tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 20:05:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

mtmd : merge llava, gemma3 and minicpmv CLI into single llama-mtmd-cli (#13012)

Browse Source 
* mtmd : merge `llava-cli` and `gemma3-cli` into single `mtmd-cli`

* support for minicpmv

* remove cpp files of llava and minicpmv

* update hot topics

* mtmd : add not supported msg for qwen2vl

* Update examples/llava/mtmd.cpp

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Xuan-Son Nguyen
2025-04-21 15:32:58 +02:00
committed by GitHub
parent 2016f07bd1
commit 84a9bf2fc2
10 changed files with 360 additions and 772 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
README.md
View File

@ -16,6 +16,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
## Hot topics ## Hot topics
- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated
- **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427 - **How to use [MTLResidencySet](https://developer.apple.com/documentation/metal/mtlresidencyset?language=objc) to keep the GPU memory active?** https://github.com/ggml-org/llama.cpp/pull/11427
- **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode - **VS Code extension for FIM completions:** https://github.com/ggml-org/llama.vscode
- Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639 - Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639

2
common/arg.cpp
View File

@ -2726,7 +2726,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
[](common_params & params, const std::string & value) { [](common_params & params, const std::string & value) {
params.chat_template = value; params.chat_template = value;
} }
).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER}).set_env("LLAMA_ARG_CHAT_TEMPLATE")); ).set_examples({LLAMA_EXAMPLE_MAIN, LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_LLAVA}).set_env("LLAMA_ARG_CHAT_TEMPLATE"));
add_opt(common_arg( add_opt(common_arg(
{"--chat-template-file"}, "JINJA_TEMPLATE_FILE", {"--chat-template-file"}, "JINJA_TEMPLATE_FILE",
string_format( string_format(

22
examples/llava/CMakeLists.txt
View File

@ -61,19 +61,9 @@ if(TARGET BUILD_INFO)
add_dependencies(mtmd BUILD_INFO) add_dependencies(mtmd BUILD_INFO)
endif() endif()
set(TARGET llama-llava-cli) add_executable(llama-llava-cli deprecation-warning.cpp)
add_executable(${TARGET} llava-cli.cpp) add_executable(llama-gemma3-cli deprecation-warning.cpp)
set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-llava-cli) add_executable(llama-minicpmv-cli deprecation-warning.cpp)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)
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_17)
set(TARGET llama-qwen2vl-cli) set(TARGET llama-qwen2vl-cli)
add_executable(${TARGET} qwen2vl-cli.cpp) add_executable(${TARGET} qwen2vl-cli.cpp)
@ -82,9 +72,9 @@ install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT}) target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17) target_compile_features(${TARGET} PRIVATE cxx_std_17)
set(TARGET llama-gemma3-cli) set(TARGET llama-mtmd-cli)
add_executable(${TARGET} gemma3-cli.cpp) add_executable(${TARGET} mtmd-cli.cpp)
set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-gemma3-cli) set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
install(TARGETS ${TARGET} RUNTIME) install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT}) target_link_libraries(${TARGET} PRIVATE common mtmd ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17) target_compile_features(${TARGET} PRIVATE cxx_std_17)

22
examples/llava/deprecation-warning.cpp Normal file
View File

@ -0,0 +1,22 @@
#include <cstdio>
#include <string>
int main(int argc, char** argv) {
std::string filename = "main";
if (argc >= 1) {
filename = argv[0];
}
// Get only the program name from the full path
size_t pos = filename.find_last_of("/\\");
if (pos != std::string::npos) {
filename = filename.substr(pos+1);
}
fprintf(stdout, "\n");
fprintf(stdout, "WARNING: The binary '%s' is deprecated.\n", filename.c_str());
fprintf(stdout, "Please use 'llama-mtmd-cli' instead.\n");
fprintf(stdout, "\n");
return EXIT_FAILURE;
}

