tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-07-01 05:05:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into compilade/bitnet-ternary

Browse Source
This commit is contained in:
Francis Couture-Harpin
2024-08-11 15:52:29 -04:00
parent 96b3d411e0 4134999e01
commit d911cd1f13
138 changed files with 7065 additions and 1937 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
examples/baby-llama/baby-llama.cpp
View File

@ -1,7 +1,6 @@
#include "ggml.h"
#include "train.h"
#include <vector>
#include <cassert>
#include <cstdlib>
#include <cstring>

2
examples/batched-bench/batched-bench.cpp
View File

@ -69,7 +69,7 @@ int main(int argc, char ** argv) {
llama_context_params ctx_params = llama_context_params_from_gpt_params(params);
// ensure enough sequences are available
ctx_params.n_seq_max = *std::max_element(n_pl.begin(), n_pl.end());
ctx_params.n_seq_max = n_pl.empty() ? 1 : *std::max_element(n_pl.begin(), n_pl.end());
llama_context * ctx = llama_new_context_with_model(model, ctx_params);

7
examples/cvector-generator/cvector-generator.cpp
View File

@ -414,9 +414,10 @@ int main(int argc, char ** argv) {
llama_numa_init(params.numa);
// load the model to get hparams
llama_model * model;
llama_context * ctx;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
// int n_ctx = llama_n_ctx(ctx);
int n_layers = llama_n_layer(model);

8
examples/embedding/README.md
View File

@ -9,13 +9,13 @@ To get started right away, run the following command, making sure to use the cor
### Unix-based systems (Linux, macOS, etc.):
```bash
./llama-embedding -m ./path/to/model --log-disable -p "Hello World!" 2>/dev/null
./llama-embedding -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>/dev/null
```
### Windows:
```powershell
llama-embedding.exe -m ./path/to/model --log-disable -p "Hello World!" 2>$null
llama-embedding.exe -m ./path/to/model --pooling mean --log-disable -p "Hello World!" 2>$null
```
The above command will output space-separated float values.
@ -50,11 +50,11 @@ The above command will output space-separated float values.
### Unix-based systems (Linux, macOS, etc.):
```bash
./embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
./llama-embedding -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
```
### Windows:
```powershell
embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
llama-embedding.exe -p 'Castle<#sep#>Stronghold<#sep#>Dog<#sep#>Cat' --pooling mean --embd-separator '<#sep#>' --embd-normalize 2 --embd-output-format '' -m './path/to/model.gguf' --n-gpu-layers 99 --log-disable 2>/dev/null
```

152
examples/embedding/embedding.cpp
View File

@ -31,13 +31,24 @@ static void batch_add_seq(llama_batch & batch, const std::vector<int32_t> & toke
}
static void batch_decode(llama_context * ctx, llama_batch & batch, float * output, int n_seq, int n_embd, int embd_norm) {
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
const struct llama_model * model = llama_get_model(ctx);
// clear previous kv_cache values (irrelevant for embeddings)
llama_kv_cache_clear(ctx);
// run model
fprintf(stderr, "%s: n_tokens = %d, n_seq = %d\n", __func__, batch.n_tokens, n_seq);
if (llama_decode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to decode\n", __func__);
if (llama_model_has_encoder(model) && !llama_model_has_decoder(model)) {
// encoder-only model
if (llama_encode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to encode\n", __func__);
}
} else if (!llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
// decoder-only model
if (llama_decode(ctx, batch) < 0) {
fprintf(stderr, "%s : failed to decode\n", __func__);
}
}
for (int i = 0; i < batch.n_tokens; i++) {
@ -45,11 +56,22 @@ static void batch_decode(llama_context * ctx, llama_batch & batch, float * outpu
continue;
}
// try to get sequence embeddings - supported only when pooling_type is not NONE
const float * embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
const float * embd = nullptr;
int embd_pos = 0;
float * out = output + batch.seq_id[i][0] * n_embd;
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
// try to get token embeddings
embd = llama_get_embeddings_ith(ctx, i);
embd_pos = i;
GGML_ASSERT(embd != NULL && "failed to get token embeddings");
} else {
// try to get sequence embeddings - supported only when pooling_type is not NONE
embd = llama_get_embeddings_seq(ctx, batch.seq_id[i][0]);
embd_pos = batch.seq_id[i][0];
GGML_ASSERT(embd != NULL && "failed to get sequence embeddings");
}
float * out = output + embd_pos * n_embd;
llama_embd_normalize(embd, out, n_embd, embd_norm);
}
}
@ -79,11 +101,11 @@ int main(int argc, char ** argv) {
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
return 1;
@ -93,8 +115,9 @@ int main(int argc, char ** argv) {
const int n_ctx = llama_n_ctx(ctx);
const enum llama_pooling_type pooling_type = llama_pooling_type(ctx);
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
fprintf(stderr, "%s: error: pooling type NONE not supported\n", __func__);
if (llama_model_has_encoder(model) && llama_model_has_decoder(model)) {
fprintf(stderr, "%s: error: computing embeddings in encoder-decoder models is not supported\n", __func__);
return 1;
}
@ -153,13 +176,23 @@ int main(int argc, char ** argv) {
const int n_prompts = prompts.size();
struct llama_batch batch = llama_batch_init(n_batch, 0, 1);
// count number of embeddings
int n_embd_count = 0;
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
for (int k = 0; k < n_prompts; k++) {
n_embd_count += inputs[k].size();
}
} else {
n_embd_count = n_prompts;
}
// allocate output
const int n_embd = llama_n_embd(model);
std::vector<float> embeddings(n_prompts * n_embd, 0);
std::vector<float> embeddings(n_embd_count * n_embd, 0);
float * emb = embeddings.data();
// break into batches
int p = 0; // number of prompts processed already
int e = 0; // number of embeddings already stored
int s = 0; // number of prompts in current batch
for (int k = 0; k < n_prompts; k++) {
// clamp to n_batch tokens
@ -169,11 +202,11 @@ int main(int argc, char ** argv) {
// encode if at capacity
if (batch.n_tokens + n_toks > n_batch) {
float * out = emb + p * n_embd;
float * out = emb + e * n_embd;
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
llama_batch_clear(batch);
p += s;
e += pooling_type == LLAMA_POOLING_TYPE_NONE ? batch.n_tokens : s;
s = 0;
llama_batch_clear(batch);
}
// add to batch
@ -182,40 +215,63 @@ int main(int argc, char ** argv) {
}
// final batch
float * out = emb + p * n_embd;
float * out = emb + e * n_embd;
batch_decode(ctx, batch, out, s, n_embd, params.embd_normalize);
if (params.embd_out.empty()) {
// print the first part of the embeddings or for a single prompt, the full embedding
fprintf(stdout, "\n");
for (int j = 0; j < n_prompts; j++) {
fprintf(stdout, "embedding %d: ", j);
for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, "\n");
}
// print cosine similarity matrix
if (n_prompts > 1) {
fprintf(stdout, "\n");
printf("cosine similarity matrix:\n\n");
for (int i = 0; i < n_prompts; i++) {
fprintf(stdout, "%6.6s ", prompts[i].c_str());
}
fprintf(stdout, "\n");
for (int i = 0; i < n_prompts; i++) {
for (int j = 0; j < n_prompts; j++) {
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f ", sim);
if (pooling_type == LLAMA_POOLING_TYPE_NONE) {
for (int j = 0; j < n_embd_count; j++) {
fprintf(stdout, "embedding %d: ", j);
for (int i = 0; i < std::min(3, n_embd); i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, " ... ");
for (int i = n_embd - 3; i < n_embd; i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, "%1.10s", prompts[i].c_str());
fprintf(stdout, "\n");
}
} else {
// print the first part of the embeddings or for a single prompt, the full embedding
for (int j = 0; j < n_prompts; j++) {
fprintf(stdout, "embedding %d: ", j);
for (int i = 0; i < (n_prompts > 1 ? std::min(16, n_embd) : n_embd); i++) {
if (params.embd_normalize == 0) {
fprintf(stdout, "%6.0f ", emb[j * n_embd + i]);
} else {
fprintf(stdout, "%9.6f ", emb[j * n_embd + i]);
}
}
fprintf(stdout, "\n");
}
// print cosine similarity matrix
if (n_prompts > 1) {
fprintf(stdout, "\n");
printf("cosine similarity matrix:\n\n");
for (int i = 0; i < n_prompts; i++) {
fprintf(stdout, "%6.6s ", prompts[i].c_str());
}
fprintf(stdout, "\n");
for (int i = 0; i < n_prompts; i++) {
for (int j = 0; j < n_prompts; j++) {
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f ", sim);
}
fprintf(stdout, "%1.10s", prompts[i].c_str());
fprintf(stdout, "\n");
}
}
}
}
@ -233,23 +289,23 @@ int main(int argc, char ** argv) {
}
fprintf(stdout, notArray ? "]\n }" : "]");
j++;
if (j < n_prompts) fprintf(stdout, notArray ? ",\n" : ","); else break;
if (j < n_embd_count) fprintf(stdout, notArray ? ",\n" : ","); else break;
}
fprintf(stdout, notArray ? "\n ]" : "]\n");
if (params.embd_out == "json+" && n_prompts > 1) {
fprintf(stdout, ",\n \"cosineSimilarity\": [\n");
for (int i = 0;;) { // at least two iteration (n_prompts > 1)
for (int i = 0;;) { // at least two iteration (n_embd_count > 1)
fprintf(stdout, " [");
for (int j = 0;;) { // at least two iteration (n_prompts > 1)
for (int j = 0;;) { // at least two iteration (n_embd_count > 1)
float sim = llama_embd_similarity_cos(emb + i * n_embd, emb + j * n_embd, n_embd);
fprintf(stdout, "%6.2f", sim);
j++;
if (j < n_prompts) fprintf(stdout, ", "); else break;
if (j < n_embd_count) fprintf(stdout, ", "); else break;
}
fprintf(stdout, " ]");
i++;
if (i < n_prompts) fprintf(stdout, ",\n"); else break;
if (i < n_embd_count) fprintf(stdout, ",\n"); else break;
}
fprintf(stdout, "\n ]");
}

7
examples/eval-callback/eval-callback.cpp
View File

@ -163,9 +163,10 @@ int main(int argc, char ** argv) {
params.warmup = false;
// init
llama_model * model;
llama_context * ctx;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
return 1;

24
examples/export-lora/export-lora.cpp
View File

@ -50,20 +50,6 @@ static struct gguf_context * load_gguf(std::string & fname, struct ggml_context
return ctx_gguf;
}
static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
std::string result;
for (size_t pos = 0; ; pos += search.length()) {
auto new_pos = s.find(search, pos);
if (new_pos == std::string::npos) {
result += s.substr(pos, s.size() - pos);
break;
}
result += s.substr(pos, new_pos - pos) + replace;
pos = new_pos;
}
s = std::move(result);
}
struct file_input {
struct ggml_context * ctx_meta = nullptr;
struct gguf_context * ctx_gguf = nullptr;
@ -135,7 +121,7 @@ struct lora_merge_ctx {
lora_merge_ctx(
std::string & base_fname,
std::vector<std::tuple<std::string, float>> & lora_files,
std::vector<llama_lora_adapter_info> & lora_files,
std::string & outfile,
int n_threads) : base_model(base_fname, 0), n_threads(n_threads), fout(outfile, std::ios::binary) {
fout.exceptions(std::ofstream::failbit); // fail fast on write errors
@ -144,9 +130,9 @@ struct lora_merge_ctx {
throw std::runtime_error("split model is not yet supported");
}
for (auto lora_inp : lora_files) {
auto fname = std::get<0>(lora_inp);
auto scale = std::get<1>(lora_inp);
for (auto & lora_inp : lora_files) {
auto fname = lora_inp.path;
auto scale = lora_inp.scale;
std::unique_ptr<file_input> adapter(new file_input(fname, scale));
check_metadata_lora(adapter.get());
adapters.push_back(std::move(adapter));
@ -407,7 +393,7 @@ int main(int argc, char ** argv) {
g_verbose = (params.verbosity == 1);
try {
lora_merge_ctx ctx(params.model, params.lora_adapter, params.lora_outfile, params.n_threads);
lora_merge_ctx ctx(params.model, params.lora_adapters, params.lora_outfile, params.n_threads);
ctx.run_merge();
} catch (const std::exception & err) {
fprintf(stderr, "%s\n", err.what());

6
examples/imatrix/imatrix.cpp
View File

@ -611,10 +611,10 @@ int main(int argc, char ** argv) {
params.warmup = false;
// init
llama_model * model;
llama_context * ctx;
llama_init_result llama_init = llama_init_from_gpt_params(params);
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
return 1;

5
examples/infill/infill.cpp
View File

@ -179,7 +179,10 @@ int main(int argc, char ** argv) {
// load the model and apply lora adapter, if any
LOG("%s: load the model and apply lora adapter, if any\n", __func__);
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
model = llama_init.model;
ctx = llama_init.context;
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n", __func__);

29
examples/llama-bench/llama-bench.cpp
View File

@ -27,6 +27,14 @@
#include "ggml-cann.h"
#endif
#ifdef _WIN32
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
# define NOMINMAX
#endif
#include <windows.h>
#endif
// utils
static uint64_t get_time_ns() {
using clock = std::chrono::high_resolution_clock;
@ -96,6 +104,27 @@ static std::string get_cpu_info() {
}
fclose(f);
}
#elif defined(_WIN32)
HKEY hKey;
if (RegOpenKeyEx(HKEY_LOCAL_MACHINE,
TEXT("HARDWARE\\DESCRIPTION\\System\\CentralProcessor\\0"),
0,
KEY_READ,
&hKey) != ERROR_SUCCESS) {
// fail to open registry key
return "";
}
char cpu_brand[256];
DWORD cpu_brand_size = sizeof(cpu_brand);
if (RegQueryValueExA(hKey,
TEXT("ProcessorNameString"),
NULL,
NULL,
(LPBYTE)cpu_brand,
&cpu_brand_size) == ERROR_SUCCESS) {
id.assign(cpu_brand, cpu_brand_size);
}
RegCloseKey(hKey);
#endif
// TODO: other platforms
return id;

