ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-28 10:21:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Port mdmd from mainline + Qwen2/2.5-VL support (#798)

Browse Source 
* Add mtmd: the beginning

* Add mtmd: mtmd.cpp compiles

* Add mtmd: clip initialization compiles

* Add mtmd: clip.cpp compiles

* Add mtmd: builds successfully

* Add CPU implementation for GGML_OP_GLU

* Add CUDA implementation for GGML_OP_GLU

* Add CPU implementation for GGML_OP_CONV_2D and GGML_OP_CONV_2D_DW

* Add CUDA implementation for GGML_OP_CONV_2D and GGML_OP_CONV_2D_DW

* Add mtmd: refresh CPU rope

* Add mtmd: refresh CUDA rope

* Add mtmd: add Qwen2-VL

* Add mtmd: Qwen2.5-VL text seems to work with this change

* Add mtmd: fix swiglu

* Add mtmd: use LOG_TEE so generated tokens show up in terminal

* Add mtmd: do not attempt to load a GPU backend if none are available

* GLU, not GPU

* Fix typo

* Fix new/free mismatch

* LOG stuff

* Add mtmd: this fixes gibberish on second image

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
This commit is contained in:
Kawrakow
2025-09-27 08:45:29 +02:00
committed by GitHub
parent 367654f99e
commit 87e4762720
51 changed files with 115141 additions and 432 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
examples/CMakeLists.txt
View File

@@ -29,7 +29,6 @@ else()
add_subdirectory(imatrix)
add_subdirectory(infill)
add_subdirectory(llama-bench)
add_subdirectory(llava)
add_subdirectory(lookahead)
add_subdirectory(lookup)
add_subdirectory(main)
@@ -39,6 +38,7 @@ else()
add_subdirectory(quantize-stats)
add_subdirectory(quantize)
add_subdirectory(retrieval)
add_subdirectory(mtmd)
if (GGML_RPC)
add_subdirectory(rpc)
endif()

2
examples/main/main.cpp
View File

@@ -233,7 +233,7 @@ int main(int argc, char ** argv) {
if (params.conversation) {
if (params.enable_chat_template) {
//LOG_TEE("%s: chat template example: %s\n", __func__, common_chat_format_example(model, *chat_templates.template_default, params.use_jinja).c_str());
LOG_TEE("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja).c_str());
LOG_TEE("%s: chat template example:\n%s\n", __func__, common_chat_format_example(chat_templates.get(), params.use_jinja, {}).c_str());
} else {
LOG_TEE("%s: in-suffix/prefix is specified, chat template will be disabled\n", __func__);
}

62
examples/mtmd/CMakeLists.txt Normal file
View File

@@ -0,0 +1,62 @@
# mtmd
find_package(Threads REQUIRED)
add_library(mtmd
mtmd.cpp
mtmd-audio.cpp
mtmd.h
clip.cpp
clip.h
clip-impl.h
mtmd-helper.cpp
mtmd-helper.h
)
target_link_libraries (mtmd PUBLIC ggml llama)
target_link_libraries (mtmd PRIVATE Threads::Threads)
target_include_directories(mtmd PUBLIC .)
target_include_directories(mtmd PRIVATE ../..)
target_include_directories(mtmd PRIVATE ../../vendor)
target_compile_features (mtmd PRIVATE cxx_std_17)
if (BUILD_SHARED_LIBS)
set_target_properties (mtmd PROPERTIES POSITION_INDEPENDENT_CODE ON)
target_compile_definitions(mtmd PRIVATE LLAMA_BUILD)
target_compile_definitions(mtmd PUBLIC LLAMA_SHARED)
endif()
set(MTMD_PUBLIC_HEADERS
${CMAKE_CURRENT_SOURCE_DIR}/mtmd.h
${CMAKE_CURRENT_SOURCE_DIR}/mtmd-helper.h
)
set_target_properties(mtmd
PROPERTIES
PUBLIC_HEADER "${MTMD_PUBLIC_HEADERS}")
install(TARGETS mtmd LIBRARY PUBLIC_HEADER)
if (NOT MSVC)
# for stb_image.h and miniaudio.h
target_compile_options(mtmd PRIVATE -Wno-cast-qual)
endif()
if (TARGET BUILD_INFO)
add_dependencies(mtmd BUILD_INFO)
add_dependencies(mtmd-helper BUILD_INFO)
endif()
add_executable(llama-llava-cli deprecation-warning.cpp)
add_executable(llama-gemma3-cli deprecation-warning.cpp)
add_executable(llama-minicpmv-cli deprecation-warning.cpp)
add_executable(llama-qwen2vl-cli deprecation-warning.cpp)
set(TARGET llama-mtmd-cli)
add_executable (${TARGET} mtmd-cli.cpp)
set_target_properties (${TARGET} PROPERTIES OUTPUT_NAME llama-mtmd-cli)
if(LLAMA_TOOLS_INSTALL)
install(TARGETS ${TARGET} RUNTIME)
endif()
target_link_libraries (${TARGET} PRIVATE common mtmd Threads::Threads)
target_compile_features(${TARGET} PRIVATE cxx_std_17)

