Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-19 09:15:18 +02:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into concedo_experimental

Browse Source 
# Conflicts:
#	CMakeLists.txt
#	Makefile
#	README.md
#	build.zig
#	tests/test-backend-ops.cpp
This commit is contained in:
Concedo
2024-02-21 18:59:13 +08:00
parent 2d71256d21 88c46cbdac
commit 1a490e87c8
17 changed files with 734 additions and 110 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/llava/README.md
+32 -6
View File
@@ -59,14 +59,40 @@ python ./convert.py ../llava-v1.5-7b --skip-unknown
Now both the LLaMA part and the image encoder is in the `llava-v1.5-7b` directory.
## LLaVA 1.6 gguf conversion
1) Backup your pth/safetensor model files as llava-surgery modifies them
2) Use `python llava-surgery-v2.py -C -m /path/to/hf-model` which also supports llava-1.5 variants pytorch as well as safetensor models:
1) First clone a LLaVA 1.6 model:
```console
git clone https://huggingface.co/liuhaotian/llava-v1.6-vicuna-7b
```
2) Backup your pth/safetensor model files as llava-surgery modifies them
3) Use `llava-surgery-v2.py` which also supports llava-1.5 variants pytorch as well as safetensor models:
```console
python examples/llava/llava-surgery-v2.py -C -m ../llava-v1.6-vicuna-7b/
```
- you will find a llava.projector and a llava.clip file in your model directory
3) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory (https://huggingface.co/cmp-nct/llava-1.6-gguf/blob/main/config_vit.json) and rename it to config.json.
4) Create the visual gguf model: `python ./examples/llava/convert-image-encoder-to-gguf.py -m ../path/to/vit --llava-projector ../path/to/llava.projector --output-dir ../path/to/output --clip-model-is-vision`
4) Copy the llava.clip file into a subdirectory (like vit), rename it to pytorch_model.bin and add a fitting vit configuration to the directory:
```console
mkdir vit
cp ../llava-v1.6-vicuna-7b/llava.clip vit/pytorch_model.bin
cp ../llava-v1.6-vicuna-7b/llava.projector vit/
curl -s -q https://huggingface.co/cmp-nct/llava-1.6-gguf/raw/main/config_vit.json -o vit/config.json
```
5) Create the visual gguf model:
```console
python ./examples/llava/convert-image-encoder-to-gguf.py -m vit --llava-projector vit/llava.projector --output-dir vit --clip-model-is-vision
```
- This is similar to llava-1.5, the difference is that we tell the encoder that we are working with the pure vision model part of CLIP
5) Everything else as usual: convert.py the hf model, quantize as needed
6) Then convert the model to gguf format:
```console
python ./convert.py ../llava-v1.6-vicuna-7b/
```
7) And finally we can run the llava-cli using the 1.6 model version:
```console
./llava-cli -m ../llava-v1.6-vicuna-7b/ggml-model-f16.gguf --mmproj vit/mmproj-model-f16.gguf --image some-image.jpg -c 4096
```
**note** llava-1.6 needs more context than llava-1.5, at least 3000 is needed (just run it at -c 4096)
**note** llava-1.6 greatly benefits from batched prompt processing (defaults work)
examples/llava/llava.cpp
+1 -1
View File
@@ -311,7 +311,7 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
return true;
}
static bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out) {
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*6); // TODO: base on gridsize/llava model
if (!image_embd) {
fprintf(stderr, "Unable to allocate memory for image embeddings\n");
examples/llava/llava.h
+2
View File
@@ -31,6 +31,8 @@ struct llava_image_embed {
/** sanity check for clip <-> llava embed size match */
LLAVA_API bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx * ctx_clip);
LLAVA_API bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float ** image_embd_out, int * n_img_pos_out);
/** build an image embed from image file bytes */
LLAVA_API struct llava_image_embed * llava_image_embed_make_with_bytes(struct clip_ctx * ctx_clip, int n_threads, const unsigned char * image_bytes, int image_bytes_length);
/** build an image embed from a path to an image filename */
examples/quantize/quantize.cpp
+1
View File
@@ -33,6 +33,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
{ "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },
{ "Q3_K_L", LLAMA_FTYPE_MOSTLY_Q3_K_L, " 3.35G, +0.1764 ppl @ LLaMA-v1-7B", },
{ "IQ4_NL", LLAMA_FTYPE_MOSTLY_IQ4_NL, " 4.25 bpw non-linear quantization", },
{ "Q4_K", LLAMA_FTYPE_MOSTLY_Q4_K_M, "alias for Q4_K_M", },
{ "Q4_K_S", LLAMA_FTYPE_MOSTLY_Q4_K_S, " 3.59G, +0.0992 ppl @ LLaMA-v1-7B", },
{ "Q4_K_M", LLAMA_FTYPE_MOSTLY_Q4_K_M, " 3.80G, +0.0532 ppl @ LLaMA-v1-7B", },
examples/server/oai.hpp
+2 -4
View File
@@ -15,13 +15,11 @@
using json = nlohmann::json;
inline static json oaicompat_completion_params_parse(
const struct llama_model * model,
const json &body, /* openai api json semantics */
const std::string &chat_template)
{
json llama_params;
std::string formatted_prompt = chat_template == "chatml"
? format_chatml(body["messages"]) // OpenAI 'messages' to chatml (with <|im_start|>,...)