7
examples/llava/CMakeLists.txt
View File

@ -36,3 +36,10 @@ set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)
set(TARGET llama-minicpmv-cli)
add_executable(${TARGET} minicpmv-cli.cpp)
set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-minicpmv-cli)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_11)

99
examples/llava/README-minicpmv2.5.md Normal file
View File

@ -0,0 +1,99 @@
## MiniCPM-Llama3-V 2.5
### Prepare models and code
Download [MiniCPM-Llama3-V-2_5](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5) PyTorch model from huggingface to "MiniCPM-Llama3-V-2_5" folder.
Clone llama.cpp:
```bash
git clone https://github.com/ggerganov/llama.cpp
cd llama.cpp
```
### Usage
Convert PyTorch model to gguf files (You can also download the converted [gguf](https://huggingface.co/openbmb/MiniCPM-Llama3-V-2_5-gguf) by us)
```bash
python ./examples/minicpmv/minicpmv-surgery.py -m ../MiniCPM-Llama3-V-2_5
python ./examples/minicpmv/minicpmv-convert-image-encoder-to-gguf.py -m ../MiniCPM-Llama3-V-2_5 --minicpmv-projector ../MiniCPM-Llama3-V-2_5/minicpmv.projector --output-dir ../MiniCPM-Llama3-V-2_5/ --image-mean 0.5 0.5 0.5 --image-std 0.5 0.5 0.5
python ./convert-hf-to-gguf.py ../MiniCPM-Llama3-V-2_5/model
# quantize int4 version
./llama-quantize ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf Q4_K_M
```
Build for Linux or Mac
```bash
make
make llama-minicpmv-cli
```
Inference on Linux or Mac
```
# run f16 version
./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/model-8B-F16.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
# run quantized int4 version
./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
# or run in interactive mode
./llama-minicpmv-cli -m ../MiniCPM-Llama3-V-2_5/model/ggml-model-Q4_K_M.gguf --mmproj ../MiniCPM-Llama3-V-2_5/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -i
```
### Android
#### Build on Android device using Termux
We found that build on Android device would bring better runtime performance, so we recommend to build on device.
[Termux](https://github.com/termux/termux-app#installation) is a terminal app on Android device (no root required).
Install tools in Termux:
```
apt update && apt upgrade -y
apt install git make cmake
```
It's recommended to move your model inside the `~/` directory for best performance:
```
cd storage/downloads
mv model.gguf ~/
```
#### Building the Project using Android NDK
Obtain the [Android NDK](https://developer.android.com/ndk) and then build with CMake.
Execute the following commands on your computer to avoid downloading the NDK to your mobile. Alternatively, you can also do this in Termux:
```bash
mkdir build-android
cd build-android
export NDK=/your_ndk_path
cmake -DCMAKE_TOOLCHAIN_FILE=$NDK/build/cmake/android.toolchain.cmake -DANDROID_ABI=arm64-v8a -DANDROID_PLATFORM=android-23 -DCMAKE_C_FLAGS=-march=armv8.4a+dotprod ..
make
```
Install [termux](https://github.com/termux/termux-app#installation) on your device and run `termux-setup-storage` to get access to your SD card (if Android 11+ then run the command twice).
Finally, copy these built `llama` binaries and the model file to your device storage. Because the file permissions in the Android sdcard cannot be changed, you can copy the executable files to the `/data/data/com.termux/files/home/bin` path, and then execute the following commands in Termux to add executable permission:
(Assumed that you have pushed the built executable files to the /sdcard/llama.cpp/bin path using `adb push`)
```
$cp -r /sdcard/llama.cpp/bin /data/data/com.termux/files/home/
$cd /data/data/com.termux/files/home/bin
$chmod +x ./*
```
Download models and push them to `/sdcard/llama.cpp/`, then move it to `/data/data/com.termux/files/home/model/`
```
$mv /sdcard/llama.cpp/ggml-model-Q4_K_M.gguf /data/data/com.termux/files/home/model/
$mv /sdcard/llama.cpp/mmproj-model-f16.gguf /data/data/com.termux/files/home/model/
```
Now, you can start chatting:
```
$cd /data/data/com.termux/files/home/bin
$./llama-minicpmv-cli -m ../model/ggml-model-Q4_K_M.gguf --mmproj ../model/mmproj-model-f16.gguf -c 4096 --temp 0.7 --top-p 0.8 --top-k 100 --repeat-penalty 1.05 --image xx.jpg -p "What is in the image?"
```