332
examples/llava/llava-cli.cpp
View File

@ -1,332 +0,0 @@
#include "arg.h"
#include "base64.hpp"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "clip.h"
#include "llava.h"
#include "llama.h"
#include "ggml.h"
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <vector>
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))) {
LOG_ERR("%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 = common_tokenize(ctx_llama, str2, add_bos, true);
eval_tokens(ctx_llama, embd_inp, n_batch, n_past);
return true;
}
static const char * sample(struct common_sampler * smpl,
struct llama_context * ctx_llama,
int * n_past) {
const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
common_sampler_accept(smpl, id, true);
const llama_model * model = llama_get_model(ctx_llama);
const llama_vocab * vocab = llama_model_get_vocab(model);
static std::string ret;
if (llama_vocab_is_eog(vocab, id)) {
ret = "</s>";
} else {
ret = common_token_to_piece(ctx_llama, id);
}
eval_id(ctx_llama, id, n_past);
return ret.c_str();
}
static const char* IMG_BASE64_TAG_BEGIN = "<img src=\"data:image/jpeg;base64,";
static const char* IMG_BASE64_TAG_END = "\">";
static void find_image_tag_in_prompt(const std::string& prompt, size_t& begin_out, size_t& end_out) {
begin_out = prompt.find(IMG_BASE64_TAG_BEGIN);
end_out = prompt.find(IMG_BASE64_TAG_END, (begin_out == std::string::npos) ? 0UL : begin_out);
}
static bool prompt_contains_image(const std::string& prompt) {
size_t begin, end;
find_image_tag_in_prompt(prompt, begin, end);
return (begin != std::string::npos);
}
// replaces the base64 image tag in the prompt with `replacement`
static llava_image_embed * llava_image_embed_make_with_prompt_base64(struct clip_ctx * ctx_clip, int n_threads, const std::string& prompt) {
size_t img_base64_str_start, img_base64_str_end;
find_image_tag_in_prompt(prompt, img_base64_str_start, img_base64_str_end);
if (img_base64_str_start == std::string::npos || img_base64_str_end == std::string::npos) {
LOG_ERR("%s: invalid base64 image tag. must be %s<base64 byte string>%s\n", __func__, IMG_BASE64_TAG_BEGIN, IMG_BASE64_TAG_END);
return NULL;
}
auto base64_bytes_start = img_base64_str_start + strlen(IMG_BASE64_TAG_BEGIN);
auto base64_bytes_count = img_base64_str_end - base64_bytes_start;
auto base64_str = prompt.substr(base64_bytes_start, base64_bytes_count );
auto required_bytes = base64::required_encode_size(base64_str.size());
auto img_bytes = std::vector<unsigned char>(required_bytes);
base64::decode(base64_str.begin(), base64_str.end(), img_bytes.begin());
auto embed = llava_image_embed_make_with_bytes(ctx_clip, n_threads, img_bytes.data(), img_bytes.size());
if (!embed) {
LOG_ERR("%s: could not load image from base64 string.\n", __func__);
return NULL;
}
return embed;
}
static std::string remove_image_from_prompt(const std::string& prompt, const char * replacement = "") {
size_t begin, end;
find_image_tag_in_prompt(prompt, begin, end);
if (begin == std::string::npos || end == std::string::npos) {
return prompt;
}
auto pre = prompt.substr(0, begin);
auto post = prompt.substr(end + strlen(IMG_BASE64_TAG_END));
return pre + replacement + post;
}
struct llava_context {
struct clip_ctx * ctx_clip = NULL;
struct llama_context * ctx_llama = NULL;
struct llama_model * model = NULL;
};
static void print_usage(int, char ** argv) {
LOG("\n example usage:\n");
LOG("\n %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("\n note: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static struct llava_image_embed * load_image(llava_context * ctx_llava, common_params * params, const std::string & fname) {
// load and preprocess the image
llava_image_embed * embed = NULL;
auto prompt = params->prompt;
if (prompt_contains_image(prompt)) {
if (!params->image.empty()) {
LOG_INF("using base64 encoded image instead of command line image path\n");
}
embed = llava_image_embed_make_with_prompt_base64(ctx_llava->ctx_clip, params->cpuparams.n_threads, prompt);
if (!embed) {
LOG_ERR("%s: can't load image from prompt\n", __func__);
return NULL;
}
params->prompt = remove_image_from_prompt(prompt);
} else {
embed = llava_image_embed_make_with_filename(ctx_llava->ctx_clip, params->cpuparams.n_threads, fname.c_str());
if (!embed) {
fprintf(stderr, "%s: is %s really an image file?\n", __func__, fname.c_str());
return NULL;
}
}
return embed;
}
static void process_prompt(struct llava_context * ctx_llava, struct llava_image_embed * image_embed, common_params * params, const std::string & prompt) {
int n_past = 0;
const int max_tgt_len = params->n_predict < 0 ? 256 : params->n_predict;
std::string system_prompt, user_prompt;
size_t image_pos = prompt.find("<image>");
if (image_pos != std::string::npos) {
// new templating mode: Provide the full prompt including system message and use <image> as a placeholder for the image
system_prompt = prompt.substr(0, image_pos);
user_prompt = prompt.substr(image_pos + std::string("<image>").length());
LOG_INF("system_prompt: %s\n", system_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = common_tokenize(ctx_llava->ctx_llama, system_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
LOG_INF("user_prompt: %s\n", user_prompt.c_str());
if (params->verbose_prompt) {
auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
} else {
// llava-1.5 native mode
system_prompt = "A chat between a curious human and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the human's questions.\nUSER:";
user_prompt = prompt + "\nASSISTANT:";
if (params->verbose_prompt) {
auto tmp = common_tokenize(ctx_llava->ctx_llama, user_prompt, true, true);
for (int i = 0; i < (int) tmp.size(); i++) {
LOG_INF("%6d -> '%s'\n", tmp[i], common_token_to_piece(ctx_llava->ctx_llama, tmp[i]).c_str());
}
}
}
eval_string(ctx_llava->ctx_llama, system_prompt.c_str(), params->n_batch, &n_past, true);
llava_eval_image_embed(ctx_llava->ctx_llama, image_embed, params->n_batch, &n_past);
eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
// generate the response
LOG("\n");
struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
if (!smpl) {
LOG_ERR("%s: failed to initialize sampling subsystem\n", __func__);
exit(1);
}
std::string response = "";
for (int i = 0; i < max_tgt_len; i++) {
const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
if (strstr(tmp, "###")) break; // Yi-VL behavior
LOG("%s", tmp);
if (strstr(response.c_str(), "<|im_end|>")) break; // Yi-34B llava-1.6 - for some reason those decode not as the correct token (tokenizer works)
if (strstr(response.c_str(), "<|im_start|>")) break; // Yi-34B llava-1.6
if (strstr(response.c_str(), "USER:")) break; // mistral llava-1.6
fflush(stdout);
}
common_sampler_free(smpl);
LOG("\n");
}
static struct llama_model * llava_init(common_params * params) {
llama_backend_init();
llama_numa_init(params->numa);
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
}
return model;
}
static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
const char * clip_path = params->mmproj.path.c_str();
auto prompt = params->prompt;
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
auto ctx_clip = clip_model_load(clip_path, GGML_LOG_LEVEL_INFO);
llama_context_params ctx_params = common_context_params_to_llama(*params);
ctx_params.n_ctx = params->n_ctx < 2048 ? 2048 : params->n_ctx; // we need a longer context size to process image embeddings
llama_context * ctx_llama = llama_init_from_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_ERR("%s: 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->ctx_clip = ctx_clip;
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_model_free(ctx_llava->model);
llama_backend_free();
}
int main(int argc, char ** argv) {
ggml_time_init();
common_params params;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, print_usage)) {
return 1;
}
common_init();
if (params.mmproj.path.empty() || (params.image.empty() && !prompt_contains_image(params.prompt))) {
print_usage(argc, argv);
return 1;
}
auto * model = llava_init(&params);
if (model == NULL) {
fprintf(stderr, "%s: error: failed to init llava model\n", __func__);
return 1;
}
if (prompt_contains_image(params.prompt)) {
auto * ctx_llava = llava_init_context(&params, model);
auto * image_embed = load_image(ctx_llava, &params, "");
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_perf_context_print(ctx_llava->ctx_llama);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);
} else {
for (auto & image : params.image) {
auto * ctx_llava = llava_init_context(&params, model);
auto * image_embed = load_image(ctx_llava, &params, image);
if (!image_embed) {
LOG_ERR("%s: failed to load image %s. Terminating\n\n", __func__, image.c_str());
return 1;
}
// process the prompt
process_prompt(ctx_llava, image_embed, &params, params.prompt);
llama_perf_context_print(ctx_llava->ctx_llama);
llava_image_embed_free(image_embed);
ctx_llava->model = NULL;
llava_free(ctx_llava);
}
}
llama_model_free(model);
return 0;
}