63
examples/mtmd/README.md Normal file
View File

@@ -0,0 +1,63 @@
# Multimodal Support in llama.cpp
This directory provides multimodal capabilities for `llama.cpp`. Initially intended as a showcase for running LLaVA models, its scope has expanded significantly over time to include various other vision-capable models. As a result, LLaVA is no longer the only multimodal architecture supported.
> [!IMPORTANT]
>
> Multimodal support can be viewed as a sub-project within `llama.cpp`. It is under **very heavy development**, and **breaking changes are expected**.
The naming and structure related to multimodal support have evolved, which might cause some confusion. Here's a brief timeline to clarify:
- [#3436](https://github.com/ggml-org/llama.cpp/pull/3436): Initial support for LLaVA 1.5 was added, introducing `llava.cpp` and `clip.cpp`. The `llava-cli` binary was created for model interaction.
- [#4954](https://github.com/ggml-org/llama.cpp/pull/4954): Support for MobileVLM was added, becoming the second vision model supported. This built upon the existing `llava.cpp`, `clip.cpp`, and `llava-cli` infrastructure.
- **Expansion & Fragmentation:** Many new models were subsequently added (e.g., [#7599](https://github.com/ggml-org/llama.cpp/pull/7599), [#10361](https://github.com/ggml-org/llama.cpp/pull/10361), [#12344](https://github.com/ggml-org/llama.cpp/pull/12344), and others). However, `llava-cli` lacked support for the increasingly complex chat templates required by these models. This led to the creation of model-specific binaries like `qwen2vl-cli`, `minicpmv-cli`, and `gemma3-cli`. While functional, this proliferation of command-line tools became confusing for users.
- [#12849](https://github.com/ggml-org/llama.cpp/pull/12849): `libmtmd` was introduced as a replacement for `llava.cpp`. Its goals include providing a single, unified command-line interface, improving the user/developer experience (UX/DX), and supporting both audio and image inputs.
- [#13012](https://github.com/ggml-org/llama.cpp/pull/13012): `mtmd-cli` was added, consolidating the various model-specific CLIs into a single tool powered by `libmtmd`.
## Pre-quantized models
See the list of pre-quantized model [here](../../docs/multimodal.md)
## How it works and what is `mmproj`?
Multimodal support in `llama.cpp` works by encoding images into embeddings using a separate model component, and then feeding these embeddings into the language model.
This approach keeps the multimodal components distinct from the core `libllama` library. Separating these allows for faster, independent development cycles. While many modern vision models are based on Vision Transformers (ViTs), their specific pre-processing and projection steps can vary significantly. Integrating this diverse complexity directly into `libllama` is currently challenging.
Consequently, running a multimodal model typically requires two GGUF files:
1. The standard language model file.
2. A corresponding **multimodal projector (`mmproj`)** file, which handles the image encoding and projection.
## What is `libmtmd`?
As outlined in the history, `libmtmd` is the modern library designed to replace the original `llava.cpp` implementation for handling multimodal inputs.
Built upon `clip.cpp` (similar to `llava.cpp`), `libmtmd` offers several advantages:
- **Unified Interface:** Aims to consolidate interaction for various multimodal models.
- **Improved UX/DX:** Features a more intuitive API, inspired by the `Processor` class in the Hugging Face `transformers` library.
- **Flexibility:** Designed to support multiple input types (text, audio, images) while respecting the wide variety of chat templates used by different models.
## How to obtain `mmproj`
Multimodal projector (`mmproj`) files are specific to each model architecture.
For the following models, you can use `convert_hf_to_gguf.py` with `--mmproj` flag to get the `mmproj` file:
- [Gemma 3](https://huggingface.co/collections/google/gemma-3-release-67c6c6f89c4f76621268bb6d) ; See the guide [here](../../docs/multimodal/gemma3.md) - Note: 1B variant does not have vision support
- SmolVLM (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
- SmolVLM2 (from [HuggingFaceTB](https://huggingface.co/HuggingFaceTB))
- [Pixtral 12B](https://huggingface.co/mistral-community/pixtral-12b) - only works with `transformers`-compatible checkpoint
- Qwen 2 VL and Qwen 2.5 VL (from [Qwen](https://huggingface.co/Qwen))
- [Mistral Small 3.1 24B](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503)
- InternVL 2.5 and InternVL 3 from [OpenGVLab](https://huggingface.co/OpenGVLab) (note: we don't support conversion of `InternVL3-*-hf` model, only non-HF version is supported ; `InternLM2Model` **text** model is not supported)
For older models, please refer to the relevant guide for instructions on how to obtain or create them:
NOTE: conversion scripts are located under `tools/mtmd/legacy-models`
- [LLaVA](../../docs/multimodal/llava.md)
- [MobileVLM](../../docs/multimodal/MobileVLM.md)
- [GLM-Edge](../../docs/multimodal/glmedge.md)
- [MiniCPM-V 2.5](../../docs/multimodal/minicpmv2.5.md)
- [MiniCPM-V 2.6](../../docs/multimodal/minicpmv2.6.md)
- [MiniCPM-o 2.6](../../docs/multimodal/minicpmo2.6.md)
- [IBM Granite Vision](../../docs/multimodal/granitevision.md)

