: Sure!\nAI: What would you like to talk about?\nUser:"
```
The following command generates "infinite" text from a starting prompt (you can use `Ctrl-C` to stop it):

24
examples/quantize/quantize.cpp
View file

@ -33,6 +33,8 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q8_0", LLAMA_FTYPE_MOSTLY_Q8_0, " 6.70G, +0.0004 ppl @ LLaMA-v1-7B", },
{ "F16", LLAMA_FTYPE_MOSTLY_F16, "13.00G @ 7B", },
{ "F32", LLAMA_FTYPE_ALL_F32, "26.00G @ 7B", },
// Note: Ensure COPY comes after F32 to avoid ftype 0 from matching.
{ "COPY", LLAMA_FTYPE_ALL_F32, "only copy tensors, no quantizing", },
};
@ -69,12 +71,17 @@ bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftype, std:
// ./quantize [--allow-requantize] [--leave-output-tensor] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
//
void usage(const char * executable) {
fprintf(stderr, "usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
fprintf(stderr, " --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
fprintf(stderr, " --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
fprintf(stderr, "\nAllowed quantization types:\n");
printf("usage: %s [--help] [--allow-requantize] [--leave-output-tensor] model-f32.gguf [model-quant.gguf] type [nthreads]\n\n", executable);
printf(" --allow-requantize: Allows requantizing tensors that have already been quantized. Warning: This can severely reduce quality compared to quantizing from 16bit or 32bit\n");
printf(" --leave-output-tensor: Will leave output.weight un(re)quantized. Increases model size but may also increase quality, especially when requantizing\n");
printf("\nAllowed quantization types:\n");
for (auto & it : QUANT_OPTIONS) {
printf(" %2d or %-6s : %s\n", it.ftype, it.name.c_str(), it.desc.c_str());
if (it.name != "COPY") {
printf(" %2d or ", it.ftype);
} else {
printf(" ");
}
printf("%-6s : %s\n", it.name.c_str(), it.desc.c_str());
}
exit(1);
}
@ -119,6 +126,9 @@ int main(int argc, char ** argv) {
// export as [inp path]/ggml-model-[ftype].gguf
fname_out = fpath + "ggml-model-" + ftype_str + ".gguf";
arg_idx++;
if (ftype_str == "COPY") {
params.only_copy = true;
}
}
else {
fname_out = argv[arg_idx];
@ -131,6 +141,10 @@ int main(int argc, char ** argv) {
if (!try_parse_ftype(argv[arg_idx], params.ftype, ftype_str)) {
fprintf(stderr, "%s: invalid ftype '%s'\n", __func__, argv[3]);
return 1;
} else {
if (ftype_str == "COPY") {
params.only_copy = true;
}
}
arg_idx++;
}

22
examples/server/server.cpp
View file

@ -17,6 +17,8 @@
#include "completion.js.hpp"
#include "json-schema-to-grammar.mjs.hpp"
#include <cstddef>
#ifndef SERVER_VERBOSE
#define SERVER_VERBOSE 1
#endif
@ -1038,7 +1040,7 @@ static json format_timings(llama_server_context &llama)
{
const auto timings = llama_get_timings(llama.ctx);
assert(timings.n_eval == llama.num_tokens_predicted);
assert(timings.n_eval == ptrdiff_t(llama.num_tokens_predicted));
return json{
{"prompt_n", timings.n_p_eval},
@ -1239,7 +1241,7 @@ void beam_search_callback(void * callback_data, llama_beams_state beams_state) {
const llama_token * tokens = beams_state.beam_views[0].tokens;
const auto map = [](llama_token tok) { return completion_token_output{{},tok}; };
std::transform(tokens, tokens + n, llama.generated_token_probs.end() - n, map);
printf("%lu", n);
printf("%zu", n);
}
fflush(stdout);
#if 0 // DEBUG: print current beams for this iteration
@ -1377,7 +1379,13 @@ int main(int argc, char **argv)
}
}
const json data = format_final_response(llama, llama.generated_text, llama.generated_token_probs);
auto probs = llama.generated_token_probs;
if (llama.params.n_probs > 0 && llama.stopped_word) {
const std::vector<llama_token> stop_word_toks = llama_tokenize(llama.ctx, llama.stopping_word, false);
probs = std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.end() - stop_word_toks.size());
}
const json data = format_final_response(llama, llama.generated_text, probs);
llama_print_timings(llama.ctx);
@ -1454,7 +1462,11 @@ int main(int argc, char **argv)
if (!llama.has_next_token) {
// Generation is done, send extra information.
const json data = format_final_response(llama, "", llama.generated_token_probs);
const json data = format_final_response(
llama,
"",
std::vector<completion_token_output>(llama.generated_token_probs.begin(), llama.generated_token_probs.begin() + sent_token_probs_index)
);
const std::string str =
"data: " +
@ -1548,7 +1560,7 @@ int main(int argc, char **argv)
svr.set_exception_handler([](const Request &, Response &res, std::exception_ptr ep)
{
const auto * fmt = "500 Internal Server Error\n%s";
const char fmt[] = "500 Internal Server Error\n%s";
char buf[BUFSIZ];
try {
std::rethrow_exception(std::move(ep));

5
examples/train-text-from-scratch/convert-train-checkpoint-to-gguf.py
View file

@ -2,13 +2,16 @@
# train-text-from-scratch checkpoint --> gguf conversion
import argparse
import gguf
import os
import struct
import sys
import numpy as np
from pathlib import Path
if 'NO_LOCAL_GGUF' not in os.environ:
sys.path.insert(1, str(Path(__file__).parent / '..' / '..' / 'gguf-py' / 'gguf'))
import gguf
# gguf constants
LLM_KV_OPTIMIZER_TYPE = "optimizer.type"
LLM_KV_OPTIMIZER_TYPE_ADAM = "adam"