354
examples/llava/minicpmv-cli.cpp
View File

@ -1,354 +0,0 @@
#include "arg.h"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "clip.h"
#include "llava.h"
#include "llama.h"
#include "ggml.h"
#include <algorithm>
#include <cstdio>
#include <cstdlib>
#include <cstring>
#include <vector>
#include <iostream> // TODO: remove me
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("\nexample usage:\n\n%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("\nnote: a lower temperature value like 0.1 is recommended for better quality.\n");
}
static struct llama_model * llava_init(common_params * params) {
llama_backend_init();
llama_numa_init(params->numa);
llama_model_params model_params = common_model_params_to_llama(*params);
llama_model * model = llama_model_load_from_file(params->model.path.c_str(), model_params);
if (model == NULL) {
LOG_ERR("%s: unable to load model\n" , __func__);
return NULL;
}
return model;
}
static struct llava_context * llava_init_context(common_params * params, llama_model * model) {
auto prompt = params->prompt;
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
llama_context_params ctx_params = common_context_params_to_llama(*params);
if (params->n_ctx < 2048) {
// warn user here, "Image processing requires at least 2048 context, setting context to 2048"
LOG_WRN("%s: 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_init_from_model(model, ctx_params);
if (ctx_llama == NULL) {
LOG_ERR("%s: 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_model_free(ctx_llava->model);
llama_backend_free();
}
static struct clip_ctx * clip_init_context(common_params * params) {
const char * clip_path = params->mmproj.path.c_str();
auto prompt = params->prompt;
if (prompt.empty()) {
prompt = "describe the image in detail.";
}
struct clip_context_params clip_params = {
/* use_gpu */ params->n_gpu_layers != 0,
/* verbosity */ GGML_LOG_LEVEL_INFO, // TODO: make this configurable
};
auto * ctx_clip = clip_init(clip_path, clip_params);
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))) {
LOG_ERR("%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 = common_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, common_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);
int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
if (has_minicpmv_projector == 2) {
system_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n";
}
else if (has_minicpmv_projector == 3) {
system_prompt = "<|im_start|>user\n";
}
else if (has_minicpmv_projector == 4) {
system_prompt = "<|im_start|>user\n";
}
LOG_INF("%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) {
if (has_minicpmv_projector == 2) {
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);
}
else if (has_minicpmv_projector == 3 || has_minicpmv_projector == 4) {
size_t num_image_embeds_col = clip_uhd_num_image_embeds_col(ctx_llava->ctx_clip);
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("<slice>").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("</slice>").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);
}
}
}
}
}
LOG_INF("%s: image token past: %d\n", __func__, n_past);
}
static const char * sample(struct common_sampler * smpl,
struct llama_context * ctx_llama,
int * n_past) {
const llama_token id = common_sampler_sample(smpl, ctx_llama, -1);
common_sampler_accept(smpl, id, true);
const llama_model * model = llama_get_model(ctx_llama);
const llama_vocab * vocab = llama_model_get_vocab(model);
static std::string ret;
if (llama_vocab_is_eog(vocab, id)) {
ret = "</s>";
} else {
ret = common_token_to_piece(ctx_llama, id);
}
eval_id(ctx_llama, id, n_past);
return ret.c_str();
}
static struct llava_context * minicpmv_init(common_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->cpuparams.n_threads, fname.c_str());
if (!embeds) {
LOG_ERR("failed to load image %s. Terminating\n\n", fname.c_str());
return NULL;
}
// process the prompt
if (params->prompt.empty() && params->interactive == false) {
LOG_ERR("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_INF("%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_INF("%s: llama process image in %8.2f ms.\n", __func__, t_process_image_ms);
llava_image_embed_free(embeds);
return ctx_llava;
}
static struct common_sampler * llama_init(struct llava_context * ctx_llava, common_params * params, const std::string & prompt, int & n_past, bool is_first = false){
std::string user_prompt = prompt;
int has_minicpmv_projector = clip_is_minicpmv(ctx_llava->ctx_clip);
if (!is_first) {
if (has_minicpmv_projector == 2) {
user_prompt = "<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n" + prompt;
}
else if (has_minicpmv_projector == 3) {
user_prompt = "<|im_start|>user\n" + prompt;
}
else if (has_minicpmv_projector == 4) {
user_prompt = "<|im_start|>user\n" + prompt;
}
}
eval_string(ctx_llava->ctx_llama, user_prompt.c_str(), params->n_batch, &n_past, false);
if (has_minicpmv_projector == 2) {
eval_string(ctx_llava->ctx_llama, "<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", params->n_batch, &n_past, false);
}
else if (has_minicpmv_projector == 3) {
eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
}
else if (has_minicpmv_projector == 4) {
eval_string(ctx_llava->ctx_llama, "<|im_end|><|im_start|>assistant\n", params->n_batch, &n_past, false);
}
// generate the response
LOG_INF("\n");
struct common_sampler * smpl = common_sampler_init(ctx_llava->model, params->sampling);
return smpl;
}
static const char * llama_loop(struct llava_context * ctx_llava,struct common_sampler * smpl, int &n_past){
const char * tmp = sample(smpl, ctx_llava->ctx_llama, &n_past);
return tmp;
}
int main(int argc, char ** argv) {
ggml_time_init();
common_params params;
if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_LLAVA, show_additional_info)) {
return 1;
}
common_init();
if (params.mmproj.path.empty() || (params.image.empty())) {
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("<user>%s\n", params.prompt.c_str());
LOG("<assistant>");
auto * smpl = llama_init(ctx_llava, &params, params.prompt, 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++) {
const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0){
if (!have_tmp) {
continue;
}
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);
}
common_sampler_free(smpl);
}else {
while (true) {
LOG("<user>");
std::string prompt;
std::getline(std::cin, prompt);
LOG("<assistant>");
auto * smpl = 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++) {
const auto * tmp = llama_loop(ctx_llava, smpl, n_past);
response += tmp;
if (strcmp(tmp, "</s>") == 0) break;
printf("%s", tmp);// mistral llava-1.6
if (strstr(response.c_str(), "<user>")) break; // minicpm-v
fflush(stdout);
}
common_sampler_free(smpl);
}
}
printf("\n");
llama_perf_context_print(ctx_llava->ctx_llama);
ctx_llava->model = NULL;
llava_free(ctx_llava);
}
return 0;
}