474
examples/mtmd/clip-impl.h Normal file
View File

@@ -0,0 +1,474 @@
#include "ggml.h"
#include "clip.h"
#include <climits>
#include <cstdarg>
#include <cinttypes>
#include <string>
#include <map>
#include <sstream>
#include <vector>
#include <memory>
// Internal header for clip.cpp
#define KEY_FTYPE "general.file_type"
#define KEY_NAME "general.name"
#define KEY_DESCRIPTION "general.description"
#define KEY_PROJ_TYPE "clip.projector_type"
#define KEY_HAS_AUDIO_ENC "clip.has_audio_encoder"
#define KEY_HAS_VISION_ENC "clip.has_vision_encoder"
#define KEY_USE_GELU "clip.use_gelu"
#define KEY_USE_SILU "clip.use_silu"
#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_PROJ_DIM "clip.%s.projection_dim"
#define KEY_N_HEAD "clip.%s.attention.head_count"
#define KEY_LAYER_NORM_EPS "clip.%s.attention.layer_norm_epsilon"
// vision-specific
#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_FEATURE_LAYER "clip.vision.feature_layer"
#define KEY_PROJ_SCALE_FACTOR "clip.vision.projector.scale_factor"
#define KEY_SPATIAL_MERGE_SIZE "clip.vision.spatial_merge_size"
#define KEY_MM_PATCH_MERGE_TYPE "clip.vision.mm_patch_merge_type"
#define KEY_IMAGE_GRID_PINPOINTS "clip.vision.image_grid_pinpoints"
#define KEY_IMAGE_CROP_RESOLUTION "clip.vision.image_crop_resolution"
#define KEY_WIN_ATTN_PATTERN "clip.vision.n_wa_pattern"
#define KEY_ATTN_WINDOW_SIZE "clip.vision.window_size"
#define KEY_MINICPMV_VERSION "clip.minicpmv_version"
#define KEY_MINICPMV_QUERY_NUM "clip.minicpmv_query_num"
// audio-specific
#define KEY_A_NUM_MEL_BINS "clip.audio.num_mel_bins"
#define KEY_A_PROJ_STACK_FACTOR "clip.audio.projector.stack_factor"
//
// tensor name constants
//
#define TN_POS_EMBD "%s.position_embd.weight"
#define TN_CLASS_EMBD "v.class_embd"
#define TN_PATCH_EMBD "v.patch_embd.weight" // not rename tensor with ".0" postfix for backwrad compat
#define TN_PATCH_EMBD_1 "v.patch_embd.weight.1"
#define TN_PATCH_BIAS "v.patch_embd.bias"
#define TN_ATTN_K "%s.blk.%d.attn_k.%s"
#define TN_ATTN_Q "%s.blk.%d.attn_q.%s"
#define TN_ATTN_V "%s.blk.%d.attn_v.%s"
#define TN_ATTN_OUTPUT "%s.blk.%d.attn_out.%s"
#define TN_ATTN_K_NORM "%s.blk.%d.attn_k_norm.%s"
#define TN_ATTN_Q_NORM "%s.blk.%d.attn_q_norm.%s"
#define TN_FFN_DOWN "%s.blk.%d.ffn_down.%s"
#define TN_FFN_GATE "%s.blk.%d.ffn_gate.%s"
#define TN_FFN_UP "%s.blk.%d.ffn_up.%s"
#define TN_FFN_GATE "%s.blk.%d.ffn_gate.%s"
#define TN_LN_1 "%s.blk.%d.ln1.%s" // layer norm
#define TN_LN_2 "%s.blk.%d.ln2.%s" // layer norm
#define TN_LS_1 "%s.blk.%d.ls1.%s" // layer scale
#define TN_LS_2 "%s.blk.%d.ls2.%s" // layer scale
#define TN_LN_PRE "%s.pre_ln.%s"
#define TN_LN_POST "%s.post_ln.%s"
#define TN_LLAVA_PROJ "mm.%d.%s"
#define TN_MVLM_PROJ_MLP "mm.model.mlp.%d.%s"
#define TN_MVLM_PROJ_BLOCK "mm.model.mb_block.%d.block.%d.%s"
#define TN_MVLM_PROJ_PEG "mm.model.peg.%d.%s"
#define TN_IMAGE_NEWLINE "model.image_newline"
#define TN_MM_INP_NORM "mm.input_norm.weight"
#define TN_MM_INP_NORM_B "mm.input_norm.bias"
#define TN_MM_INP_PROJ "mm.input_projection.weight" // gemma3
#define TN_MM_SOFT_EMB_N "mm.soft_emb_norm.weight" // gemma3
#define TN_MM_PROJECTOR "mm.model.fc.weight" // idefics3
#define TN_MM_PATCH_MERGER "mm.patch_merger.weight" // mistral small 3.1
#define TN_TOK_IMG_BREAK "v.token_embd.img_break" // pixtral
#define TN_TOK_GLM_BOI "adapter.boi" // glm-edge (these embeddings are not in text model)
#define TN_TOK_GLM_EOI "adapter.eoi" // glm-edge (these embeddings are not in text model)
// mimicpmv
#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"
#define TN_GLM_ADAPER_CONV "adapter.conv.%s"
#define TN_GLM_ADAPTER_LINEAR "adapter.linear.linear.%s"
#define TN_GLM_ADAPTER_NORM_1 "adapter.linear.norm1.%s"
#define TN_GLM_ADAPTER_D_H_2_4H "adapter.linear.dense_h_to_4h.%s"
#define TN_GLM_ADAPTER_GATE "adapter.linear.gate.%s"
#define TN_GLM_ADAPTER_D_4H_2_H "adapter.linear.dense_4h_to_h.%s"
// ultravox
#define TN_CONV1D "a.conv1d.%d.%s"
#define TN_MM_AUDIO_MLP "mm.a.mlp.%d.%s"
#define TN_MM_AUDIO_FC "mm.a.fc.%s" // fully connected layer
#define TN_MM_NORM_PRE "mm.a.norm_pre.%s"
#define TN_MM_NORM_MID "mm.a.norm_mid.%s"
// align x to upper multiple of n
#define CLIP_ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
enum projector_type {
PROJECTOR_TYPE_MLP,
PROJECTOR_TYPE_MLP_NORM,
PROJECTOR_TYPE_LDP,
PROJECTOR_TYPE_LDPV2,
PROJECTOR_TYPE_MINICPMV,
PROJECTOR_TYPE_GLM_EDGE,
PROJECTOR_TYPE_QWEN2VL,
PROJECTOR_TYPE_GEMMA3,
PROJECTOR_TYPE_IDEFICS3,
PROJECTOR_TYPE_PIXTRAL,
PROJECTOR_TYPE_QWEN25VL,
PROJECTOR_TYPE_ULTRAVOX,
PROJECTOR_TYPE_INTERNVL,
PROJECTOR_TYPE_LLAMA4,
PROJECTOR_TYPE_QWEN2A,