633
examples/llava/clip.cpp
View File

@ -74,26 +74,27 @@ static std::string format(const char * fmt, ...) {
// key constants
//
#define KEY_FTYPE "general.file_type"
#define KEY_NAME "general.name"
#define KEY_DESCRIPTION "general.description"
#define KEY_HAS_TEXT_ENC "clip.has_text_encoder"
#define KEY_HAS_VIS_ENC "clip.has_vision_encoder"
#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
#define KEY_USE_GELU "clip.use_gelu"
#define KEY_N_EMBD "clip.%s.embedding_length"
#define KEY_N_FF "clip.%s.feed_forward_length"
#define KEY_N_BLOCK "clip.%s.block_count"
#define KEY_N_HEAD "clip.%s.attention.head_count"
#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
#define KEY_PROJ_DIM "clip.%s.projection_dim"
#define KEY_TOKENS "tokenizer.ggml.tokens"
#define KEY_N_POSITIONS "clip.text.context_length"
#define KEY_IMAGE_SIZE "clip.vision.image_size"
#define KEY_PATCH_SIZE "clip.vision.patch_size"
#define KEY_IMAGE_MEAN "clip.vision.image_mean"
#define KEY_IMAGE_STD "clip.vision.image_std"
#define KEY_PROJ_TYPE "clip.projector_type"
#define KEY_FTYPE "general.file_type"
#define KEY_NAME "general.name"
#define KEY_DESCRIPTION "general.description"
#define KEY_HAS_TEXT_ENC "clip.has_text_encoder"
#define KEY_HAS_VIS_ENC "clip.has_vision_encoder"
#define KEY_HAS_LLAVA_PROJ "clip.has_llava_projector"
#define KEY_HAS_MINICPMV_PROJ "clip.has_minicpmv_projector"
#define KEY_USE_GELU "clip.use_gelu"
#define KEY_N_EMBD "clip.%s.embedding_length"
#define KEY_N_FF "clip.%s.feed_forward_length"
#define KEY_N_BLOCK "clip.%s.block_count"
#define KEY_N_HEAD "clip.%s.attention.head_count"
#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
#define KEY_PROJ_DIM "clip.%s.projection_dim"
#define KEY_TOKENS "tokenizer.ggml.tokens"
#define KEY_N_POSITIONS "clip.text.context_length"
#define KEY_IMAGE_SIZE "clip.vision.image_size"
#define KEY_PATCH_SIZE "clip.vision.patch_size"
#define KEY_IMAGE_MEAN "clip.vision.image_mean"
#define KEY_IMAGE_STD "clip.vision.image_std"
#define KEY_PROJ_TYPE "clip.projector_type"
#define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type"
#define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints"
@ -127,12 +128,20 @@ static std::string format(const char * fmt, ...) {
#define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s"
#define TN_IMAGE_NEWLINE "model.image_newline"
#define TN_MINICPMV_POS_EMBD_K "resampler.pos_embed_k"
#define TN_MINICPMV_QUERY "resampler.query"
#define TN_MINICPMV_PROJ "resampler.proj.weight"
#define TN_MINICPMV_KV_PROJ "resampler.kv.weight"
#define TN_MINICPMV_ATTN "resampler.attn.%s.%s"
#define TN_MINICPMV_LN "resampler.ln_%s.%s"
enum projector_type {
PROJECTOR_TYPE_MLP,
PROJECTOR_TYPE_MLP_NORM,
PROJECTOR_TYPE_LDP,
PROJECTOR_TYPE_LDPV2,
PROJECTOR_TYPE_RESAMPLER,
PROJECTOR_TYPE_UNKNOWN,
};
@ -140,6 +149,7 @@ static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_MLP, "mlp" },
{ PROJECTOR_TYPE_LDP, "ldp" },
{ PROJECTOR_TYPE_LDPV2, "ldpv2"},
{ PROJECTOR_TYPE_RESAMPLER, "resampler"},
};
@ -200,17 +210,14 @@ static std::string gguf_data_to_str(enum gguf_type type, const void * data, int
}
static void replace_all(std::string & s, const std::string & search, const std::string & replace) {
std::string result;
for (size_t pos = 0; ; pos += search.length()) {
auto new_pos = s.find(search, pos);
if (new_pos == std::string::npos) {
result += s.substr(pos, s.size() - pos);
break;
}
result += s.substr(pos, new_pos - pos) + replace;
pos = new_pos;
if (search.empty()) {
return; // Avoid infinite loop if 'search' is an empty string
}
size_t pos = 0;
while ((pos = s.find(search, pos)) != std::string::npos) {
s.replace(pos, search.length(), replace);
pos += replace.length();
}
s = std::move(result);
}
static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
@ -492,12 +499,33 @@ struct clip_vision_model {
struct ggml_tensor * mm_model_mlp_2_b;
struct ggml_tensor * mm_model_peg_0_w;
struct ggml_tensor * mm_model_peg_0_b;
// MINICPMV projection
struct ggml_tensor * mm_model_pos_embed_k;
struct ggml_tensor * mm_model_query;
struct ggml_tensor * mm_model_proj;
struct ggml_tensor * mm_model_kv_proj;
struct ggml_tensor * mm_model_attn_q_w;
struct ggml_tensor * mm_model_attn_q_b;
struct ggml_tensor * mm_model_attn_k_w;
struct ggml_tensor * mm_model_attn_k_b;
struct ggml_tensor * mm_model_attn_v_w;
struct ggml_tensor * mm_model_attn_v_b;
struct ggml_tensor * mm_model_attn_o_w;
struct ggml_tensor * mm_model_attn_o_b;
struct ggml_tensor * mm_model_ln_q_w;
struct ggml_tensor * mm_model_ln_q_b;
struct ggml_tensor * mm_model_ln_kv_w;
struct ggml_tensor * mm_model_ln_kv_b;
struct ggml_tensor * mm_model_ln_post_w;
struct ggml_tensor * mm_model_ln_post_b;
};
struct clip_ctx {
bool has_text_encoder = false;
bool has_vision_encoder = false;
bool has_llava_projector = false;
bool has_minicpmv_projector = false;
struct clip_vision_model vision_model;
projector_type proj_type = PROJECTOR_TYPE_MLP;
@ -522,9 +550,11 @@ struct clip_ctx {
ggml_backend_t backend = NULL;
ggml_gallocr_t compute_alloc = NULL;
struct clip_image_size * load_image_size;
};
static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs) {
static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32_batch * imgs, struct clip_image_size * load_image_size, bool is_inf = false) {
if (!ctx->has_vision_encoder) {
LOG_TEE("This gguf file seems to have no vision encoder\n");
return nullptr;
@ -533,20 +563,33 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
const auto & model = ctx->vision_model;
const auto & hparams = model.hparams;
const int image_size = hparams.image_size;
const int image_size = hparams.image_size;
int image_size_width = image_size;
int image_size_height = image_size;
if (ctx->has_minicpmv_projector) {
if (load_image_size == nullptr) {
load_image_size = clip_image_size_init();
}
LOG_TEE("%s: %d %d\n", __func__, load_image_size->width, load_image_size->height);
image_size_width = load_image_size->width;
image_size_height = load_image_size->height;
if (is_inf) {
image_size_width = imgs->data->nx;
image_size_height = imgs->data->ny;
}
}
const int patch_size = hparams.patch_size;
const int num_patches = ((image_size / patch_size) * (image_size / patch_size));
const int num_patches_per_side = image_size / patch_size; GGML_UNUSED(num_patches_per_side);
const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
const int hidden_size = hparams.hidden_size;
const int n_head = hparams.n_head;
const int d_head = hidden_size / n_head;
const int n_layer = hparams.n_layer;
int n_layer = hparams.n_layer;
const float eps = hparams.eps;
const int batch_size = imgs->size;
if (ctx->has_llava_projector) {
if (ctx->has_llava_projector || ctx->has_minicpmv_projector) {
GGML_ASSERT(batch_size == 1);
}
@ -559,7 +602,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph * gf = ggml_new_graph(ctx0);
struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size, image_size, 3, batch_size);
struct ggml_tensor * inp_raw = ggml_new_tensor_4d(ctx0, GGML_TYPE_F32, image_size_width, image_size_height, 3, batch_size);
ggml_set_name(inp_raw, "inp_raw");
ggml_set_input(inp_raw);
@ -572,19 +615,21 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
// inp = ggml_add(ctx0, inp, ggml_repeat(ctx0, model.patch_bias, inp));
inp = ggml_add(ctx0, inp, model.patch_bias);
}
// concat class_embeddings and patch_embeddings
struct ggml_tensor * embeddings = inp;
if (ctx->has_class_embedding) {
embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
ggml_set_name(embeddings, "embeddings");
ggml_set_input(embeddings);
embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
embeddings = ggml_acc(ctx0, embeddings, inp,
embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
}
struct ggml_tensor * pos_embed = nullptr;
if (ctx->has_llava_projector) {
// concat class_embeddings and patch_embeddings
if (ctx->has_class_embedding) {
embeddings = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, hidden_size, num_positions, batch_size);
ggml_set_name(embeddings, "embeddings");
ggml_set_input(embeddings);
embeddings = ggml_acc(ctx0, embeddings, model.class_embedding,
embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], 0);
embeddings = ggml_acc(ctx0, embeddings, inp,
embeddings->nb[1], embeddings->nb[2], embeddings->nb[3], model.class_embedding->nb[1]);
}
}
struct ggml_tensor * positions = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_positions);
ggml_set_name(positions, "positions");
@ -593,6 +638,14 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
embeddings =
ggml_add(ctx0, embeddings, ggml_get_rows(ctx0, model.position_embeddings, positions));
if (ctx->has_minicpmv_projector) {
int pos_w = image_size_width/patch_size;
int pos_h = image_size_height/patch_size;
pos_embed = ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 4096, pos_w * pos_h, 1);
ggml_set_name(pos_embed, "pos_embed");
ggml_set_input(pos_embed);
}
// pre-layernorm
if (ctx->has_pre_norm) {
embeddings = ggml_norm(ctx0, embeddings, eps);
@ -602,6 +655,9 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
}
// loop over layers
if (ctx->has_minicpmv_projector) {
n_layer += 1;
}
for (int il = 0; il < n_layer - 1; il++) {
struct ggml_tensor * cur = embeddings; // embeddings = residual, cur = hidden_states
@ -691,7 +747,7 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
}
// llava projector
{
if (ctx->has_llava_projector) {
embeddings = ggml_reshape_2d(ctx0, embeddings, embeddings->ne[0], embeddings->ne[1]);
struct ggml_tensor * patches = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, num_patches);
@ -872,6 +928,65 @@ static ggml_cgraph * clip_image_build_graph(clip_ctx * ctx, const clip_image_f32
GGML_ABORT("fatal error");
}
}
// minicpmv projector
else if (ctx->has_minicpmv_projector)
{
if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
struct ggml_tensor * q = model.mm_model_query;
{ // layernorm
q = ggml_norm(ctx0, q, eps);
q = ggml_add(ctx0, ggml_mul(ctx0, q, model.mm_model_ln_q_w), model.mm_model_ln_q_b);
}
struct ggml_tensor * v = ggml_mul_mat(ctx0, model.mm_model_kv_proj, embeddings);
{ // layernorm
v = ggml_norm(ctx0, v, eps);
v = ggml_add(ctx0, ggml_mul(ctx0, v, model.mm_model_ln_kv_w), model.mm_model_ln_kv_b);
}
struct ggml_tensor * k;
{ // position
// q = ggml_add(ctx0, q, model.mm_model_pos_embed);
k = ggml_add(ctx0, v, pos_embed);
}
{ // attention
const int hidden_size = 4096;
const int d_head = 128;
const int n_head = hidden_size/d_head;
const int num_query = 96;
struct ggml_tensor * Q = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_q_w, q), model.mm_model_attn_q_b);
Q = ggml_scale_inplace(ctx0, Q, 1.0f / sqrt((float)d_head));
struct ggml_tensor * K = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_k_w, k), model.mm_model_attn_k_b);
struct ggml_tensor * V = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_v_w, v), model.mm_model_attn_v_b);
// permute
Q = ggml_reshape_4d(ctx0, Q, d_head, n_head, num_query, batch_size);
Q = ggml_cont(ctx0, ggml_permute(ctx0, Q, 0, 2, 1, 3));
Q = ggml_reshape_3d(ctx0, Q, d_head, num_query, n_head * batch_size);
K = ggml_reshape_4d(ctx0, K, d_head, n_head, num_positions, batch_size);
K = ggml_cont(ctx0, ggml_permute(ctx0, K, 0, 2, 1, 3));
K = ggml_reshape_3d(ctx0, K, d_head, num_positions, n_head * batch_size);
V = ggml_reshape_4d(ctx0, V, d_head, n_head, num_positions, batch_size);
V = ggml_cont(ctx0, ggml_permute(ctx0, V, 1, 2, 0, 3));
V = ggml_reshape_3d(ctx0, V, num_positions, d_head, n_head * batch_size);
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
KQ = ggml_soft_max_inplace(ctx0, KQ);
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ);
KQV = ggml_reshape_4d(ctx0, KQV, d_head, num_query, n_head, batch_size);
KQV = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
KQV = ggml_cont_3d(ctx0, KQV, hidden_size, num_query, batch_size);
embeddings = ggml_add(ctx0, ggml_mul_mat(ctx0, model.mm_model_attn_o_w, KQV), model.mm_model_attn_o_b);
}
{ // layernorm
embeddings = ggml_norm(ctx0, embeddings, eps);
embeddings = ggml_add(ctx0, ggml_mul(ctx0, embeddings, model.mm_model_ln_post_w), model.mm_model_ln_post_b);
}
embeddings = ggml_mul_mat(ctx0, model.mm_model_proj, embeddings);
}
else {
GGML_ASSERT(false);
}
}
// build the graph
ggml_build_forward_expand(gf, embeddings);
@ -1029,7 +1144,13 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
new_clip->has_llava_projector = gguf_get_val_bool(ctx, idx);
}
GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
idx = gguf_find_key(ctx, KEY_HAS_MINICPMV_PROJ);
if (idx != -1) {
new_clip->has_minicpmv_projector = gguf_get_val_bool(ctx, idx);
}
// GGML_ASSERT(new_clip->has_llava_projector); // see monatis/clip.cpp for image and/or text encoding for semantic search
GGML_ASSERT(new_clip->has_vision_encoder);
GGML_ASSERT(!new_clip->has_text_encoder);
@ -1040,6 +1161,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
LOG_TEE("%s: text_encoder: %d\n", __func__, new_clip->has_text_encoder);
LOG_TEE("%s: vision_encoder: %d\n", __func__, new_clip->has_vision_encoder);
LOG_TEE("%s: llava_projector: %d\n", __func__, new_clip->has_llava_projector);
LOG_TEE("%s: minicpmv_projector: %d\n", __func__, new_clip->has_minicpmv_projector);
LOG_TEE("%s: model size: %.2f MB\n", __func__, model_size / 1024.0 / 1024.0);
LOG_TEE("%s: metadata size: %.2f MB\n", __func__, ggml_get_mem_size(meta) / 1024.0 / 1024.0);
}
@ -1281,6 +1403,27 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
vision_model.mm_model_peg_0_w = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "weight"));
vision_model.mm_model_peg_0_b = get_tensor(new_clip->ctx_data, format(TN_MVLM_PROJ_PEG, 0, "bias"));
}
else if (new_clip->proj_type == PROJECTOR_TYPE_RESAMPLER) {
// vision_model.mm_model_pos_embed = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD);
vision_model.mm_model_pos_embed_k = get_tensor(new_clip->ctx_data, TN_MINICPMV_POS_EMBD_K);
vision_model.mm_model_query = get_tensor(new_clip->ctx_data, TN_MINICPMV_QUERY);
vision_model.mm_model_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_PROJ);
vision_model.mm_model_kv_proj = get_tensor(new_clip->ctx_data, TN_MINICPMV_KV_PROJ);
vision_model.mm_model_attn_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "weight"));
vision_model.mm_model_attn_k_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "weight"));
vision_model.mm_model_attn_v_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "weight"));
vision_model.mm_model_attn_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "q", "bias"));
vision_model.mm_model_attn_k_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "k", "bias"));
vision_model.mm_model_attn_v_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "v", "bias"));
vision_model.mm_model_attn_o_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "weight"));
vision_model.mm_model_attn_o_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_ATTN, "out", "bias"));
vision_model.mm_model_ln_q_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "weight"));
vision_model.mm_model_ln_q_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "q", "bias"));
vision_model.mm_model_ln_kv_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "weight"));
vision_model.mm_model_ln_kv_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "kv", "bias"));
vision_model.mm_model_ln_post_w = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "weight"));
vision_model.mm_model_ln_post_b = get_tensor(new_clip->ctx_data, format(TN_MINICPMV_LN, "post", "bias"));
}
else {
std::string proj_type = PROJECTOR_TYPE_NAMES[new_clip->proj_type];
throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
@ -1319,7 +1462,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
new_clip->compute_alloc = ggml_gallocr_new(ggml_backend_get_default_buffer_type(new_clip->backend));
clip_image_f32_batch batch;
batch.size = 1;
ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch);
ggml_cgraph * gf = clip_image_build_graph(new_clip, &batch, nullptr, false);
ggml_gallocr_reserve(new_clip->compute_alloc, gf);
size_t compute_memory_buffer_size = ggml_gallocr_get_buffer_size(new_clip->compute_alloc, 0);
LOG_TEE("%s: compute allocated memory: %.2f MB\n", __func__, compute_memory_buffer_size /1024.0/1024.0);
@ -1328,6 +1471,17 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
return new_clip;
}
void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size) {
ctx_clip->load_image_size = load_image_size;
}
struct clip_image_size * clip_image_size_init() {
struct clip_image_size * load_image_size = new struct clip_image_size();