50
examples/llava/gemma3-cli.cpp → examples/llava/mtmd-cli.cpp
View File

@ -28,15 +28,16 @@ static bool g_is_generating = false;
/** /**
* Please note that this is NOT a production-ready stuff. * Please note that this is NOT a production-ready stuff.
* It is a playground for trying Gemma 3 vision capabilities. * It is a playground for trying multimodal support in llama.cpp.
* For contributors: please keep this code simple and easy to understand. * For contributors: please keep this code simple and easy to understand.
*/ */
static void show_additional_info(int /*argc*/, char ** argv) { static void show_additional_info(int /*argc*/, char ** argv) {
LOG( LOG(
"Experimental CLI for using Gemma 3 vision model\n\n" "Experimental CLI for multimodal\n\n"
"Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n" "Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> -p <prompt>\n\n"
" -m and --mmproj are required\n" " -m and --mmproj are required\n"
" -hf user/repo can replace both -m and --mmproj in most cases\n"
" --image and -p are optional, if NOT provided, the CLI will run in chat mode\n", " --image and -p are optional, if NOT provided, the CLI will run in chat mode\n",
argv[0] argv[0]
); );
@ -56,7 +57,7 @@ static void sigint_handler(int signo) {
} }
#endif #endif
struct gemma3_context { struct mtmd_cli_context {
mtmd_context_ptr ctx_vision; mtmd_context_ptr ctx_vision;
common_init_result llama_init; common_init_result llama_init;
@ -70,18 +71,38 @@ struct gemma3_context {
// so here we don't need to keep track of chat history // so here we don't need to keep track of chat history
common_chat_templates_ptr tmpls; common_chat_templates_ptr tmpls;
// support for legacy templates (models not having EOT token)
llama_tokens antiprompt_tokens;
int n_threads = 1; int n_threads = 1;
llama_pos n_past = 0; llama_pos n_past = 0;
gemma3_context(common_params & params) : llama_init(common_init_from_params(params)) { mtmd_cli_context(common_params & params) : llama_init(common_init_from_params(params)) {
model = llama_init.model.get(); model = llama_init.model.get();
lctx = llama_init.context.get(); lctx = llama_init.context.get();
vocab = llama_model_get_vocab(model); vocab = llama_model_get_vocab(model);
n_threads = params.cpuparams.n_threads; n_threads = params.cpuparams.n_threads;
batch = llama_batch_init(params.n_batch, 0, 1); batch = llama_batch_init(params.n_batch, 0, 1);
n_batch = params.n_batch; n_batch = params.n_batch;
if (!llama_model_chat_template(model, nullptr) && params.chat_template.empty()) {
LOG_ERR("Model does not have chat template.\n");
LOG_ERR(" For old llava models, you may need to use '--chat-template vicuna'\n");
LOG_ERR(" For MobileVLM models, use '--chat-template deepseek'\n");
exit(1);
}
tmpls = common_chat_templates_init(model, params.chat_template); tmpls = common_chat_templates_init(model, params.chat_template);
LOG_INF("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja).c_str());
init_vision_context(params); init_vision_context(params);
// load antiprompt tokens for legacy templates
if (params.chat_template == "vicuna") {
antiprompt_tokens = common_tokenize(lctx, "ASSISTANT:", false, true);
} else if (params.chat_template == "deepseek") {
antiprompt_tokens = common_tokenize(lctx, "###", false, true);
}
} }
void init_vision_context(common_params & params) { void init_vision_context(common_params & params) {
@ -97,6 +118,17 @@ struct gemma3_context {
exit(1); exit(1);
} }
} }
bool check_antiprompt(const llama_tokens & generated_tokens) {
if (antiprompt_tokens.empty() || generated_tokens.size() < antiprompt_tokens.size()) {
return false;
}
return std::equal(
generated_tokens.end() - antiprompt_tokens.size(),
generated_tokens.end(),
antiprompt_tokens.begin()
);
}
}; };
struct decode_embd_batch { struct decode_embd_batch {
@ -132,7 +164,8 @@ struct decode_embd_batch {
} }
}; };
static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_predict) { static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
llama_tokens generated_tokens;
for (int i = 0; i < n_predict; i++) { for (int i = 0; i < n_predict; i++) {
if (i > n_predict || !g_is_generating) { if (i > n_predict || !g_is_generating) {
printf("\n"); printf("\n");
@ -140,9 +173,10 @@ static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_
} }
llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1); llama_token token_id = common_sampler_sample(smpl, ctx.lctx, -1);
generated_tokens.push_back(token_id);
common_sampler_accept(smpl, token_id, true); common_sampler_accept(smpl, token_id, true);
if (llama_vocab_is_eog(ctx.vocab, token_id)) { if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
printf("\n"); printf("\n");
break; // end of generation break; // end of generation
} }
@ -161,7 +195,7 @@ static int generate_response(gemma3_context & ctx, common_sampler * smpl, int n_
return 0; return 0;
} }
static int eval_message(gemma3_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) { static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vector<std::string> & images_fname, bool add_bos = false) {
std::vector<mtmd_bitmap> bitmaps; std::vector<mtmd_bitmap> bitmaps;
common_chat_templates_inputs tmpl_inputs; common_chat_templates_inputs tmpl_inputs;
@ -218,7 +252,7 @@ int main(int argc, char ** argv) {
return 1; return 1;
} }
gemma3_context ctx(params); mtmd_cli_context ctx(params);
printf("%s: %s\n", __func__, params.model.path.c_str()); printf("%s: %s\n", __func__, params.model.path.c_str());
bool is_single_turn = !params.prompt.empty() && !params.image.empty(); bool is_single_turn = !params.prompt.empty() && !params.image.empty();