PROJECTOR_TYPE_QWEN25O, // will be replaced by QWEN2A or QWEN25VL depending on clip_ctx
PROJECTOR_TYPE_VOXTRAL,
PROJECTOR_TYPE_LFM2,
PROJECTOR_TYPE_KIMIVL,
PROJECTOR_TYPE_UNKNOWN,
};
static std::map<projector_type, std::string> PROJECTOR_TYPE_NAMES = {
{ PROJECTOR_TYPE_MLP, "mlp" },
{ PROJECTOR_TYPE_LDP, "ldp" },
{ PROJECTOR_TYPE_LDPV2, "ldpv2"},
{ PROJECTOR_TYPE_MINICPMV, "resampler"},
{ PROJECTOR_TYPE_GLM_EDGE, "adapter"},
{ PROJECTOR_TYPE_QWEN2VL, "qwen2vl_merger"},
{ PROJECTOR_TYPE_QWEN25VL, "qwen2.5vl_merger"},
{ PROJECTOR_TYPE_GEMMA3, "gemma3"},
{ PROJECTOR_TYPE_IDEFICS3, "idefics3"},
{ PROJECTOR_TYPE_PIXTRAL, "pixtral"},
{ PROJECTOR_TYPE_ULTRAVOX, "ultravox"},
{ PROJECTOR_TYPE_INTERNVL, "internvl"},
{ PROJECTOR_TYPE_LLAMA4, "llama4"},
{ PROJECTOR_TYPE_QWEN2A, "qwen2a"},
{ PROJECTOR_TYPE_QWEN25O, "qwen2.5o"},
{ PROJECTOR_TYPE_VOXTRAL, "voxtral"},
{ PROJECTOR_TYPE_LFM2, "lfm2"},
{ PROJECTOR_TYPE_KIMIVL, "kimivl"},
};
static projector_type clip_projector_type_from_string(const std::string & str) {
for (const auto & pair : PROJECTOR_TYPE_NAMES) {
if (pair.second == str) {
return pair.first;
}
}
return PROJECTOR_TYPE_UNKNOWN;
}
// RGB uint8 image
struct clip_image_u8 {
int nx;
int ny;
std::vector<uint8_t> buf;
};
// For images, buf.size() == nx*ny*3
// Memory layout: RGBRGBRGB...
// For audio, only one channel is used, buf.size() == nx*ny
// nx will be n_frames and ny will be n_mel
struct clip_image_f32 {
int nx;
int ny;
std::vector<float> buf;
};
//
// logging
//
static void clip_log_callback_default(enum ggml_log_level level, const char * text, void * user_data) {
(void) level;
(void) user_data;
fputs(text, stderr);
fflush(stderr);
}
struct clip_logger_state {
ggml_log_level verbosity_thold;
ggml_log_callback log_callback;
void * log_callback_user_data;
};
extern struct clip_logger_state g_logger_state;
static void clip_log_internal_v(enum ggml_log_level level, const char * format, va_list args) {
if (format == NULL) {
return;
}
va_list args_copy;
va_copy(args_copy, args);
char buffer[128];
int len = vsnprintf(buffer, 128, format, args);
if (len < 128) {
g_logger_state.log_callback(level, buffer, g_logger_state.log_callback_user_data);
} else {
char * buffer2 = (char *) calloc(len + 1, sizeof(char));
vsnprintf(buffer2, len + 1, format, args_copy);
buffer2[len] = 0;
g_logger_state.log_callback(level, buffer2, g_logger_state.log_callback_user_data);
free(buffer2);
}
va_end(args_copy);
}
static void clip_log_internal(enum ggml_log_level level, const char * format, ...) {
va_list args;
va_start(args, format);
clip_log_internal_v(level, format, args);
va_end(args);
}
#define LOG_TMPL(level, ...) \
do { \
if ((level) >= g_logger_state.verbosity_thold) { \
clip_log_internal((level), __VA_ARGS__); \
} \
} while (0)
#define LOG_INF(...) LOG_TMPL(GGML_LOG_LEVEL_INFO, __VA_ARGS__)
#define LOG_WRN(...) LOG_TMPL(GGML_LOG_LEVEL_WARN, __VA_ARGS__)
#define LOG_ERR(...) LOG_TMPL(GGML_LOG_LEVEL_ERROR, __VA_ARGS__)
#define LOG_DBG(...) LOG_TMPL(GGML_LOG_LEVEL_DEBUG, __VA_ARGS__)
#define LOG_CNT(...) LOG_TMPL(GGML_LOG_LEVEL_CONT, __VA_ARGS__)
//
// cpp wrappers
//
// wrapper for clip_image_size
struct clip_image_size_deleter {
void operator()(clip_image_size * val) { clip_image_size_free(val); }
};
typedef std::unique_ptr<clip_image_size, clip_image_size_deleter> clip_image_size_ptr;
// wrapper for clip_image_u8
struct clip_image_u8_deleter {
void operator()(clip_image_u8 * val) { clip_image_u8_free(val); }
};
typedef std::unique_ptr<clip_image_u8, clip_image_u8_deleter> clip_image_u8_ptr;
// wrapper for clip_image_f32
struct clip_image_f32_deleter {
void operator()(clip_image_f32 * val) { clip_image_f32_free(val); }
};
typedef std::unique_ptr<clip_image_f32, clip_image_f32_deleter> clip_image_f32_ptr;
struct clip_image_u8_batch {
std::vector<clip_image_u8_ptr> entries;
};
struct clip_image_f32_batch {
std::vector<clip_image_f32_ptr> entries;
bool is_audio = false;
// for llava-uhd style models, we need to know the grid size
// note: entries.size() == grid_x * grid_y + 1 (one overview image)
int grid_x = 0;
int grid_y = 0;
clip_image_f32_batch clone() const {
clip_image_f32_batch new_batch{
/* entries */ {},
/* is_audio */ is_audio,
/* grid_x */ grid_x,
/* grid_y */ grid_y,
};
new_batch.entries.reserve(entries.size());
for (const auto & entry : entries) {
new_batch.entries.emplace_back(new clip_image_f32(*entry));
}
return new_batch;
}
};
//
// common utils
//
static std::string string_format(const char * fmt, ...) {
va_list ap;
va_list ap2;
va_start(ap, fmt);
va_copy(ap2, ap);
int size = vsnprintf(NULL, 0, fmt, ap);