load_image_size->width = 448;
load_image_size->height = 448;
return load_image_size;
}
struct clip_image_u8 * clip_image_u8_init() {
return new clip_image_u8();
}
@ -1598,9 +1752,184 @@ static std::vector<clip_image_u8*> divide_to_patches_u8(const clip_image_u8 & im
return patches;
}
static int ensure_divide(int length, int patch_size) {
return std::max(static_cast<int>(std::round(static_cast<float>(length) / patch_size) * patch_size), patch_size);
}
static std::pair<int, int> uhd_find_best_resize(std::pair<int, int> original_size, int scale_resolution, int patch_size, bool allow_upscale = false) {
int width = original_size.first;
int height = original_size.second;
if ((width * height > scale_resolution * scale_resolution) || allow_upscale) {
float r = static_cast<float>(width) / height;
height = static_cast<int>(scale_resolution / std::sqrt(r));
width = static_cast<int>(height * r);
}
int best_width = ensure_divide(width, patch_size);
int best_height = ensure_divide(height, patch_size);
return std::make_pair(best_width, best_height);
}
static std::pair<int, int> uhd_get_refine_size(std::pair<int, int> original_size, std::pair<int, int> grid, int scale_resolution, int patch_size, bool allow_upscale = false) {
int width, height;
std::tie(width, height) = original_size;
int grid_x, grid_y;
std::tie(grid_x, grid_y) = grid;
int refine_width = ensure_divide(width, grid_x);
int refine_height = ensure_divide(height, grid_y);
int grid_width = refine_width / grid_x;
int grid_height = refine_height / grid_y;
// auto best_grid_size = find_best_resize(std::make_tuple(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); (old line)
auto best_grid_size = uhd_find_best_resize(std::make_pair(grid_width, grid_height), scale_resolution, patch_size, allow_upscale); // (new line) => fixes conversion for make_tuple to make_pair
int best_grid_width, best_grid_height;
std::tie(best_grid_width, best_grid_height) = best_grid_size;
// std::pair<int, int> refine_size = std::make_tuple(best_grid_width * grid_x, best_grid_height * grid_y); (old line)
std::pair<int, int> refine_size = std::make_pair(best_grid_width * grid_x, best_grid_height * grid_y); // (new line)
return refine_size;
}
inline int clip(int x, int lower, int upper) {
return std::max(lower, std::min(x, upper));
}
static std::pair<int, int> uhd_best_grid(const int max_slice_nums, const int multiple, const float log_ratio) {
std::vector<int> candidate_split_grids_nums;
for (int i : {multiple - 1, multiple, multiple + 1}) {
if (i == 1 || i > max_slice_nums) {
continue;
}
candidate_split_grids_nums.push_back(i);
}
std::vector<std::pair<int, int>> candidate_grids;
for (int split_grids_nums : candidate_split_grids_nums) {
int m = 1;
while (m <= split_grids_nums) {
if (split_grids_nums % m == 0) {
candidate_grids.emplace_back(m, split_grids_nums / m);
}
++m;
}
}
std::pair<int, int> best_grid{1, 1};
float min_error = std::numeric_limits<float>::infinity();
for (const auto& grid : candidate_grids) {
float error = std::abs(log_ratio - std::log(1.0 * grid.first / grid.second));
if (error < min_error) {
best_grid = grid;
min_error = error;
}
}
return best_grid;
}
// inspired from LLaVA-UHD:
// -> https://arxiv.org/pdf/2403.11703
// -> https://github.com/thunlp/LLaVA-UHD
// -> https://github.com/thunlp/LLaVA-UHD/blob/302301bc2175f7e717fb8548516188e89f649753/llava_uhd/train/llava-uhd/slice_logic.py#L118
static std::vector<std::vector<clip_image_u8 *>> uhd_slice_image(const clip_image_u8 * img, const int max_slice_nums=9, const int scale_resolution=448, const int patch_size=14) {
const std::pair<int, int> original_size={img->nx,img->ny};
const int original_width = img->nx;
const int original_height = img->ny;
const float log_ratio = log(1.0*original_width/original_height);
const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
const int multiple = fmin(ceil(ratio), max_slice_nums);
std::vector<std::vector<clip_image_u8 *>> images;
LOG_TEE("%s: multiple %d\n", __func__, multiple);
images.push_back(std::vector<clip_image_u8 *>());
if (multiple <= 1) {
auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size, true);
clip_image_u8 * source_image = clip_image_u8_init();
bicubic_resize(*img, *source_image, best_size.first, best_size.second);
// source_image = image.resize(best_size, Image.Resampling.BICUBIC)
images[images.size()-1].push_back(source_image);
}
else if (multiple > 1) {
auto best_size = uhd_find_best_resize(original_size, scale_resolution, patch_size);
clip_image_u8 * source_image = clip_image_u8_init();
bicubic_resize(*img, *source_image, best_size.first, best_size.second);
// source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
LOG_TEE("%s: image_size: %d %d; source_image size: %d %d\n", __func__, img->nx, img->ny, best_size.first, best_size.second);
images[images.size()-1].push_back(source_image);
std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
LOG_TEE("%s: image_size: %d %d; best_grid: %d %d\n", __func__, img->nx, img->ny, best_grid.first, best_grid.second);
auto refine_size = uhd_get_refine_size(original_size, best_grid, scale_resolution, patch_size, true);
clip_image_u8 * refine_image = clip_image_u8_init();
bicubic_resize(*img, *refine_image, refine_size.first, refine_size.second);
LOG_TEE("%s: refine_image_size: %d %d; refine_size: %d %d\n", __func__, refine_image->nx, refine_image->ny, refine_size.first, refine_size.second);
// split_to_patches
int width = refine_image->nx;
int height = refine_image->ny;
int grid_x = int(width / best_grid.first);
int grid_y = int(height / best_grid.second);
for (int patches_i = 0, ic = 0; patches_i < height && ic < best_grid.second; patches_i += grid_y, ic += 1){
images.push_back(std::vector<clip_image_u8 *>());
for(int patches_j = 0, jc = 0; patches_j < width && jc < best_grid.first; patches_j += grid_x, jc += 1){
clip_image_u8 * patch = clip_image_u8_init();
patch->nx = grid_x;
patch->ny = grid_y;
patch->buf.resize(3 * patch->nx * patch->ny);
for (int y = patches_i; y < patches_i + grid_y; ++y) {
for (int x = patches_j; x < patches_j + grid_x; ++x) {
const int i = 3 * (y * refine_image->nx + x);
const int j = 3 * ((y-patches_i) * patch->nx + (x-patches_j));
patch->buf[j] = refine_image->buf[i];
patch->buf[j+1] = refine_image->buf[i+1];
patch->buf[j+2] = refine_image->buf[i+2];
}
}
images[images.size()-1].push_back(patch);
}
}
}
return images;
}
int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip) {
const int max_slice_nums=9;
const int scale_resolution=448;
const int original_width = ctx_clip->load_image_size->width;
const int original_height = ctx_clip->load_image_size->height;
const float log_ratio = log(1.0*original_width/original_height);
const float ratio = 1.0 * original_width * original_height/ (scale_resolution * scale_resolution);
const int multiple = fmin(ceil(ratio), max_slice_nums);
std::pair<int, int> best_grid = uhd_best_grid(max_slice_nums, multiple, log_ratio);
return best_grid.first;
}
// returns the normalized float tensor for llava-1.5, for spatial_unpad with anyres processing for llava-1.6 it returns the normalized image patch tensors as a vector
// res_imgs memory is being allocated here, previous allocations will be freed if found
bool clip_image_preprocess(struct clip_ctx * ctx, const clip_image_u8 * img, clip_image_f32_batch * res_imgs) {
if (clip_is_minicpmv(ctx)) {
std::vector<std::vector<clip_image_u8 *>> imgs = uhd_slice_image(img);
res_imgs->size = 0;
for (size_t i = 0; i < imgs.size(); ++i) {
res_imgs->size += imgs[i].size();
}
res_imgs->data = new clip_image_f32[res_imgs->size];
int idx = 0;
for (size_t i = 0; i < imgs.size(); ++i) {
for (size_t j = 0; j < imgs[i].size(); ++j) {
LOG_TEE("%s: %d %d\n", __func__,imgs[i][j]->nx,imgs[i][j]->ny);
clip_image_f32 * res = clip_image_f32_init();
normalize_image_u8_to_f32(imgs[i][j], res, ctx->image_mean, ctx->image_std);
res_imgs->data[idx++] = *res;
clip_image_f32_free(res);
}
}
return true;
}
bool pad_to_square = true;
if (!ctx->has_vision_encoder) {
LOG_TEE("This gguf file seems to have no vision encoder\n");
@ -1816,11 +2145,99 @@ int clip_n_patches(const struct clip_ctx * ctx) {
if (ctx->proj_type == PROJECTOR_TYPE_LDP || ctx->proj_type == PROJECTOR_TYPE_LDPV2) {
n_patches /= 4;
} else if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
n_patches = 96;
}
return n_patches;
}
static std::vector<std::vector<std::vector<float>>> get_1d_sincos_pos_embed_from_grid_new(int embed_dim, const std::vector<std::vector<float>> & pos) {
assert(embed_dim % 2 == 0);
int H = pos.size();
int W = pos[0].size();
std::vector<float> omega(embed_dim / 2);
for (int i = 0; i < embed_dim / 2; ++i) {
omega[i] = 1.0 / pow(10000.0, static_cast<float>(i) / (embed_dim / 2));
}
std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
for (int h = 0; h < H; ++h) {
for (int w = 0; w < W; ++w) {
for (int d = 0; d < embed_dim / 2; ++d) {
float out_value = pos[h][w] * omega[d];
emb[h][w][d] = sin(out_value);
emb[h][w][d + embed_dim / 2] = cos(out_value);
}
}
}
return emb;
}
static std::vector<std::vector<std::vector<float>>> get_2d_sincos_pos_embed_from_grid(int embed_dim, const std::vector<std::vector<std::vector<float>>> & grid) {
assert(embed_dim % 2 == 0);
std::vector<std::vector<std::vector<float>>> emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[0]); // (H, W, D/2)
std::vector<std::vector<std::vector<float>>> emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim / 2, grid[1]); // (H, W, D/2)
int H = emb_h.size();
int W = emb_h[0].size();
std::vector<std::vector<std::vector<float>>> emb(H, std::vector<std::vector<float>>(W, std::vector<float>(embed_dim)));
for (int h = 0; h < H; ++h) {
for (int w = 0; w < W; ++w) {
for (int d = 0; d < embed_dim / 2; ++d) {
emb[h][w][d] = emb_h[h][w][d];
emb[h][w][d + embed_dim / 2] = emb_w[h][w][d];
}
}
}
return emb;
}
static std::vector<std::vector<float>> get_2d_sincos_pos_embed(int embed_dim, const std::pair<int, int> image_size) {
int grid_h_size = image_size.first;
int grid_w_size = image_size.second;
std::vector<float> grid_h(grid_h_size);
std::vector<float> grid_w(grid_w_size);
for (int i = 0; i < grid_h_size; ++i) {
grid_h[i] = static_cast<float>(i);
}
for (int i = 0; i < grid_w_size; ++i) {
grid_w[i] = static_cast<float>(i);
}
std::vector<std::vector<float>> grid(grid_h_size, std::vector<float>(grid_w_size));
for (int h = 0; h < grid_h_size; ++h) {
for (int w = 0; w < grid_w_size; ++w) {
grid[h][w] = grid_w[w];
}
}
std::vector<std::vector<std::vector<float>>> grid_2d = {grid, grid};
for (int h = 0; h < grid_h_size; ++h) {
for (int w = 0; w < grid_w_size; ++w) {
grid_2d[0][h][w] = grid_h[h];
grid_2d[1][h][w] = grid_w[w];
}
}
std::vector<std::vector<std::vector<float>>> pos_embed_3d = get_2d_sincos_pos_embed_from_grid(embed_dim, grid_2d);
int H = image_size.first;
int W = image_size.second;
std::vector<std::vector<float>> pos_embed_2d(H * W, std::vector<float>(embed_dim));
for (int h = 0; h < H; ++h) {
for (int w = 0; w < W; ++w) {
pos_embed_2d[w * H + h] = pos_embed_3d[h][w];
}
}
return pos_embed_2d;
}
bool clip_image_encode(struct clip_ctx * ctx, const int n_threads, clip_image_f32 * img, float * vec) {
if (!ctx->has_vision_encoder) {
LOG_TEE("This gguf file seems to have no vision encoder\n");
@ -1843,18 +2260,27 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
if (ctx->has_llava_projector) {
GGML_ASSERT(batch_size == 1); // TODO: support multiple images
}
if (ctx->has_minicpmv_projector) {
GGML_ASSERT(batch_size == 1);
}
// build the inference graph
ggml_cgraph * gf = clip_image_build_graph(ctx, imgs);
ggml_cgraph * gf = clip_image_build_graph(ctx, imgs, ctx->load_image_size, true);
ggml_gallocr_alloc_graph(ctx->compute_alloc, gf);
// set inputs
const auto & model = ctx->vision_model;
const auto & hparams = model.hparams;
const int image_size = hparams.image_size;
const int image_size = hparams.image_size;
int image_size_width = image_size;
int image_size_height = image_size;
if (ctx->has_minicpmv_projector) {
image_size_width = imgs->data[0].nx;
image_size_height = imgs->data[0].ny;
}
const int patch_size = hparams.patch_size;
const int num_patches = ((image_size / patch_size) * (image_size / patch_size));
const int num_patches = ((image_size_width / patch_size) * (image_size_height / patch_size));
const int num_positions = num_patches + (ctx->has_class_embedding ? 1 : 0);
{
@ -1864,7 +2290,9 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
for (size_t i = 0; i < imgs->size; i++) {
const int nx = imgs->data[i].nx;
const int ny = imgs->data[i].ny;
GGML_ASSERT(nx == image_size && ny == image_size);
if (!ctx->has_minicpmv_projector) {
GGML_ASSERT(nx == image_size && ny == image_size);
}
const int n = nx * ny;
@ -1881,37 +2309,75 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
ggml_backend_tensor_set(inp_raw, data, 0, ggml_nbytes(inp_raw));
free(data);
}
{
if (ctx->has_class_embedding) {
struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
void* zero_mem = malloc(ggml_nbytes(embeddings));
memset(zero_mem, 0, ggml_nbytes(embeddings));
ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
free(zero_mem);
if (ctx->has_minicpmv_projector) {
{
// inspired from siglip:
// -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
// -> https://huggingface.co/HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit/blob/d66538faeba44480d0bfaa42145eef26f9423199/modeling_siglip.py#L316
struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
int* positions_data = (int*)malloc(ggml_nbytes(positions));
for (int i = 0; i < num_positions; i++) {
positions_data[i] = std::floor(70.0*i/num_positions);
}
ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
free(positions_data);
}
}
{
struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
{
// inspired from resampler of Qwen-VL:
// -> https://huggingface.co/Qwen/Qwen-VL/tree/main
// -> https://huggingface.co/Qwen/Qwen-VL/blob/0547ed36a86561e2e42fecec8fd0c4f6953e33c4/visual.py#L23
struct ggml_tensor * pos_embed = ggml_graph_get_tensor(gf, "pos_embed");
if(ctx->load_image_size==nullptr){
ctx->load_image_size= clip_image_size_init();
}
int pos_w = ctx->load_image_size->width/patch_size;
int pos_h = ctx->load_image_size->height/patch_size;
int embed_dim = 4096;
auto pos_embed_t = get_2d_sincos_pos_embed(embed_dim, std::make_pair(pos_w, pos_h));
int* positions_data = (int*)malloc(ggml_nbytes(positions));
for (int i = 0; i < num_positions; i++) {
positions_data[i] = i;
float * pos_embed_data = (float *)malloc(ggml_nbytes(pos_embed));
for(int i=0;i<pos_w * pos_h;++i){
for(int j=0;j<embed_dim;++j){
pos_embed_data[i*embed_dim+j]=pos_embed_t[i][j];
}
}
ggml_backend_tensor_set(pos_embed, pos_embed_data, 0, ggml_nbytes(pos_embed));
free(pos_embed_data);
}
ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
free(positions_data);
}
} else {
{
if (ctx->has_class_embedding) {
struct ggml_tensor * embeddings = ggml_graph_get_tensor(gf, "embeddings");
{
struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
int* patches_data = (int*)malloc(ggml_nbytes(patches));
for (int i = 0; i < num_patches; i++) {
patches_data[i] = i + 1;
void* zero_mem = malloc(ggml_nbytes(embeddings));
memset(zero_mem, 0, ggml_nbytes(embeddings));
ggml_backend_tensor_set(embeddings, zero_mem, 0, ggml_nbytes(embeddings));
free(zero_mem);
}
}
{
struct ggml_tensor * positions = ggml_graph_get_tensor(gf, "positions");
int* positions_data = (int*)malloc(ggml_nbytes(positions));
for (int i = 0; i < num_positions; i++) {
positions_data[i] = i;
}
ggml_backend_tensor_set(positions, positions_data, 0, ggml_nbytes(positions));
free(positions_data);
}
{
struct ggml_tensor * patches = ggml_graph_get_tensor(gf, "patches");
int* patches_data = (int*)malloc(ggml_nbytes(patches));
for (int i = 0; i < num_patches; i++) {
patches_data[i] = i + 1;
}
ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
free(patches_data);
}
ggml_backend_tensor_set(patches, patches_data, 0, ggml_nbytes(patches));
free(patches_data);
}
if (ggml_backend_is_cpu(ctx->backend)) {
@ -2081,7 +2547,14 @@ int clip_n_mmproj_embd(const struct clip_ctx * ctx) {
if (ctx->proj_type == PROJECTOR_TYPE_MLP_NORM) {
return ctx->vision_model.mm_3_b->ne[0];
}
if (ctx->proj_type == PROJECTOR_TYPE_RESAMPLER) {
return 4096;
}
std::string proj_type = PROJECTOR_TYPE_NAMES[ctx->proj_type];
throw std::runtime_error(format("%s: don't support projector with: %s currently\n", __func__, proj_type.c_str()));
}
bool clip_is_minicpmv(const struct clip_ctx * ctx) {
return ctx->has_minicpmv_projector;
}