239
examples/llava/mtmd.cpp
View File

@ -12,6 +12,15 @@
#include <limits> #include <limits>
#include <vector> #include <vector>
// slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings
// models not having it (llava-1.6) will process embeddings without any special tokens in-between
enum mtmd_slice_tmpl {
MTMD_SLICE_TMPL_NONE,
MTMD_SLICE_TMPL_MINICPMV_2_5,
MTMD_SLICE_TMPL_MINICPMV_2_6,
// TODO @ngxson : add support for idefics (SmolVLM)
};
struct mtmd_context { struct mtmd_context {
struct clip_ctx * ctx_clip; struct clip_ctx * ctx_clip;
const struct llama_model * text_model; const struct llama_model * text_model;
@ -21,6 +30,16 @@ struct mtmd_context {
int n_threads; int n_threads;
std::string image_marker; std::string image_marker;
// for minicpmv, we need special tokens in-between slices
mtmd_slice_tmpl slice_tmpl = MTMD_SLICE_TMPL_NONE;
llama_token tok_ov_img_start = LLAMA_TOKEN_NULL; // overview image
llama_token tok_ov_img_end = LLAMA_TOKEN_NULL; // overview image
llama_token tok_slices_start = LLAMA_TOKEN_NULL; // start of all slices
llama_token tok_slices_end = LLAMA_TOKEN_NULL; // end of all slices
llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice
llama_token tok_sli_img_end = LLAMA_TOKEN_NULL; // single slice
llama_token tok_row_end = LLAMA_TOKEN_NULL; // end of row
// TODO @ngxson : add timings // TODO @ngxson : add timings
mtmd_context(const char * mmproj_fname, mtmd_context(const char * mmproj_fname,
@ -38,11 +57,66 @@ struct mtmd_context {
throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname)); throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
} }
this->text_model = text_model; this->text_model = text_model;
GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead");
int minicpmv_version = clip_is_minicpmv(ctx_clip);
if (minicpmv_version == 2) {
// minicpmv 2.5 format:
// <image> (overview) </image><slice><image> (slice) </image><image> (slice) </image>\n ... </slice>
slice_tmpl = MTMD_SLICE_TMPL_MINICPMV_2_5;
tok_ov_img_start = lookup_token("<image>");
tok_ov_img_end = lookup_token("</image>");
tok_slices_start = lookup_token("<slice>");
tok_slices_end = lookup_token("</slice>");
tok_sli_img_start = tok_ov_img_start;
tok_sli_img_end = tok_ov_img_end;
tok_row_end = lookup_token("\n");
} else if (minicpmv_version == 3 || minicpmv_version == 4) {
// minicpmv 2.6 format:
// <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
slice_tmpl = MTMD_SLICE_TMPL_MINICPMV_2_6;
tok_ov_img_start = lookup_token("<image>");
tok_ov_img_end = lookup_token("</image>");
tok_sli_img_start = lookup_token("<slice>");
tok_sli_img_end = lookup_token("</slice>");
tok_row_end = lookup_token("\n");
} else if (minicpmv_version != 0) {
GGML_ASSERT(false && "unsupported minicpmv version");
}
} }
~mtmd_context() { ~mtmd_context() {
clip_free(ctx_clip); clip_free(ctx_clip);
} }
private:
llama_token lookup_token(const std::string & token_text) {
const llama_vocab * vocab = llama_model_get_vocab(text_model);
const int n_vocab = llama_vocab_n_tokens(vocab);
for (int i = 0; i < n_vocab; i++) {
if (token_to_piece(vocab, i, true) == token_text) {
return i;
}
}
return LLAMA_TOKEN_NULL;
}
std::string token_to_piece(const llama_vocab * vocab, llama_token token, bool special) {
std::string piece;
piece.resize(piece.capacity()); // using string internal cache, 15 bytes + '\n'
const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
if (n_chars < 0) {
piece.resize(-n_chars);
int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
GGML_ASSERT(check == -n_chars);
} else {
piece.resize(n_chars);
}
return piece;
}
}; };
struct mtmd_image_tokens_data { struct mtmd_image_tokens_data {
@ -102,21 +176,58 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
std::string prompt_modified(text.text); std::string prompt_modified(text.text);
std::string marker_modified(ctx->image_marker); std::string marker_modified(ctx->image_marker);
projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
// a bit hacky here, but works for now // a bit hacky here, but works for now
// for some models, we need to add prefix and suffix to the image embeddings // for some models, we need to add prefix and suffix to the image embeddings
if (proj_type == PROJECTOR_TYPE_GEMMA3) { if (clip_is_gemma3(ctx->ctx_clip)) {
// gemma 3
// <start_of_image> ... (image embeddings) ... <end_of_image> // <start_of_image> ... (image embeddings) ... <end_of_image>
marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>"; marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
string_replace_all(prompt_modified, ctx->image_marker, marker_modified); string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
} }
// llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
// for glm-edge, we don't need to add because the tokens are already in the returned embeddings
// TODO @ngxson : glm-edge : remove BOI / EOI tokens embeddings, decode them as normal tokens