GGML_ASSERT(size >= 0 && size < INT_MAX); // NOLINT
std::vector<char> buf(size + 1);
int size2 = vsnprintf(buf.data(), size + 1, fmt, ap2);
GGML_ASSERT(size2 == size);
va_end(ap2);
va_end(ap);
return std::string(buf.data(), buf.size());
}
static void string_replace_all(std::string & s, const std::string & search, const std::string & replace) {
if (search.empty()) {
return;
}
std::string builder;
builder.reserve(s.length());
size_t pos = 0;
size_t last_pos = 0;
while ((pos = s.find(search, last_pos)) != std::string::npos) {
builder.append(s, last_pos, pos - last_pos);
builder.append(replace);
last_pos = pos + search.length();
}
builder.append(s, last_pos, std::string::npos);
s = std::move(builder);
}
// split string by a `std::string delim` instead of `char delim`
static std::vector<std::string> string_split_str(std::string s, const std::string & delimiter) {
std::vector<std::string> tokens;
size_t pos = 0;
std::string token;
while ((pos = s.find(delimiter)) != std::string::npos) {
token = s.substr(0, pos);
tokens.push_back(token);
s.erase(0, pos + delimiter.length());
}
tokens.push_back(s);
return tokens;
}
//
// gguf utils
//
static std::string gguf_data_to_str(enum gguf_type type, const void * data, int i) {
switch (type) {
case GGUF_TYPE_UINT8: return std::to_string(((const uint8_t *)data)[i]);
case GGUF_TYPE_INT8: return std::to_string(((const int8_t *)data)[i]);
case GGUF_TYPE_UINT16: return std::to_string(((const uint16_t *)data)[i]);
case GGUF_TYPE_INT16: return std::to_string(((const int16_t *)data)[i]);
case GGUF_TYPE_UINT32: return std::to_string(((const uint32_t *)data)[i]);
case GGUF_TYPE_INT32: return std::to_string(((const int32_t *)data)[i]);
case GGUF_TYPE_UINT64: return std::to_string(((const uint64_t *)data)[i]);
case GGUF_TYPE_INT64: return std::to_string(((const int64_t *)data)[i]);
case GGUF_TYPE_FLOAT32: return std::to_string(((const float *)data)[i]);
case GGUF_TYPE_FLOAT64: return std::to_string(((const double *)data)[i]);
case GGUF_TYPE_BOOL: return ((const bool *)data)[i] ? "true" : "false";
default: return string_format("unknown type %d", type);
}
}
static std::string gguf_kv_to_str(const struct gguf_context * ctx_gguf, int i) {
const enum gguf_type type = gguf_get_kv_type(ctx_gguf, i);
switch (type) {
case GGUF_TYPE_STRING:
return gguf_get_val_str(ctx_gguf, i);
case GGUF_TYPE_ARRAY:
{
const enum gguf_type arr_type = gguf_get_arr_type(ctx_gguf, i);
int arr_n = gguf_get_arr_n(ctx_gguf, i);
const void * data = arr_type == GGUF_TYPE_STRING ? nullptr : gguf_get_arr_data(ctx_gguf, i);
std::stringstream ss;
ss << "[";
for (int j = 0; j < arr_n; j++) {
if (arr_type == GGUF_TYPE_STRING) {
std::string val = gguf_get_arr_str(ctx_gguf, i, j);
// escape quotes
string_replace_all(val, "\\", "\\\\");
string_replace_all(val, "\"", "\\\"");
ss << '"' << val << '"';
} else if (arr_type == GGUF_TYPE_ARRAY) {
ss << "???";
} else {
ss << gguf_data_to_str(arr_type, data, j);
}
if (j < arr_n - 1) {
ss << ", ";
}
}
ss << "]";
return ss.str();
}
default:
return gguf_data_to_str(type, gguf_get_val_data(ctx_gguf, i), 0);
}
}
//
// debugging
//
static void print_tensor_shape(ggml_tensor * t) {
printf("%s.shape = [", t->name);
for (int i = 0; i < ggml_n_dims(t); ++i) {
printf("%" PRId64, t->ne[i]);
if (i < ggml_n_dims(t) - 1) {
printf(", ");
}
}
printf("]\n");
}
static void print_tensor_data(ggml_tensor * t, uint8_t * data, int64_t n) {
ggml_type type = t->type;
int64_t * ne = t->ne;
size_t * nb = t->nb;
for (int64_t i3 = 0; i3 < ne[3]; i3++) {
printf("%s.data: [\n", t->name);
for (int64_t i2 = 0; i2 < ne[2]; i2++) {
if (i2 == n && ne[2] > 2*n) {
printf(" ..., \n");
i2 = ne[2] - n;
}
printf(" [\n");
for (int64_t i1 = 0; i1 < ne[1]; i1++) {
if (i1 == n && ne[1] > 2*n) {
printf(" ..., \n");
i1 = ne[1] - n;
}
printf(" [");
for (int64_t i0 = 0; i0 < ne[0]; i0++) {
if (i0 == n && ne[0] > 2*n) {
printf("..., ");
i0 = ne[0] - n;
}
size_t i = i3 * nb[3] + i2 * nb[2] + i1 * nb[1] + i0 * nb[0];
float v;
if (type == GGML_TYPE_F16) {
v = ggml_fp16_to_fp32(*(ggml_fp16_t *) &data[i]);
} else if (type == GGML_TYPE_F32) {
v = *(float *) &data[i];
} else if (type == GGML_TYPE_I32) {
v = (float) *(int32_t *) &data[i];
} else if (type == GGML_TYPE_I16) {
v = (float) *(int16_t *) &data[i];
} else if (type == GGML_TYPE_I8) {
v = (float) *(int8_t *) &data[i];
} else {
GGML_ABORT("fatal error");
}
printf("%8.4f", v);
if (i0 < ne[0] - 1) printf(", ");
}
printf("],\n");
}
printf(" ],\n");
}
printf(" ]\n");
}
}
//
// API used internally with mtmd
//
projector_type clip_get_projector_type(const struct clip_ctx * ctx);