: format_llama2(body["messages"]); // OpenAI 'messages' to llama2 (with [INST],...)
llama_params["__oaicompat"] = true;
@@ -34,7 +32,7 @@ inline static json oaicompat_completion_params_parse(
// https://platform.openai.com/docs/api-reference/chat/create
llama_sampling_params default_sparams;
llama_params["model"] = json_value(body, "model", std::string("unknown"));
llama_params["prompt"] = formatted_prompt;
llama_params["prompt"] = format_chat(model, chat_template, body["messages"]);
llama_params["cache_prompt"] = json_value(body, "cache_prompt", false);
llama_params["temperature"] = json_value(body, "temperature", 0.0);
llama_params["top_k"] = json_value(body, "top_k", default_sparams.top_k);
examples/server/server.cpp
+12 -41
View File
@@ -6,6 +6,7 @@
#include "oai.hpp"
#include "../llava/clip.h"
#include "../llava/llava.h"
#include "stb_image.h"
@@ -38,7 +39,7 @@ struct server_params
std::string hostname = "127.0.0.1";
std::vector<std::string> api_keys;
std::string public_path = "examples/server/public";
std::string chat_template = "chatml";
std::string chat_template = "";
int32_t port = 8080;
int32_t read_timeout = 600;
int32_t write_timeout = 600;
@@ -998,43 +999,12 @@ struct llama_server_context
{
continue;
}
clip_image_f32_batch img_res_v;
img_res_v.size = 0;
img_res_v.data = nullptr;
if (!clip_image_preprocess(clp_ctx, img.img_data, img_res_v))
{
LOG_TEE("Error processing the given image");
clip_free(clp_ctx);
clip_image_f32_batch_free(img_res_v);
return false;
}
if (img_res_v.size == 0)
{
if (!llava_image_embed_make_with_clip_img(clp_ctx, params.n_threads, img.img_data, &img.image_embedding, &img.image_tokens)) {
LOG_TEE("Error processing the given image");
return false;
}
// note: assumes only one image was returned by clip_image_preprocess
clip_image_f32 * img_res = img_res_v.data;
img.image_tokens = clip_n_patches(clp_ctx);
img.image_embedding = (float *)malloc(clip_embd_nbytes(clp_ctx));
if (!img.image_embedding)
{
LOG_TEE("Unable to allocate memory for image embeddings\n");
clip_image_f32_batch_free(img_res_v);
clip_free(clp_ctx);
return false;
}
LOG_TEE("slot %i - encoding image [id: %i]\n", slot.id, img.id);
if (!clip_image_encode(clp_ctx, params.n_threads, img_res, img.image_embedding))
{
LOG_TEE("Unable to encode image\n");
clip_image_f32_batch_free(img_res_v);
return false;
}
clip_image_f32_batch_free(img_res_v);
img.request_encode_image = false;
}
@@ -1938,8 +1908,9 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
printf(" types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
printf(" -gan N, --grp-attn-n N set the group attention factor to extend context size through self-extend(default: 1=disabled), used together with group attention width `--grp-attn-w`");
printf(" -gaw N, --grp-attn-w N set the group attention width to extend context size through self-extend(default: 512), used together with group attention factor `--grp-attn-n`");
printf(" --chat-template FORMAT_NAME");
printf(" set chat template, possible value is: llama2, chatml (default %s)", sparams.chat_template.c_str());
printf(" --chat-template JINJA_TEMPLATE\n");
printf(" set custom jinja chat template (default: template taken from model's metadata)\n");
printf(" Note: only commonly used templates are accepted, since we don't have jinja parser\n");
printf("\n");
}
@@ -2390,13 +2361,13 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
invalid_param = true;
break;
}
std::string value(argv[i]);
if (value != "chatml" && value != "llama2") {
fprintf(stderr, "error: chat template can be \"llama2\" or \"chatml\", but got: %s\n", value.c_str());
if (!verify_custom_template(argv[i])) {