std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker); std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
output.clear(); output.clear();
output.reserve(parts.size()); output.reserve(parts.size());
size_t i_img = 0; size_t i_img = 0;
// utility for adding raw tokens
auto add_text_chunk = [&output](std::vector<llama_token> && tokens) {
mtmd_input_chunk chunk{
MTMD_INPUT_CHUNK_TYPE_TEXT,
std::move(tokens),
{},
};
output.emplace_back(std::move(chunk));
};
// utility for splitting batch of multiple images into chunks of batch having single images
auto split_batch_to_chunk = [&ctx](clip_image_f32_batch && batch_f32, const std::string & id) {
std::vector<mtmd_input_chunk> chunks;
for (auto & entry : batch_f32.entries) {
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
image_tokens->nx = clip_n_patches(ctx->ctx_clip);
image_tokens->ny = 1;
image_tokens->batch_f32.entries.push_back(std::move(entry));
image_tokens->id = id;
mtmd_input_chunk chunk{
MTMD_INPUT_CHUNK_TYPE_IMAGE,
{},
std::move(image_tokens),
};
chunks.emplace_back(std::move(chunk));
}
return chunks;
};
for (const auto & part : parts) { for (const auto & part : parts) {
//printf("tokenizing part: %s\n", part.c_str()); //printf("tokenizing part: %s\n", part.c_str());
bool add_bos = &parts.front() == &part; bool add_bos = &parts.front() == &part;
@ -139,12 +250,13 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
return 1; return 1;
} }
// shim layer // convert mtmd_bitmap to clip_image_u8
clip_image_u8_ptr img_u8(clip_image_u8_init()); clip_image_u8_ptr img_u8(clip_image_u8_init());
img_u8->nx = bitmaps[i_img].nx; img_u8->nx = bitmaps[i_img].nx;
img_u8->ny = bitmaps[i_img].ny; img_u8->ny = bitmaps[i_img].ny;
img_u8->buf.resize(bitmaps[i_img].data.size()); img_u8->buf.resize(bitmaps[i_img].data.size());
std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3); std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);
clip_image_size img_u8_size{img_u8->nx, img_u8->ny};
// preprocess image // preprocess image
clip_image_f32_batch batch_f32; clip_image_f32_batch batch_f32;
@ -154,19 +266,70 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
return 2; return 2;
} }
if (ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6) {
// split batch into chunks of single images
auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img].id);
GGML_ASSERT(chunks.size() > 0);
// add overview image
add_text_chunk({ctx->tok_ov_img_start});
output.emplace_back(std::move(chunks.front()));
chunks.erase(chunks.begin());
add_text_chunk({ctx->tok_ov_img_end});
// add slices
if (!chunks.empty()) {
clip_add_load_image_size(ctx->ctx_clip, &img_u8_size);
int n_col = clip_uhd_num_image_embeds_col(ctx->ctx_clip);
int n_row = (int)chunks.size() / n_col;
GGML_ASSERT(n_row * n_col == (int)chunks.size());
if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
add_text_chunk({ctx->tok_slices_start});
}
for (int y = 0; y < n_row; y++) {
for (int x = 0; x < n_col; x++) {
if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) {
add_text_chunk({ctx->tok_sli_img_start});
}
output.emplace_back(std::move(chunks[y * n_col + x]));
if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) {
add_text_chunk({ctx->tok_sli_img_end});
}
}
if (ctx->tok_row_end != LLAMA_TOKEN_NULL && y != n_row - 1) {
add_text_chunk({ctx->tok_row_end});
}
}
if (ctx->tok_slices_end != LLAMA_TOKEN_NULL) {
add_text_chunk({ctx->tok_slices_end});
}
}
} else {
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens); mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
image_tokens->nx = clip_n_patches(ctx->ctx_clip); // TODO @ngxson : use clip_n_patches_by_image image_tokens->nx = clip_n_patches(ctx->ctx_clip) * batch_f32.entries.size(); // TODO @ngxson : use clip_n_patches_by_image
image_tokens->ny = 1; // TODO image_tokens->ny = 1; // TODO
image_tokens->batch_f32 = std::move(batch_f32); image_tokens->batch_f32 = std::move(batch_f32);
image_tokens->id = bitmaps[i_img].id; // optional image_tokens->id = bitmaps[i_img].id; // optional
LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx);
LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size());
if (clip_is_glm(ctx->ctx_clip)) {
// glm-edge
image_tokens->nx += 2; // add 2 for the begin_of_image and end_of_image token embeddings
}
mtmd_input_chunk chunk{ mtmd_input_chunk chunk{
MTMD_INPUT_CHUNK_TYPE_IMAGE, MTMD_INPUT_CHUNK_TYPE_IMAGE,
{}, {},
std::move(image_tokens), std::move(image_tokens),
}; };
output.emplace_back(std::move(chunk)); output.emplace_back(std::move(chunk));
i_img++; }
i_img++; // move to next image
} }
} }
@ -198,11 +361,35 @@ std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) { int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip); int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd); ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