4446
examples/mtmd/clip.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

106
examples/mtmd/clip.h Normal file
View File

@@ -0,0 +1,106 @@
#pragma once
#include "ggml.h"
#include <stddef.h>
#include <stdint.h>
// !!! Internal header, to be used by mtmd only !!!
struct clip_ctx;
struct clip_image_size {
int width;
int height;
};
struct clip_image_f32;
struct clip_image_u8_batch;
struct clip_image_f32_batch;
enum clip_modality {
CLIP_MODALITY_VISION,
CLIP_MODALITY_AUDIO,
};
struct clip_context_params {
bool use_gpu;
enum ggml_log_level verbosity;
};
struct clip_init_result {
struct clip_ctx * ctx_v; // vision context
struct clip_ctx * ctx_a; // audio context
};
struct clip_init_result clip_init(const char * fname, struct clip_context_params ctx_params);
void clip_free(struct clip_ctx * ctx);
size_t clip_embd_nbytes(const struct clip_ctx * ctx);
size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h);
int32_t clip_get_image_size (const struct clip_ctx * ctx);
int32_t clip_get_patch_size (const struct clip_ctx * ctx);
int32_t clip_get_hidden_size(const struct clip_ctx * ctx);
// TODO: should be enum, not string
const char * clip_patch_merge_type(const struct clip_ctx * ctx);
int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img);
// for M-RoPE, this will be the number of token positions in X and Y directions
// for other models, X will be the total number of tokens and Y will be 1
int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img);
// this should be equal to the embedding dimension of the text model
int clip_n_mmproj_embd(const struct clip_ctx * ctx);
struct clip_image_size * clip_image_size_init(void);
struct clip_image_u8 * clip_image_u8_init (void);
struct clip_image_f32 * clip_image_f32_init(void);
struct clip_image_f32_batch * clip_image_f32_batch_init(void); // only used by libllava
// nx, ny are the output image dimensions
unsigned char * clip_image_u8_get_data(struct clip_image_u8 * img, uint32_t * nx, uint32_t * ny);
void clip_image_size_free (struct clip_image_size * img_size);
void clip_image_u8_free (struct clip_image_u8 * img);
void clip_image_f32_free(struct clip_image_f32 * img);
void clip_image_u8_batch_free (struct clip_image_u8_batch * batch);
void clip_image_f32_batch_free(struct clip_image_f32_batch * batch);
// use for accessing underlay data of clip_image_f32_batch
size_t clip_image_f32_batch_n_images(const struct clip_image_f32_batch * batch); // equivalent to batch->size()
size_t clip_image_f32_batch_nx(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->nx
size_t clip_image_f32_batch_ny(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->ny
struct clip_image_f32 * clip_image_f32_get_img(const struct clip_image_f32_batch * batch, int idx); // equivalent to batch[idx]->data
/**
* Build image from pixels decoded by other libraries instead of stb_image.h for better performance.
* The memory layout is RGBRGBRGB..., input buffer length must be 3*nx*ny bytes
*/
void clip_build_img_from_pixels(const unsigned char * rgb_pixels, int nx, int ny, struct clip_image_u8 * img);
/** preprocess img and store the result in res_imgs, pad_to_square may be overridden to false depending on model configuration */
bool clip_image_preprocess(struct clip_ctx * ctx, const struct clip_image_u8 * img, struct clip_image_f32_batch * res_imgs );
struct ggml_tensor * clip_get_newline_tensor(const struct clip_ctx * ctx);
bool clip_image_encode (struct clip_ctx * ctx, int n_threads, struct clip_image_f32 * img, float * vec);
bool clip_image_batch_encode(struct clip_ctx * ctx, int n_threads, const struct clip_image_f32_batch * imgs, float * vec);
int clip_is_minicpmv(const struct clip_ctx * ctx);
bool clip_is_glm(const struct clip_ctx * ctx);
bool clip_is_qwen2vl(const struct clip_ctx * ctx);
bool clip_is_llava(const struct clip_ctx * ctx);
bool clip_is_gemma3(const struct clip_ctx * ctx);
bool clip_encode_float_image (struct clip_ctx * ctx, int n_threads, float * img, int h, int w, float * vec);
// use by audio input
void clip_image_f32_batch_add_mel(struct clip_image_f32_batch * batch, int n_mel, int n_frames, float * mel);
bool clip_has_vision_encoder(const struct clip_ctx * ctx);
bool clip_has_audio_encoder(const struct clip_ctx * ctx);
bool clip_has_whisper_encoder(const struct clip_ctx * ctx);

22
examples/mtmd/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;
}

412
examples/mtmd/legacy-models/convert_image_encoder_to_gguf.py Normal file
View File