13
examples/llava/clip.h
View File

@ -18,14 +18,17 @@
# define CLIP_API
#endif
struct clip_ctx;
#ifdef __cplusplus
extern "C" {
#endif
struct clip_ctx;
struct clip_image_size {
int width;
int height;
};
struct clip_image_u8_batch {
struct clip_image_u8 * data;
size_t size;
@ -55,6 +58,10 @@ CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
CLIP_API int clip_n_patches (const struct clip_ctx * ctx);
CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);
CLIP_API struct clip_image_size * clip_image_size_init();
CLIP_API struct clip_image_u8 * clip_image_u8_init ();
CLIP_API struct clip_image_f32 * clip_image_f32_init();
@ -78,6 +85,8 @@ CLIP_API bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, cons
CLIP_API bool clip_model_quantize(const char * fname_inp, const char * fname_out, int itype);
CLIP_API bool clip_is_minicpmv(const struct clip_ctx * ctx);
#ifdef __cplusplus
}
#endif

75
examples/llava/llava.cpp
View File

@ -202,6 +202,33 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
return true;
}
static clip_image_f32 * only_v2_5_reshape_by_patch(clip_image_f32 * image, int patch_size) {
int width = image->nx;
int height = image->ny;
int num_patches = (height / patch_size) * (width / patch_size);
clip_image_f32 * patch = clip_image_f32_init();
patch->nx = patch_size * num_patches;
patch->ny = patch_size;
patch->buf.resize(3 * patch->nx * patch->ny);
int patch_index = 0;
for (int i = 0; i < height; i += patch_size) {
for (int j = 0; j < width; j += patch_size) {
for (int pi = 0; pi < patch_size; ++pi) {
for (int pj = 0; pj < patch_size; ++pj) {
int input_index = ((i + pi) * width + (j + pj)) * 3;
int output_index = (pi * patch_size * num_patches + patch_index * patch_size + pj) * 3;
patch->buf[output_index] = image->buf[input_index];
patch->buf[output_index+1] = image->buf[input_index+1];
patch->buf[output_index+2] = image->buf[input_index+2];
}
}
patch_index++;
}
}
return patch;
}
static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const clip_image_u8 * img, float * image_embd, int * n_img_pos) {
// std::vector<clip_image_f32*> img_res_v; // format VectN x H x W x RGB (N x 336 x 336 x 3), so interleaved RGB - different to the python implementation which is N x 3 x 336 x 336
@ -218,7 +245,44 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
if (clip_is_minicpmv(ctx_clip)) {
std::vector<float *> image_embd_v;
image_embd_v.resize(img_res_v.size);
struct clip_image_size * load_image_size = clip_image_size_init();
for (size_t i = 0; i < img_res_v.size; i++) {
const int64_t t_img_enc_step_start_us = ggml_time_us();
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
int patch_size=14;
load_image_size->width = img_res_v.data[i].nx;
load_image_size->height = img_res_v.data[i].ny;
clip_add_load_image_size(ctx_clip, load_image_size);
const bool encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
if (!encoded) {
LOG_TEE("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
return false;
}
const int64_t t_img_enc_steop_batch_us = ggml_time_us();
LOG_TEE("%s: step %d of %d encoded in %8.2f ms\n", __func__, (int)i+1, (int)img_res_v.size, (t_img_enc_steop_batch_us - t_img_enc_step_start_us) / 1000.0);
}
const int64_t t_img_enc_batch_us = ggml_time_us();
LOG_TEE("%s: all %d segments encoded in %8.2f ms\n", __func__, (int)img_res_v.size, (t_img_enc_batch_us - t_img_enc_start_us) / 1000.0);
int n_img_pos_out = 0;
for (size_t i = 0; i < image_embd_v.size(); i++) {
std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
n_img_pos_out += clip_n_patches(ctx_clip);
}
*n_img_pos = n_img_pos_out;
for (size_t i = 0; i < image_embd_v.size(); i++) {
free(image_embd_v[i]);
}
image_embd_v.clear();
load_image_size->width = img->nx;
load_image_size->height = img->ny;
clip_add_load_image_size(ctx_clip, load_image_size);
LOG_TEE("%s: load_image_size %d %d\n", __func__, load_image_size->width, load_image_size->height);
}
else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
// flat / default llava-1.5 type embedding
*n_img_pos = clip_n_patches(ctx_clip);
bool encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[0], image_embd); // image_embd shape is 576 x 4096
@ -228,7 +292,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
return false;
}
} else {
}
else {
// spatial_unpad llava-1.6 type embedding
// TODO: CLIP needs batching support - in HF the llm projection is separate after encoding, which might be a solution to quickly get batching working
std::vector<float *> image_embd_v;
@ -297,7 +362,11 @@ bool llava_validate_embed_size(const llama_context * ctx_llama, const clip_ctx *
}
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
int num_max_patches = 6;
if (clip_is_minicpmv(ctx_clip)) {
num_max_patches = 10;
}
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
if (!image_embd) {
LOG_TEE("Unable to allocate memory for image embeddings\n");
return false;

5
examples/llava/llava.h
View File

@ -17,12 +17,11 @@
# define LLAVA_API
#endif
struct clip_ctx;
#ifdef __cplusplus
extern "C" {
#endif
struct clip_ctx;
struct llava_image_embed {
float * embed;
int n_image_pos;
@ -37,8 +36,8 @@ LLAVA_API bool llava_image_embed_make_with_clip_img(struct clip_ctx * ctx_clip,
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 */
LLAVA_API struct llava_image_embed * llava_image_embed_make_with_filename(struct clip_ctx * ctx_clip, int n_threads, const char * image_path);
LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
/** free an embedding made with llava_image_embed_make_* */
LLAVA_API void llava_image_embed_free(struct llava_image_embed * embed);
/** write the image represented by embed into the llama context with batch size n_batch, starting at context pos n_past. on completion, n_past points to the next position in the context after the image embed. */
LLAVA_API bool llava_eval_image_embed(struct llama_context * ctx_llama, const struct llava_image_embed * embed, int n_batch, int * n_past);

309
examples/llava/minicpmv-cli.cpp Normal file
View File

@ -0,0 +1,309 @@
#include "ggml.h"
#include "log.h"
#include "common.h"
#include "clip.h"
#include "llava.h"
#include "llama.h"
#include <cstdio>
#include <cstdlib>
#include <vector>
struct llava_context {
struct clip_ctx * ctx_clip = NULL;
struct llama_context * ctx_llama = NULL;
struct llama_model * model = NULL;
};
static void show_additional_info(int /*argc*/, char ** argv) {
LOG_TEE("\n example usage: %s -m <llava-v1.5-7b/ggml-model-q5_k.gguf> --mmproj <llava-v1.5-7b/mmproj-model-f16.gguf> --image <path/to/an/image.jpg> --image <path/to/another/image.jpg> [--temp 0.1] [-p \"describe the image in detail.\"]\n", argv[0]);
LOG_TEE(" note: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static void llama_log_callback_logTee(ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
LOG_TEE("%s", text);
}
static struct llama_model * llava_init(gpt_params * params) {
llama_backend_init();
llama_numa_init(params->numa);
llama_model_params model_params = llama_model_params_from_gpt_params(*params);
llama_model * model = llama_load_model_from_file(params->model.c_str(), model_params);
if (model == NULL) {
LOG_TEE("%s: error: unable to load model\n" , __func__);
return NULL;
}
return model;
}
static struct llava_context * llava_init_context(gpt_params * params, llama_model * model) {
auto prompt = params->prompt;
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
llama_context_params ctx_params = llama_context_params_from_gpt_params(*params);
if (params->n_ctx < 2048) {
// warn user here, "Image processing requires at least 2048 context, setting context to 2048"
LOG_TEE("%s: warn: Image processing requires at least 2048 context, setting context to 2048\n" , __func__);
ctx_params.n_ctx = 2048;
} else {
ctx_params.n_ctx = params->n_ctx;
}
llama_context * ctx_llama = llama_new_context_with_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_TEE("%s: error: failed to create the llama_context\n" , __func__);
return NULL;
}
auto ctx_llava = (struct llava_context *)malloc(sizeof(llava_context));
ctx_llava->ctx_llama = ctx_llama;
ctx_llava->model = model;
return ctx_llava;
}
static void llava_free(struct llava_context * ctx_llava) {
if (ctx_llava->ctx_clip) {
clip_free(ctx_llava->ctx_clip);
ctx_llava->ctx_clip = NULL;
}
llama_free(ctx_llava->ctx_llama);
llama_free_model(ctx_llava->model);
llama_backend_free();
}
static struct clip_ctx * clip_init_context(gpt_params * params) {
const char * clip_path = params->mmproj.c_str();
auto prompt = params->prompt;
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
auto ctx_clip = clip_model_load(clip_path, /*verbosity=*/ 1);
return ctx_clip;
}
static bool eval_tokens(struct llama_context * ctx_llama, std::vector<llama_token> tokens, int n_batch, int * n_past) {
int N = (int) tokens.size();
for (int i = 0; i < N; i += n_batch) {
int n_eval = (int) tokens.size() - i;
if (n_eval > n_batch) {
n_eval = n_batch;
}
if (llama_decode(ctx_llama, llama_batch_get_one(&tokens[i], n_eval, *n_past, 0))) {
LOG_TEE("%s : failed to eval. token %d/%d (batch size %d, n_past %d)\n", __func__, i, N, n_batch, *n_past);
return false;
}
*n_past += n_eval;
}
return true;
}
static bool eval_id(struct llama_context * ctx_llama, int id, int * n_past) {
std::vector<llama_token> tokens;
tokens.push_back(id);
return eval_tokens(ctx_llama, tokens, 1, n_past);
}
static bool eval_string(struct llama_context * ctx_llama, const char* str, int n_batch, int * n_past, bool add_bos){
std::string str2 = str;
std::vector<llama_token> embd_inp = ::llama_tokenize(ctx_llama, str2, add_bos, true);
return eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
}
static void process_eval_image_embed(struct llava_context * ctx_llava, const struct llava_image_embed * embeds, int n_batch, int * n_past, int idx) {
float * image_embed = (float *)malloc(clip_embd_nbytes(ctx_llava->ctx_clip));
std::memcpy(image_embed, embeds->embed + idx * clip_n_patches(ctx_llava->ctx_clip) * clip_n_mmproj_embd(ctx_llava->ctx_clip), clip_embd_nbytes(ctx_llava->ctx_clip));
auto slice_embed = (llava_image_embed*)malloc(sizeof(llava_image_embed));
slice_embed->embed = image_embed;
slice_embed->n_image_pos = clip_n_patches(ctx_llava->ctx_clip);
llava_eval_image_embed(ctx_llava->ctx_llama, slice_embed, n_batch, n_past);
llava_image_embed_free(slice_embed);
}
static void process_image(struct llava_context * ctx_llava, struct llava_image_embed * embeds, gpt_params * params, int &n_past) {
std::string system_prompt;
int idx = 0;
int num_image_embeds = embeds->n_image_pos / clip_n_patches(ctx_llava->ctx_clip);
system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
LOG_TEE("%s: image token past: %d\n", __func__, n_past);
eval_string(ctx_llava->ctx_llama, (system_prompt+"<image>").c_str(), params->n_batch, &n_past, false);
process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
if (num_image_embeds > 1) {
size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
eval_string(ctx_llava->ctx_llama, std::string("<slice>").c_str(), params->n_batch, &n_past, false);
for (size_t i = 0; i < (num_image_embeds-1)/num_image_embeds_col; ++i) {
for (size_t j = 0; j < num_image_embeds_col; ++j) {
eval_string(ctx_llava->ctx_llama, std::string("<image>").c_str(), params->n_batch, &n_past, false);
process_eval_image_embed(ctx_llava, embeds, params->n_batch, &n_past, idx++);
eval_string(ctx_llava->ctx_llama, std::string("</image>").c_str(), params->n_batch, &n_past, false);
if (j == num_image_embeds_col - 1) {
eval_string(ctx_llava->ctx_llama, std::string("\n").c_str(), params->n_batch, &n_past, false);
}
}
}
eval_string(ctx_llava->ctx_llama, std::string("</slice>").c_str(), params->n_batch, &n_past, false);
}
LOG_TEE("%s: image token past: %d\n", __func__, n_past);
}
static const char * sample(struct llama_sampling_context * ctx_sampling,
struct llama_context * ctx_llama,
int * n_past) {
const llama_token id = llama_sampling_sample(ctx_sampling, ctx_llama, NULL);
llama_sampling_accept(ctx_sampling, ctx_llama, id, true);
static std::string ret;
if (llama_token_is_eog(llama_get_model(ctx_llama), id)) {
ret = "</s>";
} else {
ret = llama_token_to_piece(ctx_llama, id);
}
eval_id(ctx_llama, id, n_past);
return ret.c_str();
}
static struct llava_context * minicpmv_init(gpt_params * params, const std::string & fname, int &n_past){
auto ctx_clip = clip_init_context(params);
auto embeds = llava_image_embed_make_with_filename(ctx_clip, params->n_threads, fname.c_str());
if (!embeds) {
std::cerr << "error: failed to load image " << fname << ". Terminating\n\n";
return NULL;
}
// process the prompt
if (params->prompt.empty() && params->interactive == false) {
LOG_TEE("prompt should be given or interactive mode should be on");
return NULL;
}
auto model = llava_init(params);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to init minicpmv model\n", __func__);
return NULL;
}
const int64_t t_llava_init_start_us = ggml_time_us();
auto ctx_llava = llava_init_context(params, model);
ctx_llava->ctx_clip = ctx_clip;
const int64_t t_llava_init_end_us = ggml_time_us();
float t_llava_init_ms = (t_llava_init_end_us - t_llava_init_start_us) / 1000.0;
LOG_TEE("\n%s: llava init in %8.2f ms.\n", __func__, t_llava_init_ms);
const int64_t t_process_image_start_us = ggml_time_us();
process_image(ctx_llava, embeds, params, n_past);
const int64_t t_process_image_end_us = ggml_time_us();
float t_process_image_ms = (t_process_image_end_us - t_process_image_start_us) / 1000.0;
LOG_TEE("\n%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
llava_image_embed_free(embeds);
return ctx_llava;
}
static struct llama_sampling_context * llama_init(struct llava_context * ctx_llava, gpt_params * params, std::string prompt, int &n_past, bool is_first = false){
std::string user_prompt = prompt;
if (!is_first) user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
// generate the response
LOG_TEE("\n");
struct llama_sampling_context * ctx_sampling = llama_sampling_init(params->sparams);
return ctx_sampling;
}
static const char * llama_loop(struct llava_context * ctx_llava,struct llama_sampling_context * ctx_sampling, int &n_past){
const char * tmp = sample(ctx_sampling, ctx_llava->ctx_llama, &n_past);
return tmp;
}
int main(int argc, char ** argv) {
ggml_time_init();
gpt_params params;
if (!gpt_params_parse(argc, argv, params)) {
show_additional_info(argc, argv);
return 1;
}
#ifndef LOG_DISABLE_LOGS
log_set_target(log_filename_generator("llava", "log"));
LOG_TEE("Log start\n");
log_dump_cmdline(argc, argv);
llama_log_set(llama_log_callback_logTee, nullptr);
#endif // LOG_DISABLE_LOGS
if (params.mmproj.empty() || (params.image.empty())) {
gpt_params_print_usage(argc, argv, params);
show_additional_info(argc, argv);
return 1;
}
for (auto & image : params.image) {
int n_past = 0;
auto ctx_llava = minicpmv_init(&params, image, n_past);
if (!params.prompt.empty()) {
LOG_TEE("<user>%s\n", params.prompt.c_str());
LOG_TEE("<assistant>");
auto ctx_sampling = llama_init(ctx_llava, &params, params.prompt.c_str(), n_past, true);
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
std::string response = "";
bool have_tmp = false;
for (int i = 0; i < max_tgt_len; i++) {
auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0){
if(!have_tmp)continue;
else break;
}
if (strstr(tmp, "###")) break; // Yi-VL behavior
have_tmp = true;
printf("%s", tmp);
if (strstr(response.c_str(), "<user>")) break; // minicpm-v
fflush(stdout);
}
llama_sampling_free(ctx_sampling);
}else {
while (true) {
LOG_TEE("<user>");
std::string prompt;
std::getline(std::cin, prompt);
LOG_TEE("<assistant>");
auto ctx_sampling = llama_init(ctx_llava, &params, prompt, n_past, true);
const int max_tgt_len = params.n_predict < 0 ? 256 : params.n_predict;
std::string response = "";
for (int i = 0; i < max_tgt_len; i++) {
auto tmp = llama_loop(ctx_llava, ctx_sampling, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
printf("%s", tmp);// mistral llava-1.6
if (strstr(response.c_str(), "<user>")) break; // minicpm-v
fflush(stdout);
}
llama_sampling_free(ctx_sampling);
}
}
printf("\n");
llama_print_timings(ctx_llava->ctx_llama);
ctx_llava->model = NULL;
llava_free(ctx_llava);
}
return 0;
}