bool ok = clip_image_batch_encode( bool ok = false;
// only effective for minicpmv and qwen2vl, other models will ignore load_image_size
{
clip_image_size slice_size{
image_tokens->batch_f32.entries[0]->nx,
image_tokens->batch_f32.entries[0]->ny};
clip_add_load_image_size(ctx->ctx_clip, &slice_size);
}
if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) {
// TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
const auto & entries = image_tokens->batch_f32.entries;
for (size_t i = 0; i < entries.size(); i++) {
int n_tokens_per_image = clip_n_patches(ctx->ctx_clip);
ok = clip_image_encode(
ctx->ctx_clip,
ctx->n_threads,
entries[i].get(),
ctx->image_embd_v.data() + i*n_mmproj_embd*n_tokens_per_image);
}
} else {
ok = clip_image_batch_encode(
ctx->ctx_clip, ctx->ctx_clip,
ctx->n_threads, ctx->n_threads,
&image_tokens->batch_f32, &image_tokens->batch_f32,
ctx->image_embd_v.data()); ctx->image_embd_v.data());
}
return ok ? 0 : 1; return ok ? 0 : 1;
} }
@ -268,13 +455,15 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
int32_t ret; int32_t ret;
llama_pos n_past = pos0; llama_pos n_past = pos0;
llama_batch text_batch = llama_batch_init(n_batch, 0, 1); llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
for (auto & chunk : chunks) { for (auto & chunk : chunks) {
bool is_last = &chunk == &chunks.back(); bool is_last = &chunk == &chunks.back();
if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) { if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
// TODO @ngxson : may need to split into smaller batches
text_batch.n_tokens = chunk.tokens_text.size(); text_batch.n_tokens = chunk.tokens_text.size();
for (size_t i = 0; i < chunk.tokens_text.size(); i++) { size_t i = 0;
while (i < chunk.tokens_text.size()) { // split into batches
for (; i < chunk.tokens_text.size() && text_batch.n_tokens < n_batch; i++) {
text_batch.token [i] = chunk.tokens_text[i]; text_batch.token [i] = chunk.tokens_text[i];
text_batch.pos [i] = n_past++; text_batch.pos [i] = n_past++;
text_batch.n_seq_id[i] = 1; text_batch.n_seq_id[i] = 1;
@ -291,13 +480,14 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
llama_batch_free(text_batch); llama_batch_free(text_batch);
return ret; return ret;
} }
}
} else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) { } else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
GGML_ASSERT(!is_last && "logits for last image chunk is not yet support"); GGML_ASSERT(!is_last && "logits for last image chunk is not yet support");
GGML_ASSERT(chunk.tokens_image != nullptr); GGML_ASSERT(chunk.tokens_image != nullptr);
int64_t t0 = ggml_time_ms(); int64_t t0 = ggml_time_ms();
if (ctx->print_timings) { if (ctx->print_timings) {
LOG_INF("encoding image...\n"); LOG_INF("encoding image or slice...\n");
} }
ret = mtmd_encode(ctx, chunk.tokens_image.get()); ret = mtmd_encode(ctx, chunk.tokens_image.get());
if (ret != 0) { if (ret != 0) {
@ -306,24 +496,47 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
return ret; return ret;
} }
if (ctx->print_timings) { if (ctx->print_timings) {
LOG_INF("image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0); LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
} }
int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get()); int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
int32_t i_batch = 0;
int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
float * embd = mtmd_get_output_embd(ctx); float * embd = mtmd_get_output_embd(ctx);
decode_embd_batch batch_img(embd, n_tokens, n_past, 0);
if (mtmd_decode_use_non_causal(ctx)) {
llama_set_causal_attn(lctx, false);
// TODO @ngxson : need to make sure only one image is processed at a time, and n_ubatch must be enough to hold the image
}
while (i_batch < n_img_batches) { // split into batches
int32_t pos_offset = i_batch*n_batch;
int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
float * embd_batch = embd + pos_offset*n_mmproj_embd;
decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0);
printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
int64_t t1 = ggml_time_ms(); int64_t t1 = ggml_time_ms();
ret = llama_decode(lctx, batch_img.batch); ret = llama_decode(lctx, batch_img.batch);
if (ret != 0) { if (ret != 0) {
LOG_ERR("failed to decode image\n"); LOG_ERR("failed to decode image\n");
llama_set_causal_attn(lctx, true); // restore causal attn
llama_batch_free(text_batch); llama_batch_free(text_batch);
return ret; return ret;
} }
if (ctx->print_timings) { if (ctx->print_timings) {
LOG_INF("image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1); LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
} }
n_past += n_tokens; i_batch++;
n_past += n_tokens_batch;
}
if (mtmd_decode_use_non_causal(ctx)) {
llama_set_causal_attn(lctx, true);
}
} else { } else {
GGML_ASSERT(false && "chunk type not supported"); GGML_ASSERT(false && "chunk type not supported");

