vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Merge remote-tracking branch 'origin/master' into bins

Browse source
This commit is contained in:
Olivier Chafik 2024-06-10 15:38:41 +01:00
commit daeaeb1222
30 changed files with 1233 additions and 607 deletions
Download patch file Download diff file Expand all files Collapse all files

19
examples/alpaca.sh
View file

@ -1,19 +0,0 @@
#!/bin/bash
#
# Temporary script - will be removed in the future
#
cd `dirname $0`
cd ..
./llama-cli -m ./models/alpaca.13b.ggmlv3.q8_0.bin \
--color \
-f ./prompts/alpaca.txt \
--ctx_size 2048 \
-n -1 \
-ins -b 256 \
--top_k 10000 \
--temp 0.2 \
--repeat_penalty 1.1 \
-t 7

15
examples/gpt4all.sh
View file

@ -1,15 +0,0 @@
#!/bin/bash
#
# Temporary script - will be removed in the future
#
cd `dirname $0`
cd ..
./llama-cli --color --instruct --threads 4 \
--model ./models/gpt4all-7B/gpt4all-lora-quantized.bin \
--file ./prompts/alpaca.txt \
--batch_size 8 --ctx_size 2048 -n -1 \
--repeat_last_n 64 --repeat_penalty 1.3 \
--n_predict 128 --temp 0.1 --top_k 40 --top_p 0.95

58
examples/imatrix/imatrix.cpp
View file

@ -218,20 +218,64 @@ void IMatrixCollector::save_imatrix(int ncall) const {
fname += std::to_string(ncall);
}
// avoid writing imatrix entries that do not have full data
// this can happen with MoE models where some of the experts end up not being exercised by the provided training data
int n_entries = 0;
std::vector<std::string> to_store;
bool is_first = true; // for printing
for (const auto & kv : m_stats) {
const int n_all = kv.second.counts.size();
if (n_all == 0) {
continue;
}
int n_zeros = 0;
for (const int c : kv.second.counts) {
if (c == 0) {
n_zeros++;
}
}
if (n_zeros != 0 && is_first) {
fprintf(stderr, "\n");
is_first = false;
}
if (n_zeros == n_all) {
fprintf(stderr, "%s: entry '%40s' has no data - skipping\n", __func__, kv.first.c_str());
continue;
}
if (n_zeros > 0) {
fprintf(stderr, "%s: entry '%40s' has partial data (%.2f%%) - skipping\n", __func__, kv.first.c_str(), 100.0f * (n_all - n_zeros) / n_all);
continue;
}
n_entries++;
to_store.push_back(kv.first);
}
if (to_store.size() < m_stats.size()) {
fprintf(stderr, "%s: warning: storing only %zu out of %zu entries\n", __func__, to_store.size(), m_stats.size());
}
std::ofstream out(fname, std::ios::binary);
int n_entries = m_stats.size();
out.write((const char *) &n_entries, sizeof(n_entries));
for (const auto & p : m_stats) {
int len = p.first.size();
for (const auto & name : to_store) {
const auto & stat = m_stats.at(name);
int len = name.size();
out.write((const char *) &len, sizeof(len));
out.write(p.first.c_str(), len);
out.write((const char *) &p.second.ncall, sizeof(p.second.ncall));
int nval = p.second.values.size();
out.write(name.c_str(), len);
out.write((const char *) &stat.ncall, sizeof(stat.ncall));
int nval = stat.values.size();
out.write((const char *) &nval, sizeof(nval));
if (nval > 0) {
std::vector<float> tmp(nval);
for (int i = 0; i < nval; i++) {
tmp[i] = (p.second.values[i] / static_cast<float>(p.second.counts[i])) * static_cast<float>(p.second.ncall);
tmp[i] = (stat.values[i] / static_cast<float>(stat.counts[i])) * static_cast<float>(stat.ncall);
}
out.write((const char*)tmp.data(), nval*sizeof(float));
}

18
examples/llama2-13b.sh
View file

@ -1,18 +0,0 @@
#!/bin/bash
#
# Temporary script - will be removed in the future
#
cd `dirname $0`
cd ..
./llama-cli -m models/available/Llama2/13B/llama-2-13b.ggmlv3.q4_0.bin \
--color \
--ctx_size 2048 \
-n -1 \
-ins -b 256 \
--top_k 10000 \
--temp 0.2 \
--repeat_penalty 1.1 \
-t 8