382
examples/llava/minicpmv-convert-image-encoder-to-gguf.py Normal file
View File

@ -0,0 +1,382 @@
import argparse
import os
import json
import re
import torch
import numpy as np
from gguf import *
from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer, Idefics2VisionConfig
TEXT = "clip.text"
VISION = "clip.vision"
def add_key_str(raw_key: str, arch: str) -> str:
return raw_key.format(arch=arch)
def should_skip_tensor(name: str, has_text: bool, has_vision: bool, has_minicpmv: bool) -> bool:
if name in (
"logit_scale",
"text_model.embeddings.position_ids",
"vision_model.embeddings.position_ids",
):
return True
if has_minicpmv and name in ["visual_projection.weight"]:
return True
if name.startswith("v") and not has_vision:
return True
if name.startswith("t") and not has_text:
return True
return False
def get_tensor_name(name: str) -> str:
if "projection" in name:
return name
if "mm_projector" in name:
name = name.replace("model.mm_projector", "mm")
name = re.sub(r'mm\.mlp\.mlp', 'mm.model.mlp', name, count=1)
name = re.sub(r'mm\.peg\.peg', 'mm.model.peg', name, count=1)
return name
return name.replace("text_model", "t").replace("vision_model", "v").replace("encoder.layers", "blk").replace("embeddings.", "").replace("_proj", "").replace("self_attn.", "attn_").replace("layer_norm", "ln").replace("layernorm", "ln").replace("mlp.fc1", "ffn_down").replace("mlp.fc2", "ffn_up").replace("embedding", "embd").replace("final", "post").replace("layrnorm", "ln")
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 significant 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))
ap = argparse.ArgumentParser()
ap.add_argument("-m", "--model-dir", help="Path to model directory cloned from HF Hub", required=True)
ap.add_argument("--use-f32", action="store_true", default=False, help="Use f32 instead of f16")
ap.add_argument("--text-only", action="store_true", required=False,
help="Save a text-only model. It can't be used to encode images")
ap.add_argument("--vision-only", action="store_true", required=False,
help="Save a vision-only model. It can't be used to encode texts")
ap.add_argument("--clip-model-is-vision", action="store_true", required=False,
help="The clip model is a pure vision model (ShareGPT4V vision extract for example)")
ap.add_argument("--clip-model-is-openclip", action="store_true", required=False,
help="The clip model is from openclip (for ViT-SO400M type))")
ap.add_argument("--minicpmv-projector", help="Path to minicpmv.projector file. If specified, save an image encoder for MiniCPM-V models.")
ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2"], default="mlp")
ap.add_argument("-o", "--output-dir", help="Directory to save GGUF files. Default is the original model directory", default=None)
# Example --image_mean 0.48145466 0.4578275 0.40821073 --image_std 0.26862954 0.26130258 0.27577711
# Example --image_mean 0.5 0.5 0.5 --image_std 0.5 0.5 0.5
default_image_mean = [0.48145466, 0.4578275, 0.40821073]
default_image_std = [0.26862954, 0.26130258, 0.27577711]
ap.add_argument('--image-mean', type=float, nargs='+', help='Mean of the images for normalization (overrides processor) ', default=None)
ap.add_argument('--image-std', type=float, nargs='+', help='Standard deviation of the images for normalization (overrides processor)', default=None)
# with proper
args = ap.parse_args()
if args.text_only and args.vision_only:
print("--text-only and --image-only arguments cannot be specified at the same time.")
exit(1)
if args.use_f32:
print("WARNING: Weights for the convolution op is always saved in f16, as the convolution op in GGML does not support 32-bit kernel weights yet.")
# output in the same directory as the model if output_dir is None
dir_model = args.model_dir
if args.clip_model_is_vision or not os.path.exists(dir_model + "/vocab.json") or args.clip_model_is_openclip:
vocab = None
tokens = None
else:
with open(dir_model + "/vocab.json", "r", encoding="utf-8") as f:
vocab = json.load(f)
tokens = [key for key in vocab]
# possible data types
# ftype == 0 -> float32
# ftype == 1 -> float16
#
# map from ftype to string
ftype_str = ["f32", "f16"]
ftype = 1
if args.use_f32:
ftype = 0
# if args.clip_model_is_vision or args.clip_model_is_openclip:
# model = CLIPVisionModel.from_pretrained(dir_model)
# processor = None
# else:
# model = CLIPModel.from_pretrained(dir_model)
# processor = CLIPProcessor.from_pretrained(dir_model)
default_vision_config = {
"hidden_size": 1152,
"image_size": 980,
"intermediate_size": 4304,
"model_type": "idefics2",
"num_attention_heads": 16,
"num_hidden_layers": 27,
"patch_size": 14,
}
vision_config = Idefics2VisionConfig(**default_vision_config)
model = Idefics2VisionTransformer(vision_config)
processor = None
# if model.attn_pool is not None:
# model.attn_pool = torch.nn.Identity()
# model.blocks = model.blocks[:-1]
model.load_state_dict(torch.load(os.path.join(dir_model, "minicpmv.clip")))
fname_middle = None
has_text_encoder = True
has_vision_encoder = True
has_minicpmv_projector = False
if args.text_only:
fname_middle = "text-"
has_vision_encoder = False
elif args.minicpmv_projector is not None:
fname_middle = "mmproj-"
has_text_encoder = False
has_minicpmv_projector = True
elif args.vision_only:
fname_middle = "vision-"
has_text_encoder = False
else:
fname_middle = ""
output_dir = args.output_dir if args.output_dir is not None else dir_model
os.makedirs(output_dir, exist_ok=True)
output_prefix = os.path.basename(output_dir).replace("ggml_", "")
fname_out = os.path.join(output_dir, f"{fname_middle}model-{ftype_str[ftype]}.gguf")
fout = GGUFWriter(path=fname_out, arch="clip")
fout.add_bool("clip.has_text_encoder", has_text_encoder)
fout.add_bool("clip.has_vision_encoder", has_vision_encoder)
fout.add_bool("clip.has_minicpmv_projector", has_minicpmv_projector)
fout.add_file_type(ftype)
if args.text_only:
fout.add_description("text-only CLIP model")
elif args.vision_only and not has_minicpmv_projector:
fout.add_description("vision-only CLIP model")
elif has_minicpmv_projector:
fout.add_description("image encoder for MiniCPM-V")
# add projector type
fout.add_string("clip.projector_type", "resampler")
else:
fout.add_description("two-tower CLIP model")
if has_vision_encoder:
# vision_model hparams
fout.add_uint32("clip.vision.image_size", 448)
fout.add_uint32("clip.vision.patch_size", 14)
fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), 1152)
fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), 4304)
fout.add_uint32("clip.vision.projection_dim", 0)
fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), 16)
fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
block_count = 26
fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
if processor is not None:
image_mean = processor.image_processor.image_mean if args.image_mean is None or args.image_mean == default_image_mean else args.image_mean
image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std
else:
image_mean = args.image_mean if args.image_mean is not None else default_image_mean
image_std = args.image_std if args.image_std is not None else default_image_std
fout.add_array("clip.vision.image_mean", image_mean)
fout.add_array("clip.vision.image_std", image_std)
use_gelu = True
fout.add_bool("clip.use_gelu", use_gelu)
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
"""
embed_dim: output dimension for each position
pos: a list of positions to be encoded: size (M,)
out: (M, D)
"""
assert embed_dim % 2 == 0
omega = np.arange(embed_dim // 2, dtype=np.float32)
omega /= embed_dim / 2.
omega = 1. / 10000 ** omega # (D/2,)
pos = pos.reshape(-1) # (M,)
out = np.einsum('m,d->md', pos, omega) # (M, D/2), outer product
emb_sin = np.sin(out) # (M, D/2)
emb_cos = np.cos(out) # (M, D/2)
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
return emb
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
assert embed_dim % 2 == 0
# use half of dimensions to encode grid_h
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
return emb
# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False):
"""
grid_size: int of the grid height and width
return:
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
"""
if isinstance(grid_size, int):
grid_h_size, grid_w_size = grid_size, grid_size
else:
grid_h_size, grid_w_size = grid_size[0], grid_size[1]
grid_h = np.arange(grid_h_size, dtype=np.float32)
grid_w = np.arange(grid_w_size, dtype=np.float32)
grid = np.meshgrid(grid_w, grid_h) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if cls_token:
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
return pos_embed
def _replace_name_resampler(s, v):
if re.match("resampler.pos_embed", s):
return {
s: v,
re.sub("pos_embed", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(4096, (70, 70))),
}
if re.match("resampler.proj", s):
return {
re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(4096, (70, 70))),
re.sub("proj", "proj.weight", s): v.transpose(-1, -2).contiguous(),
}
if re.match("resampler.attn.in_proj_.*", s):
return {
re.sub("attn.in_proj_", "attn.q.", s): v.chunk(3, dim=0)[0],
re.sub("attn.in_proj_", "attn.k.", s): v.chunk(3, dim=0)[1],
re.sub("attn.in_proj_", "attn.v.", s): v.chunk(3, dim=0)[2],
}
return {s: v}
if has_minicpmv_projector:
projector = torch.load(args.minicpmv_projector)
new_state_dict = {}
for k, v in projector.items():
kvs = _replace_name_resampler(k, v)
for nk, nv in kvs.items():
new_state_dict[nk] = nv
projector = new_state_dict
ftype_cur = 0
for name, data in projector.items():
name = get_tensor_name(name)
data = data.squeeze().numpy()
n_dims = len(data.shape)
if ftype == 1:
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
fout.add_tensor(name, data)
print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
print("Projector tensors added\n")
def _replace_name(s, v):
s = "vision_model." + s
if re.match("vision_model.embeddings.position_embedding", s):
v = v.unsqueeze(0)
return {s: v}
return {s: v}
state_dict = model.state_dict()
new_state_dict = {}
for k, v in state_dict.items():
kvs = _replace_name(k, v)
for nk, nv in kvs.items():
new_state_dict[nk] = nv
state_dict = new_state_dict
for name, data in state_dict.items():
if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_minicpmv_projector):
# we don't need this
print(f"skipping parameter: {name}")
continue
name = get_tensor_name(name)
data = data.squeeze().numpy()
n_dims = len(data.shape)
# ftype == 0 -> float32, ftype == 1 -> float16
ftype_cur = 0
if n_dims == 4:
print(f"tensor {name} is always saved in f16")
data = data.astype(np.float16)
ftype_cur = 1
elif ftype == 1:
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
print(f"{name} - {ftype_str[ftype_cur]} - shape = {data.shape}")
fout.add_tensor(name, data)
fout.write_header_to_file()
fout.write_kv_data_to_file()
fout.write_tensors_to_file()
fout.close()
print("Done. Output file: " + fname_out)

47
examples/llava/minicpmv-surgery.py Normal file
View File

@ -0,0 +1,47 @@
import argparse
import os
import torch
from transformers import AutoModel, AutoTokenizer
ap = argparse.ArgumentParser()
ap.add_argument("-m", "--model", help="Path to MiniCPM-V-2.5 model")
args = ap.parse_args()
# find the model part that includes the the multimodal projector weights
model = AutoModel.from_pretrained(args.model, trust_remote_code=True, local_files_only=True)
checkpoint = model.state_dict()
# get a list of mm tensor names
mm_tensors = [k for k, v in checkpoint.items() if k.startswith("resampler")]
# store these tensors in a new dictionary and torch.save them
projector = {name: checkpoint[name].float() for name in mm_tensors}
torch.save(projector, f"{args.model}/minicpmv.projector")
clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vpm")]
if len(clip_tensors) > 0:
clip = {name.replace("vpm.", ""): checkpoint[name].float() for name in clip_tensors}
torch.save(clip, f"{args.model}/minicpmv.clip")
# added tokens should be removed to be able to convert Mistral models
if os.path.exists(f"{args.model}/added_tokens.json"):
with open(f"{args.model}/added_tokens.json", "w") as f:
f.write("{}\n")
config = model.llm.config
config._name_or_path = "openbmb/MiniCPM-Llama3-V-2.5"
config.auto_map = {
"AutoConfig": "configuration_minicpm.MiniCPMConfig",
"AutoModel": "modeling_minicpm.MiniCPMModel",
"AutoModelForCausalLM": "modeling_minicpm.MiniCPMForCausalLM",
"AutoModelForSeq2SeqLM": "modeling_minicpm.MiniCPMForCausalLM",
"AutoModelForSequenceClassification": "modeling_minicpm.MiniCPMForSequenceClassification"
}
model.llm.save_pretrained(f"{args.model}/model")
tok = AutoTokenizer.from_pretrained(args.model, trust_remote_code=True)
tok.save_pretrained(f"{args.model}/model")
# os.system(f"cp {args.model}/modeling_minicpm.py {args.model}/MiniCPM_l3/modeling_minicpm.py")
print("Done!")
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
print(f"Also, use {args.model}/minicpmv.projector to prepare a minicpmv-encoder.gguf file.")