34
examples/llava/tests.sh
View File

@ -17,26 +17,30 @@ cd $PROJ_ROOT
arr_bin=() arr_bin=()
arr_hf=() arr_hf=()
arr_tmpl=() # chat template
add_test() { add_test() {
local bin=$1 local bin=$1
local hf=$2 local hf=$2
local tmpl=${3:-""} # default to empty string if not provided
arr_bin+=("$bin") arr_bin+=("$bin")
arr_hf+=("$hf") arr_hf+=("$hf")
arr_tmpl+=("$tmpl")
} }
add_test "llama-gemma3-cli" "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M" add_test "llama-mtmd-cli" "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
add_test "llama-llava-cli" "cmp-nct/Yi-VL-6B-GGUF:Q5_K" add_test "llama-mtmd-cli" "guinmoon/MobileVLM-3B-GGUF:Q4_K_M" "deepseek"
add_test "llama-llava-cli" "guinmoon/MobileVLM-3B-GGUF:Q4_K_M" add_test "llama-mtmd-cli" "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
add_test "llama-llava-cli" "THUDM/glm-edge-v-5b-gguf:Q4_K_M" add_test "llama-mtmd-cli" "second-state/Llava-v1.5-7B-GGUF:Q2_K" "vicuna"
add_test "llama-llava-cli" "second-state/Llava-v1.5-7B-GGUF:Q2_K" add_test "llama-mtmd-cli" "cjpais/llava-1.6-mistral-7b-gguf:Q3_K" "vicuna"
add_test "llama-llava-cli" "cjpais/llava-1.6-mistral-7b-gguf:Q3_K" add_test "llama-mtmd-cli" "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
add_test "llama-llava-cli" "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M" add_test "llama-mtmd-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted
add_test "llama-minicpmv-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted add_test "llama-mtmd-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
add_test "llama-minicpmv-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K" add_test "llama-mtmd-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
add_test "llama-minicpmv-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
add_test "llama-qwen2vl-cli" "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M" add_test "llama-qwen2vl-cli" "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
# add_test "llama-mtmd-cli" "cmp-nct/Yi-VL-6B-GGUF:Q5_K" # this model has broken chat template, not usable
############### ###############
cmake --build build -j --target "${arr_bin[@]}" cmake --build build -j --target "${arr_bin[@]}"
@ -46,12 +50,20 @@ arr_res=()
for i in "${!arr_bin[@]}"; do for i in "${!arr_bin[@]}"; do
bin="${arr_bin[$i]}" bin="${arr_bin[$i]}"
hf="${arr_hf[$i]}" hf="${arr_hf[$i]}"
tmpl="${arr_tmpl[$i]}"
echo "Running test with binary: $bin and HF model: $hf" echo "Running test with binary: $bin and HF model: $hf"
echo "" echo ""
echo "" echo ""
output=$("$PROJ_ROOT/build/bin/$bin" -hf "$hf" --image $SCRIPT_DIR/test-1.jpeg -p "what is the publisher name of the newspaper?" --temp 0 2>&1 | tee /dev/tty) output=$(\
"$PROJ_ROOT/build/bin/$bin" \
-hf "$hf" \
--image $SCRIPT_DIR/test-1.jpeg \
-p "what is the publisher name of the newspaper?" \
--temp 0 -n 128 \
${tmpl:+--chat-template "$tmpl"} \
2>&1 | tee /dev/tty)
echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log echo "$output" > $SCRIPT_DIR/output/$bin-$(echo "$hf" | tr '/' '-').log

2
src/llama-chat.cpp
View File

@ -121,6 +121,8 @@ llm_chat_template llm_chat_detect_template(const std::string & tmpl) {
return LLM_CHAT_TEMPLATE_PHI_3; return LLM_CHAT_TEMPLATE_PHI_3;
} else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) { } else if (tmpl_contains("<|assistant|>") && tmpl_contains("<|user|>")) {
return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE; return tmpl_contains("</s>") ? LLM_CHAT_TEMPLATE_FALCON_3 : LLM_CHAT_TEMPLATE_GLMEDGE;
} else if (tmpl_contains("<|{{ item['role'] }}|>") && tmpl_contains("<|begin_of_image|>")) {
return LLM_CHAT_TEMPLATE_GLMEDGE;
} else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) { } else if (tmpl_contains("<|user|>") && tmpl_contains("<|endoftext|>")) {
return LLM_CHAT_TEMPLATE_ZEPHYR; return LLM_CHAT_TEMPLATE_ZEPHYR;
} else if (tmpl_contains("bos_token + message['role']")) { } else if (tmpl_contains("bos_token + message['role']")) {