@@ -0,0 +1,412 @@
import argparse
import os
import json
import re
import torch
import numpy as np
from gguf import *
from transformers import CLIPModel, CLIPProcessor, CLIPVisionModel, SiglipVisionModel
TEXT = "clip.text"
VISION = "clip.vision"
def k(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_llava: bool) -> bool:
if name in (
"logit_scale",
"text_model.embeddings.position_ids",
"vision_model.embeddings.position_ids",
):
return True
if has_llava and name in ["visual_projection.weight", "vision_model.post_layernorm.weight", "vision_model.post_layernorm.bias"]:
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:
# Standardize the transformers llava next keys for
# image newline / mm projector with the classes in haotian-liu LLaVA
if name == "image_newline":
return "model.image_newline"
if name.startswith("multi_modal_projector"):
name = name.replace("multi_modal_projector", "mm")
if "linear_1" in name:
name = name.replace("linear_1", "0")
if "linear_2" in name:
name = name.replace("linear_2", "2")
return name
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('--bigendian', action="store_true", default=False, help="Model is executed on big-endian machine")
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)")
# Selectable visual encoders that are compatible with this script
encoder_group = ap.add_mutually_exclusive_group()
encoder_group.add_argument("--clip-model-is-openclip", action="store_true", required=False,
help="The clip model is from openclip (for ViT-SO400M type))")
encoder_group.add_argument("--clip-model-is-siglip", action="store_true", required=False,
help="the visual encoder is Siglip.")
ap.add_argument("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA 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 or
args.clip_model_is_siglip
):
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]
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
config = json.load(f)
if args.clip_model_is_vision:
v_hparams = config
t_hparams = None
else:
v_hparams = config["vision_config"]
t_hparams = config["text_config"]
# 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_siglip:
model = SiglipVisionModel.from_pretrained(dir_model)
processor = None
elif 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)
fname_middle = None
has_text_encoder = True
has_vision_encoder = True
has_llava_projector = False
if args.text_only:
fname_middle = "text-"
has_vision_encoder = False
elif args.llava_projector is not None:
fname_middle = "mmproj-"
has_text_encoder = False
has_llava_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", endianess=GGUFEndian.LITTLE if not args.bigendian else GGUFEndian.BIG)
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_llava_projector", has_llava_projector)
fout.add_file_type(ftype)
model_name = config["_name_or_path"] if "_name_or_path" in config else os.path.basename(dir_model)
fout.add_name(model_name)
if args.text_only:
fout.add_description("text-only CLIP model")
elif args.vision_only and not has_llava_projector:
fout.add_description("vision-only CLIP model")
elif has_llava_projector:
fout.add_description("image encoder for LLaVA")
# add projector type
fout.add_string("clip.projector_type", args.projector_type)
else:
fout.add_description("two-tower CLIP model")
if has_text_encoder:
assert t_hparams is not None
assert tokens is not None
if args.clip_model_is_siglip:
text_projection_dim = 0
else:
text_projection_dim = t_hparams.get("projection_dim", config["projection_dim"])
# text_model hparams
fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
fout.add_uint32("clip.text.projection_dim", text_projection_dim)
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
fout.add_token_list(tokens)
def get_non_negative_vision_feature_layers(v_hparams):
"""
Determine the vision feature layer(s) for the llava model, which are indices into the
hidden states of the visual encoder. Note that the hidden states array generally takes the
form:
[<emb input>, <output of enc block 0>, ... <output of enc block num_hidden_layers>]
so feature indices should be offset as n+1 to get the output of encoder block n.
We convert all vision feature layers to non-negative so that -1 can be used in
the model as an unset value. If no vision feature layer is found, we leave it unset.
"""
num_hidden_layers = v_hparams["num_hidden_layers"]
to_non_negative = lambda layer_idx: layer_idx if layer_idx >= 0 else num_hidden_layers + layer_idx + 1
feature_layers_key = None
# Key used for llava models in transformers
if "vision_feature_layer" in config:
feature_layers_key = "vision_feature_layer"
# Key used for llava models in the original format
elif "mm_vision_select_layer" in config:
feature_layers_key = "mm_vision_select_layer"
if feature_layers_key is not None:
feature_layers = config[feature_layers_key]
if isinstance(feature_layers, int):
feature_layers = [feature_layers]
return [to_non_negative(feature_layer) for feature_layer in feature_layers]
# Determine if we have explicitly specified vision feature layers in our config
feature_layers = get_non_negative_vision_feature_layers(v_hparams)
if has_vision_encoder:
# Siglip does not have a visual projector; set projection dim to 0
if args.clip_model_is_siglip:
visual_projection_dim = 0
else:
visual_projection_dim = v_hparams.get("projection_dim", config["projection_dim"])
# set vision_model hparams
fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
fout.add_uint32("clip.vision.projection_dim", visual_projection_dim)
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), v_hparams["layer_norm_eps"])
if feature_layers:
block_count = max(feature_layers)
else:
block_count = v_hparams["num_hidden_layers"] - 1 if has_llava_projector else v_hparams["num_hidden_layers"]
fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), block_count)
# /**
# "image_grid_pinpoints": [
# [
# 336,
# 672
# ],
# [
# 672,
# 336
# ],
# [
# 672,
# 672
# ],
# [
# 1008,
# 336
# ],
# [
# 336,
# 1008
# ]
# ],
# Flattened:
# [
# 336, 672,
# 672, 336,
# 672, 672,
# 1008, 336,
# 336, 1008
# ]
# *
# */
if "image_grid_pinpoints" in v_hparams:
# flatten it
image_grid_pinpoints = []
for pinpoint in v_hparams["image_grid_pinpoints"]:
for p in pinpoint:
image_grid_pinpoints.append(p)
fout.add_array("clip.vision.image_grid_pinpoints", image_grid_pinpoints)
if "image_crop_resolution" in v_hparams:
fout.add_uint32("clip.vision.image_crop_resolution", v_hparams["image_crop_resolution"])
if "image_aspect_ratio" in v_hparams:
fout.add_string("clip.vision.image_aspect_ratio", v_hparams["image_aspect_ratio"])
if "image_split_resolution" in v_hparams:
fout.add_uint32("clip.vision.image_split_resolution", v_hparams["image_split_resolution"])
if "mm_patch_merge_type" in v_hparams:
fout.add_string("clip.vision.mm_patch_merge_type", v_hparams["mm_patch_merge_type"])
if "mm_projector_type" in v_hparams:
fout.add_string("clip.vision.mm_projector_type", v_hparams["mm_projector_type"])
if feature_layers:
fout.add_array("clip.vision.feature_layer", feature_layers)
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 # pyright: ignore[reportAttributeAccessIssue]
image_std = processor.image_processor.image_std if args.image_std is None or args.image_std == default_image_std else args.image_std # pyright: ignore[reportAttributeAccessIssue]
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 = v_hparams["hidden_act"] == "gelu"
fout.add_bool("clip.use_gelu", use_gelu)
if has_llava_projector:
# By default, we drop the last layer for llava projector
# models unless we have explicitly set vision feature layers
if feature_layers is None:
model.vision_model.encoder.layers.pop(-1)
else:
model.vision_model.encoder.layers = model.vision_model.encoder.layers[:max(feature_layers)]
projector = torch.load(args.llava_projector)
for name, data in projector.items():
name = get_tensor_name(name)
# pw and dw conv ndim==4
if data.ndim == 2 or data.ndim == 4:
data = data.squeeze().numpy().astype(np.float16)
else:
data = data.squeeze().numpy().astype(np.float32)
fout.add_tensor(name, data)
print("Projector tensors added\n")
state_dict = model.state_dict()
for name, data in state_dict.items():
if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_llava_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)