1
examples/llava/requirements.txt
View File

@ -2,3 +2,4 @@
--extra-index-url https://download.pytorch.org/whl/cpu
pillow~=10.2.0
torch~=2.2.1
torchvision==0.17.1

8
examples/lookahead/lookahead.cpp
View File

@ -58,11 +58,11 @@ int main(int argc, char ** argv) {
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;
// load the target model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
// Tokenize the prompt
std::vector<llama_token> inp;

8
examples/lookup/lookup-create.cpp
View File

@ -22,11 +22,11 @@ int main(int argc, char ** argv){
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
GGML_ASSERT(model != nullptr);
// tokenize the prompt

8
examples/lookup/lookup-stats.cpp
View File

@ -26,11 +26,11 @@ int main(int argc, char ** argv){
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
// tokenize the prompt
std::vector<llama_token> inp;

8
examples/lookup/lookup.cpp
View File

@ -34,11 +34,11 @@ int main(int argc, char ** argv){
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
// tokenize the prompt
std::vector<llama_token> inp;

5
examples/main/main.cpp
View File

@ -207,7 +207,10 @@ int main(int argc, char ** argv) {
// load the model and apply lora adapter, if any
LOG("%s: load the model and apply lora adapter, if any\n", __func__);
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
model = llama_init.model;
ctx = llama_init.context;
if (sparams.cfg_scale > 1.f) {
struct llama_context_params lparams = llama_context_params_from_gpt_params(params);
ctx_guidance = llama_new_context_with_model(model, lparams);

8
examples/parallel/parallel.cpp
View File

@ -129,11 +129,11 @@ int main(int argc, char ** argv) {
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model = NULL;
llama_context * ctx = NULL;
// load the target model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
// load the prompts from an external file if there are any
if (params.prompt.empty()) {

8
examples/perplexity/perplexity.cpp
View File

@ -2018,11 +2018,11 @@ int main(int argc, char ** argv) {
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;
// load the model and apply lora adapter, if any
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
return 1;

2
examples/quantize/quantize.cpp
View File

@ -93,7 +93,7 @@ static bool try_parse_ftype(const std::string & ftype_str_in, llama_ftype & ftyp
}
// usage:
// ./quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
// ./llama-quantize [--allow-requantize] [--leave-output-tensor] [--pure] models/llama/ggml-model.gguf [models/llama/ggml-model-quant.gguf] type [nthreads]
//
[[noreturn]]
static void usage(const char * executable) {

9
examples/retrieval/retrieval.cpp
View File

@ -148,11 +148,12 @@ int main(int argc, char ** argv) {
llama_backend_init();
llama_numa_init(params.numa);
llama_model * model;
llama_context * ctx;
// load the model
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
if (model == NULL) {
fprintf(stderr, "%s: error: unable to load model\n", __func__);
return 1;

4
examples/rpc/README.md
View File

@ -1,5 +1,9 @@
## Overview
> [!IMPORTANT]
> This example and the RPC backend are currently in a proof-of-concept development stage. As such, the functionality is fragile and
> insecure. **Never run the RPC server on an open network or in a sensitive environment!**
The `rpc-server` allows running `ggml` backend on a remote host.
The RPC backend communicates with one or several instances of `rpc-server` and offloads computations to them.
This can be used for distributed LLM inference with `llama.cpp` in the following way:

13
examples/rpc/rpc-server.cpp
View File

@ -16,7 +16,7 @@
#include <stdio.h>
struct rpc_server_params {
std::string host = "0.0.0.0";
std::string host = "127.0.0.1";
int port = 50052;
size_t backend_mem = 0;
};
@ -114,6 +114,17 @@ int main(int argc, char * argv[]) {
fprintf(stderr, "Invalid parameters\n");
return 1;
}
if (params.host != "127.0.0.1") {
fprintf(stderr, "\n");
fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
fprintf(stderr, "WARNING: Host ('%s') is != '127.0.0.1'\n", params.host.c_str());
fprintf(stderr, " Never expose the RPC server to an open network!\n");
fprintf(stderr, " This is an experimental feature and is not secure!\n");
fprintf(stderr, "!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!\n");
fprintf(stderr, "\n");
}
ggml_backend_t backend = create_backend();
if (!backend) {
fprintf(stderr, "Failed to create backend\n");

7
examples/save-load-state/save-load-state.cpp
View File

@ -28,10 +28,11 @@ int main(int argc, char ** argv) {
std::string result2;
// init
llama_model * model;
llama_context * ctx;
llama_init_result llama_init = llama_init_from_gpt_params(params);
llama_model * model = llama_init.model;
llama_context * ctx = llama_init.context;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
if (model == nullptr || ctx == nullptr) {
fprintf(stderr, "%s : failed to init\n", __func__);
return 1;

130
examples/server/README.md
View File

@ -207,47 +207,12 @@ model:
-hff, --hf-file FILE Hugging Face model file (default: unused)
-hft, --hf-token TOKEN Hugging Face access token (default: value from HF_TOKEN environment variable)
retrieval:
--context-file FNAME file to load context from (repeat to specify multiple files)
--chunk-size N minimum length of embedded text chunks (default: 64)
--chunk-separator STRING
separator between chunks (default: '
')
passkey:
--junk N number of times to repeat the junk text (default: 250)
--pos N position of the passkey in the junk text (default: -1)
imatrix:
-o, --output FNAME output file (default: 'imatrix.dat')
--output-frequency N output the imatrix every N iterations (default: 10)
--save-frequency N save an imatrix copy every N iterations (default: 0)
--process-output collect data for the output tensor (default: false)
--no-ppl do not compute perplexity (default: true)
--chunk N start processing the input from chunk N (default: 0)
bench:
-pps is the prompt shared across parallel sequences (default: false)
-npp n0,n1,... number of prompt tokens
-ntg n0,n1,... number of text generation tokens
-npl n0,n1,... number of parallel prompts
embedding:
--embd-normalize normalisation for embendings (default: 2) (-1=none, 0=max absolute int16, 1=taxicab, 2=euclidean, >2=p-norm)
--embd-output-format empty = default, "array" = [[],[]...], "json" = openai style, "json+" = same "json" + cosine similarity matrix
--embd-separator separator of embendings (default \n) for example "<#sep#>"
server:
--host HOST ip address to listen (default: 127.0.0.1)
--port PORT port to listen (default: 8080)
--path PATH path to serve static files from (default: )
--embedding(s) enable embedding endpoint (default: disabled)
--embedding(s) restrict to only support embedding use case; use only with dedicated embedding models (default: disabled)
--api-key KEY API key to use for authentication (default: none)
--api-key-file FNAME path to file containing API keys (default: none)
--ssl-key-file FNAME path to file a PEM-encoded SSL private key
@ -267,7 +232,8 @@ server:
https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template
-sps, --slot-prompt-similarity SIMILARITY
how much the prompt of a request must match the prompt of a slot in order to use that slot (default: 0.50, 0.0 = disabled)
--lora-init-without-apply
load LoRA adapters without applying them (apply later via POST /lora-adapters) (default: disabled)
logging:
@ -279,15 +245,6 @@ logging:
--log-file FNAME Specify a log filename (without extension)
--log-new Create a separate new log file on start. Each log file will have unique name: "<name>.<ID>.log"
--log-append Don't truncate the old log file.
cvector:
-o, --output FNAME output file (default: 'control_vector.gguf')
--positive-file FNAME positive prompts file, one prompt per line (default: 'examples/cvector-generator/positive.txt')
--negative-file FNAME negative prompts file, one prompt per line (default: 'examples/cvector-generator/negative.txt')
--pca-batch N batch size used for PCA. Larger batch runs faster, but uses more memory (default: 100)
--pca-iter N number of iterations used for PCA (default: 1000)
--method {pca,mean} dimensionality reduction method to be used (default: pca)
```
@ -411,7 +368,8 @@ node index.js
## API Endpoints
- **GET** `/health`: Returns the current state of the server:
### GET `/health`: Returns the current state of the server
- 503 -> `{"status": "loading model"}` if the model is still being loaded.
- 500 -> `{"status": "error"}` if the model failed to load.
- 200 -> `{"status": "ok", "slots_idle": 1, "slots_processing": 2 }` if the model is successfully loaded and the server is ready for further requests mentioned below.
@ -420,7 +378,7 @@ node index.js
If the query parameter `include_slots` is passed, `slots` field will contain internal slots data except if `--slots-endpoint-disable` is set.
- **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
### POST `/completion`: Given a `prompt`, it returns the predicted completion.
*Options:*
@ -498,7 +456,7 @@ node index.js
`samplers`: The order the samplers should be applied in. An array of strings representing sampler type names. If a sampler is not set, it will not be used. If a sampler is specified more than once, it will be applied multiple times. Default: `["top_k", "tfs_z", "typical_p", "top_p", "min_p", "temperature"]` - these are all the available values.
### Result JSON
**Response format**
- Note: When using streaming mode (`stream`), only `content` and `stop` will be returned until end of completion.
@ -537,7 +495,7 @@ Notice that each `probs` is an array of length `n_probs`.
- `tokens_evaluated`: Number of tokens evaluated in total from the prompt
- `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
- **POST** `/tokenize`: Tokenize a given text.
### POST `/tokenize`: Tokenize a given text
*Options:*
@ -545,13 +503,15 @@ Notice that each `probs` is an array of length `n_probs`.
`add_special`: Boolean indicating if special tokens, i.e. `BOS`, should be inserted. Default: `false`
- **POST** `/detokenize`: Convert tokens to text.
### POST `/detokenize`: Convert tokens to text
*Options:*
`tokens`: Set the tokens to detokenize.
- **POST** `/embedding`: Generate embedding of a given text just as [the embedding example](../embedding) does.
### POST `/embedding`: Generate embedding of a given text
The same as [the embedding example](../embedding) does.
*Options:*
@ -559,7 +519,9 @@ Notice that each `probs` is an array of length `n_probs`.
`image_data`: An array of objects to hold base64-encoded image `data` and its `id`s to be reference in `content`. You can determine the place of the image in the content as in the following: `Image: [img-21].\nCaption: This is a picture of a house`. In this case, `[img-21]` will be replaced by the embeddings of the image with id `21` in the following `image_data` array: `{..., "image_data": [{"data": "<BASE64_STRING>", "id": 21}]}`. Use `image_data` only with multimodal models, e.g., LLaVA.
- **POST** `/infill`: For code infilling. Takes a prefix and a suffix and returns the predicted completion as stream.
### POST `/infill`: For code infilling.
Takes a prefix and a suffix and returns the predicted completion as stream.
*Options:*
@ -571,7 +533,7 @@ Notice that each `probs` is an array of length `n_probs`.
- **GET** `/props`: Return current server settings.
### Result JSON
**Response format**
```json
{
@ -589,7 +551,9 @@ Notice that each `probs` is an array of length `n_probs`.
- `total_slots` - the total number of slots for process requests (defined by `--parallel` option)
- `chat_template` - the model's original Jinja2 prompt template
- **POST** `/v1/chat/completions`: OpenAI-compatible Chat Completions API. Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
### POST `/v1/chat/completions`: OpenAI-compatible Chat Completions API
Given a ChatML-formatted json description in `messages`, it returns the predicted completion. Both synchronous and streaming mode are supported, so scripted and interactive applications work fine. While no strong claims of compatibility with OpenAI API spec is being made, in our experience it suffices to support many apps. Only models with a [supported chat template](https://github.com/ggerganov/llama.cpp/wiki/Templates-supported-by-llama_chat_apply_template) can be used optimally with this endpoint. By default, the ChatML template will be used.
*Options:*
@ -641,7 +605,7 @@ Notice that each `probs` is an array of length `n_probs`.
}'
```
- **POST** `/v1/embeddings`: OpenAI-compatible embeddings API.
### POST `/v1/embeddings`: OpenAI-compatible embeddings API
*Options:*
@ -675,9 +639,9 @@ Notice that each `probs` is an array of length `n_probs`.
}'
```
- **GET** `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
### GET `/slots`: Returns the current slots processing state. Can be disabled with `--slots-endpoint-disable`.
### Result JSON
**Response format**
```json
[
@ -738,7 +702,7 @@ Notice that each `probs` is an array of length `n_probs`.
]
```
- **GET** `/metrics`: [Prometheus](https://prometheus.io/) compatible metrics exporter endpoint if `--metrics` is enabled:
### GET `/metrics`: Prometheus compatible metrics exporter endpoint if `--metrics` is enabled:
Available metrics:
- `llamacpp:prompt_tokens_total`: Number of prompt tokens processed.
@ -750,13 +714,13 @@ Available metrics:
- `llamacpp:requests_processing`: Number of requests processing.
- `llamacpp:requests_deferred`: Number of requests deferred.
- **POST** `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
### POST `/slots/{id_slot}?action=save`: Save the prompt cache of the specified slot to a file.
*Options:*
`filename`: Name of the file to save the slot's prompt cache. The file will be saved in the directory specified by the `--slot-save-path` server parameter.
### Result JSON
**Response format**
```json
{
@ -770,13 +734,13 @@ Available metrics:
}
```
- **POST** `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
### POST `/slots/{id_slot}?action=restore`: Restore the prompt cache of the specified slot from a file.
*Options:*
`filename`: Name of the file to restore the slot's prompt cache from. The file should be located in the directory specified by the `--slot-save-path` server parameter.
### Result JSON
**Response format**
```json
{
@ -790,9 +754,9 @@ Available metrics:
}
```
- **POST** `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
### POST `/slots/{id_slot}?action=erase`: Erase the prompt cache of the specified slot.
### Result JSON
**Response format**
```json
{
@ -801,6 +765,42 @@ Available metrics:
}
```
### GET `/lora-adapters`: Get list of all LoRA adapters
If an adapter is disabled, the scale will be set to 0.
**Response format**
```json
[
{
"id": 0,
"path": "my_adapter_1.gguf",
"scale": 0.0
},
{
"id": 1,
"path": "my_adapter_2.gguf",
"scale": 0.0
}
]
```
### POST `/lora-adapters`: Set list of LoRA adapters
To disable an adapter, either remove it from the list below, or set scale to 0.
**Request format**
To know the `id` of the adapter, use GET `/lora-adapters`
```json
[
{"id": 0, "scale": 0.2},
{"id": 1, "scale": 0.8}
]
```
## More examples
### Change system prompt on runtime