fprintf(stderr, "error: the supplied chat template is not supported: %s\n", argv[i]);
fprintf(stderr, "note: llama.cpp does not use jinja parser, we only support commonly used templates\n");
invalid_param = true;
break;
}
sparams.chat_template = value;
sparams.chat_template = argv[i];
}
else if (arg == "--override-kv")
{
@@ -2914,7 +2885,7 @@ int main(int argc, char **argv)
if (!validate_api_key(req, res)) {
return;
}
json data = oaicompat_completion_params_parse(json::parse(req.body), sparams.chat_template);
json data = oaicompat_completion_params_parse(llama.model, json::parse(req.body), sparams.chat_template);
const int task_id = llama.queue_tasks.get_new_id();
llama.queue_results.add_waiting_task_id(task_id);
examples/server/utils.hpp
+33 -36
View File
@@ -167,50 +167,47 @@ static T json_value(const json &body, const std::string &key, const T &default_v
: default_value;
}
inline std::string format_llama2(std::vector<json> messages)
{
std::ostringstream output;
bool is_inside_turn = false;
for (auto it = messages.begin(); it != messages.end(); ++it) {
if (!is_inside_turn) {
output << "[INST] ";
}
std::string role = json_value(*it, "role", std::string("user"));
std::string content = json_value(*it, "content", std::string(""));
if (role == "system") {
output << "<<SYS>>\n" << content << "\n<<SYS>>\n\n";
is_inside_turn = true;
} else if (role == "user") {
output << content << " [/INST]";
is_inside_turn = true;
} else {
output << " " << content << " </s>";
is_inside_turn = false;
}
}
LOG_VERBOSE("format_llama2", {{"text", output.str()}});
return output.str();
// Check if the template supplied via "--chat-template" is supported or not. Returns true if it's valid
inline bool verify_custom_template(const std::string & tmpl) {
llama_chat_message chat[] = {{"user", "test"}};
std::vector<char> buf(1);
int res = llama_chat_apply_template(nullptr, tmpl.c_str(), chat, 1, true, buf.data(), buf.size());
return res >= 0;
}
inline std::string format_chatml(std::vector<json> messages)
// Format given chat. If tmpl is empty, we take the template from model metadata
inline std::string format_chat(const struct llama_model * model, const std::string & tmpl, const std::vector<json> & messages)
{
std::ostringstream chatml_msgs;
size_t alloc_size = 0;
// vector holding all allocated string to be passed to llama_chat_apply_template
std::vector<std::string> str(messages.size() * 2);
std::vector<llama_chat_message> chat(messages.size());
for (auto it = messages.begin(); it != messages.end(); ++it) {
chatml_msgs << "<|im_start|>"
<< json_value(*it, "role", std::string("user")) << '\n';
chatml_msgs << json_value(*it, "content", std::string(""))
<< "<|im_end|>\n";
for (size_t i = 0; i < messages.size(); ++i) {
auto &curr_msg = messages[i];
str[i*2 + 0] = json_value(curr_msg, "role", std::string(""));
str[i*2 + 1] = json_value(curr_msg, "content", std::string(""));
alloc_size += str[i*2 + 1].length();
chat[i].role = str[i*2 + 0].c_str();
chat[i].content = str[i*2 + 1].c_str();
}
chatml_msgs << "<|im_start|>assistant" << '\n';
const char * ptr_tmpl = tmpl.empty() ? nullptr : tmpl.c_str();
std::vector<char> buf(alloc_size * 2);
LOG_VERBOSE("format_chatml", {{"text", chatml_msgs.str()}});
// run the first time to get the total output length
int32_t res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
return chatml_msgs.str();
// if it turns out that our buffer is too small, we resize it
if ((size_t) res > buf.size()) {
buf.resize(res);
res = llama_chat_apply_template(model, ptr_tmpl, chat.data(), chat.size(), true, buf.data(), buf.size());
}
std::string formatted_chat(buf.data(), res);
LOG_VERBOSE("formatted_chat", {{"text", formatted_chat.c_str()}});
return formatted_chat;
}
//