18
examples/llama2.sh
View file

@ -1,18 +0,0 @@
#!/bin/bash
#
# Temporary script - will be removed in the future
#
cd `dirname $0`
cd ..
./llama-cli -m models/available/Llama2/7B/llama-2-7b.ggmlv3.q4_0.bin \
--color \
--ctx_size 2048 \
-n -1 \
-ins -b 256 \
--top_k 10000 \
--temp 0.2 \
--repeat_penalty 1.1 \
-t 8

10
examples/rpc/rpc-server.cpp
View file

@ -6,10 +6,6 @@
#include "ggml-metal.h"
#endif
#ifdef GGML_USE_SYCL
#include "ggml-sycl.h"
#endif
#include "ggml-rpc.h"
#ifdef _WIN32
# include <windows.h>
@ -83,12 +79,6 @@ static ggml_backend_t create_backend() {
if (!backend) {
fprintf(stderr, "%s: ggml_backend_metal_init() failed\n", __func__);
}
#elif GGML_USE_SYCL
fprintf(stderr, "%s: using SYCL backend\n", __func__);
backend = ggml_backend_sycl_init(0); // init device 0
if (!backend) {
fprintf(stderr, "%s: ggml_backend_sycl_init() failed\n", __func__);
}
#endif
// if there aren't GPU Backends fallback to CPU backend

2
examples/server/public/index-new.html
View file

@ -416,7 +416,7 @@
message = html`<${Probabilities} data=${data} />`
} else {
const text = isArrayMessage ?
data.map(msg => msg.content).join('').replace(/^\s+/, '') :
data.map(msg => msg.content).join('') :
data;
message = isCompletionMode ?
text :

30
examples/server/server.cpp
View file

@ -147,7 +147,7 @@ struct server_slot {
int32_t n_prompt_tokens = 0;
int32_t n_prompt_tokens_processed = 0;
json prompt;
std::string prompt;
// when a task is submitted, we first tokenize the prompt and store it here
std::vector<llama_token> prompt_tokens;
@ -822,13 +822,8 @@ struct server_context {
continue;
}
// skip the slot if it does not contains prompt
if (!slot.prompt.is_string()) {
continue;
}
// current slot's prompt
std::string slot_prompt = slot.prompt.get<std::string>();
std::string slot_prompt = slot.prompt;
// length of the current slot's prompt
int slot_prompt_len = slot_prompt.size();
@ -958,13 +953,16 @@ struct server_context {
if (!task.infill) {
const auto & prompt = data.find("prompt");
if (prompt == data.end()) {
send_error(task, "Either \"prompt\" or \"messages\" must be provided", ERROR_TYPE_INVALID_REQUEST);
send_error(task, "\"prompt\" must be provided", ERROR_TYPE_INVALID_REQUEST);
return false;
} else {
slot.prompt = *prompt;
}
if (slot.prompt.is_array() && slot.prompt.size() == 0) {
send_error(task, "\"prompt\" cannot be an empty array", ERROR_TYPE_INVALID_REQUEST);
if (prompt->is_string()) {
slot.prompt = prompt->get<std::string>();
} else if (prompt->is_array() && prompt->size() == 1 && prompt->at(0).is_string()) {
slot.prompt = prompt->at(0).get<std::string>();
} else {
send_error(task, "\"prompt\" must be a string or an array of strings", ERROR_TYPE_INVALID_REQUEST);
return false;
}
}
@ -1582,14 +1580,18 @@ struct server_context {
switch (task.type) {
case SERVER_TASK_TYPE_COMPLETION:
{
int id_slot = json_value(task.data, "id_slot", -1);
std::string prompt = json_value(task.data, "prompt", std::string());
const int id_slot = json_value(task.data, "id_slot", -1);
server_slot * slot;
if (id_slot != -1) {
slot = get_slot_by_id(id_slot);
} else {
std::string prompt;
if (task.data.contains("prompt") && task.data.at("prompt").is_string()) {
json_value(task.data, "prompt", std::string());
}
slot = get_available_slot(prompt);
}