75
examples/server/server.cpp
View File

@ -78,6 +78,7 @@ enum server_task_type {
SERVER_TASK_TYPE_SLOT_SAVE,
SERVER_TASK_TYPE_SLOT_RESTORE,
SERVER_TASK_TYPE_SLOT_ERASE,
SERVER_TASK_TYPE_SET_LORA,
};
struct server_task {
@ -622,6 +623,7 @@ struct server_response {
struct server_context {
llama_model * model = nullptr;
llama_context * ctx = nullptr;
std::vector<llama_lora_adapter_container> lora_adapters;
gpt_params params;
@ -677,7 +679,11 @@ struct server_context {
// dedicate one sequence to the system prompt
params.n_parallel += 1;
std::tie(model, ctx) = llama_init_from_gpt_params(params);
llama_init_result llama_init = llama_init_from_gpt_params(params);
model = llama_init.model;
ctx = llama_init.context;
lora_adapters = llama_init.lora_adapters;
params.n_parallel -= 1; // but be sneaky about it
if (model == nullptr) {
LOG_ERROR("unable to load model", {{"model", params.model}});
@ -900,7 +906,7 @@ struct server_context {
slot.params.stream = json_value(data, "stream", false);
slot.params.cache_prompt = json_value(data, "cache_prompt", false);
slot.params.n_predict = json_value(data, "n_predict", default_params.n_predict);
slot.params.n_predict = json_value(data, "n_predict", json_value(data, "max_tokens", default_params.n_predict));
slot.sparams.top_k = json_value(data, "top_k", default_sparams.top_k);
slot.sparams.top_p = json_value(data, "top_p", default_sparams.top_p);
slot.sparams.min_p = json_value(data, "min_p", default_sparams.min_p);
@ -969,6 +975,8 @@ struct server_context {
(prompt->is_array() && prompt->size() == 1 && prompt->at(0).is_string()) ||
(prompt->is_array() && !prompt->empty() && prompt->at(0).is_number_integer())) {
slot.prompt = *prompt;
} else if (prompt->is_array() && prompt->size() == 1 && prompt->at(0).is_array()) {
slot.prompt = prompt->at(0);
} else {
send_error(task, "\"prompt\" must be a string or an array of integers", ERROR_TYPE_INVALID_REQUEST);
return false;
@ -1847,6 +1855,14 @@ struct server_context {
};
queue_results.send(result);
} break;
case SERVER_TASK_TYPE_SET_LORA:
{
llama_lora_adapters_apply(ctx, lora_adapters);
server_task_result result;
result.id = task.id;
result.data = json{{ "success", true }};
queue_results.send(result);
} break;
}
}
@ -3325,6 +3341,55 @@ int main(int argc, char ** argv) {
return res.set_content(root.dump(), "application/json; charset=utf-8");
};
const auto handle_lora_adapters_list = [&](const httplib::Request & req, httplib::Response & res) {
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
json result = json::array();
for (size_t i = 0; i < ctx_server.lora_adapters.size(); ++i) {
auto & la = ctx_server.lora_adapters[i];
result.push_back({
{"id", i},
{"path", la.path},
{"scale", la.scale},
});
}
res.set_content(result.dump(), "application/json");
res.status = 200; // HTTP OK
};
const auto handle_lora_adapters_apply = [&](const httplib::Request & req, httplib::Response & res) {
res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
const std::vector<json> body = json::parse(req.body);
int max_idx = ctx_server.lora_adapters.size();
// clear existing value
for (auto & la : ctx_server.lora_adapters) {
la.scale = 0.0f;
}
// set value
for (auto entry : body) {
int id = entry.at("id");
float scale = entry.at("scale");
if (0 <= id && id < max_idx) {
ctx_server.lora_adapters[id].scale = scale;
} else {
throw std::runtime_error("invalid adapter id");
}
}
server_task task;
task.type = SERVER_TASK_TYPE_SET_LORA;
const int id_task = ctx_server.queue_tasks.post(task);
ctx_server.queue_results.add_waiting_task_id(id_task);
server_task_result result = ctx_server.queue_results.recv(id_task);
ctx_server.queue_results.remove_waiting_task_id(id_task);
res.set_content(result.data.dump(), "application/json");
res.status = 200; // HTTP OK
};
auto handle_static_file = [](unsigned char * content, size_t len, const char * mime_type) {
return [content, len, mime_type](const httplib::Request &, httplib::Response & res) {
res.set_content(reinterpret_cast<const char*>(content), len, mime_type);
@ -3363,7 +3428,6 @@ int main(int argc, char ** argv) {
// register API routes
svr->Get ("/health", handle_health);
svr->Get ("/slots", handle_slots);
svr->Get ("/metrics", handle_metrics);
svr->Get ("/props", handle_props);
svr->Get ("/v1/models", handle_models);
@ -3378,6 +3442,11 @@ int main(int argc, char ** argv) {
svr->Post("/v1/embeddings", handle_embeddings);
svr->Post("/tokenize", handle_tokenize);
svr->Post("/detokenize", handle_detokenize);
// LoRA adapters hotswap
svr->Get ("/lora-adapters", handle_lora_adapters_list);
svr->Post("/lora-adapters", handle_lora_adapters_apply);
// Save & load slots
svr->Get ("/slots", handle_slots);
if (!params.slot_save_path.empty()) {
// only enable slot endpoints if slot_save_path is set
svr->Post("/slots/:id_slot", handle_slots_action);

36
examples/server/tests/features/lora.feature Normal file
View File

@ -0,0 +1,36 @@
@llama.cpp
@lora
Feature: llama.cpp server
Background: Server startup
Given a server listening on localhost:8080
And a model url https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/stories15M_MOE-F16.gguf
And a model file stories15M_MOE-F16.gguf
And a model alias stories15M_MOE
And a lora adapter file from https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf
And 42 as server seed
And 1024 as batch size
And 1024 as ubatch size
And 2048 KV cache size
And 64 max tokens to predict
And 0.0 temperature
Then the server is starting
Then the server is healthy
Scenario: Completion LoRA disabled
Given switch off lora adapter 0
Given a prompt:
"""
Look in thy glass
"""
And a completion request with no api error
Then 64 tokens are predicted matching little|girl|three|years|old
Scenario: Completion LoRA enabled
Given switch on lora adapter 0
Given a prompt:
"""
Look in thy glass
"""
And a completion request with no api error
Then 64 tokens are predicted matching eye|love|glass|sun

21
examples/server/tests/features/steps/steps.py
View File

@ -7,6 +7,7 @@ import subprocess
import sys
import threading
import time
import requests
from collections.abc import Sequence
from contextlib import closing
from re import RegexFlag
@ -70,6 +71,7 @@ def step_server_config(context, server_fqdn: str, server_port: str):
context.user_api_key = None
context.response_format = None
context.temperature = None
context.lora_file = None
context.tasks_result = []
context.concurrent_tasks = []
@ -82,6 +84,12 @@ def step_download_hf_model(context, hf_file: str, hf_repo: str):
context.model_hf_file = hf_file
context.model_file = os.path.basename(hf_file)
@step('a lora adapter file from {lora_file_url}')
def step_download_lora_file(context, lora_file_url: str):
file_name = lora_file_url.split('/').pop()
context.lora_file = f'../../../{file_name}'
with open(context.lora_file, 'wb') as f:
f.write(requests.get(lora_file_url).content)
@step('a model file {model_file}')
def step_model_file(context, model_file: str):
@ -849,6 +857,17 @@ async def step_erase_slot(context, slot_id):
context.response = response
@step('switch {on_or_off} lora adapter {lora_id:d}')
@async_run_until_complete
async def toggle_lora_adapter(context, on_or_off: str, lora_id: int):
async with aiohttp.ClientSession() as session:
async with session.post(f'{context.base_url}/lora-adapters',
json=[{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}],
headers={"Content-Type": "application/json"}) as response:
context.response = response
print([{'id': lora_id, 'scale': 1 if on_or_off == 'on' else 0}])
@step('the server responds with status code {status_code:d}')
def step_server_responds_with_status_code(context, status_code):
assert context.response.status == status_code
@ -1326,6 +1345,8 @@ def start_server_background(context):
server_args.extend(['--grp-attn-w', context.n_ga_w])
if context.debug:
server_args.append('--verbose')
if context.lora_file:
server_args.extend(['--lora', context.lora_file])
if 'SERVER_LOG_FORMAT_JSON' not in os.environ:
server_args.extend(['--log-format', "text"])

1
examples/server/tests/requirements.txt
View File

@ -4,3 +4,4 @@ huggingface_hub~=0.20.3
numpy~=1.26.4
openai~=1.30.3
prometheus-client~=0.20.0
requests~=2.32.3

18
examples/server/utils.hpp
View File

@ -355,24 +355,6 @@ static json oaicompat_completion_params_parse(
llama_params["__oaicompat"] = true;
// Map OpenAI parameters to llama.cpp parameters
//
// For parameters that are defined by the OpenAI documentation (e.g.
// temperature), we explicitly specify OpenAI's intended default; we
// need to do that because sometimes OpenAI disagrees with llama.cpp
//
// 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["frequency_penalty"] = json_value(body, "frequency_penalty", 0.0);
llama_params["logit_bias"] = json_value(body, "logit_bias", json::object());
llama_params["n_predict"] = json_value(body, "max_tokens", -1);
llama_params["presence_penalty"] = json_value(body, "presence_penalty", 0.0);
llama_params["seed"] = json_value(body, "seed", LLAMA_DEFAULT_SEED);
llama_params["stream"] = json_value(body, "stream", false);
llama_params["temperature"] = json_value(body, "temperature", 1.0);
llama_params["top_p"] = json_value(body, "top_p", 1.0);
// Apply chat template to the list of messages
llama_params["prompt"] = format_chat(model, chat_template, body.at("messages"));

2
examples/simple/README.md
View File

@ -3,7 +3,7 @@
The purpose of this example is to demonstrate a minimal usage of llama.cpp for generating text with a given prompt.
```bash
./simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
./llama-simple -m ./models/llama-7b-v2/ggml-model-f16.gguf -p "Hello my name is"
...

8
examples/speculative/speculative.cpp
View File

@ -66,7 +66,9 @@ int main(int argc, char ** argv) {
llama_context * ctx_dft = NULL;
// load the target model
std::tie(model_tgt, ctx_tgt) = llama_init_from_gpt_params(params);
llama_init_result llama_init_tgt = llama_init_from_gpt_params(params);
model_tgt = llama_init_tgt.model;
ctx_tgt = llama_init_tgt.context;
// load the draft model
params.model = params.model_draft;
@ -75,7 +77,9 @@ int main(int argc, char ** argv) {
params.n_threads = params.n_threads_draft;
}
params.n_threads_batch = params.n_threads_batch_draft;
std::tie(model_dft, ctx_dft) = llama_init_from_gpt_params(params);
llama_init_result llama_init_dft = llama_init_from_gpt_params(params);
model_dft = llama_init_dft.model;
ctx_dft = llama_init_dft.context;
const bool vocab_type_tgt = llama_vocab_type(model_tgt);
LOG("vocab_type tgt: %d\n", vocab_type_tgt);

24
examples/sycl/README.md
View File

@ -12,9 +12,9 @@ This example program provides the tools for llama.cpp for SYCL on Intel GPU.
List all SYCL devices with ID, compute capability, max work group size, ect.
1. Build the llama.cpp for SYCL for all targets.
1. Build the llama.cpp for SYCL for the specified target *(using GGML_SYCL_TARGET)*.
2. Enable oneAPI running environment
2. Enable oneAPI running environment *(if GGML_SYCL_TARGET is set to INTEL -default-)*
```
source /opt/intel/oneapi/setvars.sh
@ -29,19 +29,13 @@ source /opt/intel/oneapi/setvars.sh
Check the ID in startup log, like:
```
found 4 SYCL devices:
Device 0: Intel(R) Arc(TM) A770 Graphics, compute capability 1.3,
max compute_units 512, max work group size 1024, max sub group size 32, global mem size 16225243136
Device 1: Intel(R) FPGA Emulation Device, compute capability 1.2,
max compute_units 24, max work group size 67108864, max sub group size 64, global mem size 67065057280
Device 2: 13th Gen Intel(R) Core(TM) i7-13700K, compute capability 3.0,
max compute_units 24, max work group size 8192, max sub group size 64, global mem size 67065057280
Device 3: Intel(R) Arc(TM) A770 Graphics, compute capability 3.0,
max compute_units 512, max work group size 1024, max sub group size 32, global mem size 16225243136
found 2 SYCL devices:
| | | | |Max | |Max |Global | |
| | | | |compute|Max work|sub |mem | |
|ID| Device Type| Name|Version|units |group |group|size | Driver version|
|--|-------------------|---------------------------------------|-------|-------|--------|-----|-------|---------------------|
| 0| [level_zero:gpu:0]| Intel Arc A770 Graphics| 1.3| 512| 1024| 32| 16225M| 1.3.29138|
| 1| [level_zero:gpu:1]| Intel UHD Graphics 750| 1.3| 32| 512| 32| 62631M| 1.3.29138|
```
|Attribute|Note|
|-|-|
|compute capability 1.3|Level-zero running time, recommended |
|compute capability 3.0|OpenCL running time, slower than level-zero in most cases|

2
examples/sycl/win-run-llama2.bat
View File

@ -6,4 +6,4 @@ set INPUT2="Building a website can be done in 10 simple steps:\nStep 1:"
@call "C:\Program Files (x86)\Intel\oneAPI\setvars.bat" intel64 --force
.\build\bin\main.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0
.\build\bin\llama-cli.exe -m models\llama-2-7b.Q4_0.gguf -p %INPUT2% -n 400 -e -ngl 33 -s 0