280
examples/mtmd/legacy-models/glmedge-convert-image-encoder-to-gguf.py Normal file
View File

@@ -0,0 +1,280 @@
import argparse
import os
import json
import re
import torch
import numpy as np
from gguf import *
TEXT = "clip.text"
VISION = "clip.vision"
from transformers import SiglipVisionModel, SiglipVisionConfig
def k(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_llava: bool) -> bool:
if name in (
"logit_scale",
"text_model.embeddings.position_ids",
"vision_model.embeddings.position_ids",
):
return True
if name in (
"vision_model.head.probe",
"vision_model.head.attention.in_proj_weight",
"vision_model.head.attention.in_proj_bias",
"vision_model.head.attention.out_proj.weight",
"vision_model.head.attention.out_proj.bias",
"vision_model.head.layernorm.weight",
"vision_model.head.layernorm.bias",
"vision_model.head.mlp.fc1.weight",
"vision_model.head.mlp.fc1.bias",
"vision_model.head.mlp.fc2.weight",
"vision_model.head.mlp.fc2.bias"
):
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("--llava-projector", help="Path to llava.projector file. If specified, save an image encoder for LLaVA models.")
ap.add_argument("--projector-type", help="Type of projector. Possible values: mlp, ldp, ldpv2", choices=["mlp", "ldp", "ldpv2","adapter"], default="adapter")
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 = [0.5, 0.5, 0.5]
default_image_std = [0.5, 0.5, 0.5]
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]
with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
config = json.load(f)
if args.clip_model_is_vision:
v_hparams = config
t_hparams = None
else:
v_hparams = config["vision_config"]
t_hparams = None
# 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
vision_config = SiglipVisionConfig(**v_hparams)
model = SiglipVisionModel(vision_config)
model.load_state_dict(torch.load(os.path.join(dir_model, "glm.clip")))
fname_middle = None
has_text_encoder = False
has_vision_encoder = True
has_glm_projector = True
if args.text_only:
fname_middle = "text-"
has_vision_encoder = False
elif args.llava_projector is not None:
fname_middle = "mmproj-"
has_text_encoder = False
has_glm_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_glm_projector", has_glm_projector)
fout.add_file_type(ftype)
model_name = config["_name_or_path"] if "_name_or_path" in config else os.path.basename(dir_model)
fout.add_name(model_name)
if has_glm_projector:
fout.add_description("image encoder for glm4v")
fout.add_string("clip.projector_type", "adapter")
else:
fout.add_description("two-tower CLIP model")
if has_text_encoder:
assert t_hparams is not None
assert tokens is not None
# text_model hparams
fout.add_uint32(k(KEY_CONTEXT_LENGTH, TEXT), t_hparams["max_position_embeddings"])
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, TEXT), t_hparams["hidden_size"])
fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, TEXT), t_hparams["intermediate_size"])
fout.add_uint32("clip.text.projection_dim", t_hparams.get("projection_dim", config["projection_dim"]))
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, TEXT), t_hparams["num_attention_heads"])
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, TEXT), t_hparams["layer_norm_eps"])
fout.add_uint32(k(KEY_BLOCK_COUNT, TEXT), t_hparams["num_hidden_layers"])
fout.add_token_list(tokens)
if has_vision_encoder:
# vision_model hparams
fout.add_uint32("clip.vision.image_size", v_hparams["image_size"])
fout.add_uint32("clip.vision.patch_size", v_hparams["patch_size"])
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), v_hparams["hidden_size"])
fout.add_uint32(k(KEY_FEED_FORWARD_LENGTH, VISION), v_hparams["intermediate_size"])
fout.add_uint32("clip.vision.projection_dim", 0)
fout.add_uint32(k(KEY_ATTENTION_HEAD_COUNT, VISION), v_hparams["num_attention_heads"])
fout.add_float32(k(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
fout.add_uint32(k(KEY_BLOCK_COUNT, VISION), v_hparams["num_hidden_layers"])
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)
fout.add_bool("clip.use_gelu", True)
if has_glm_projector:
# model.vision_model.encoder.layers.pop(-1) # pyright: ignore[reportAttributeAccessIssue]
projector = torch.load(args.llava_projector)
for name, data in projector.items():
name = get_tensor_name(name)
# pw and dw conv ndim==4
if data.ndim == 2 or data.ndim == 4:
data = data.squeeze().numpy().astype(np.float16)
else:
data = data.squeeze().numpy().astype(np.float32)
if name.startswith("vision."):
name=name.replace("vision.","")
fout.add_tensor(name, data)
print(f"Projector {name} - {data.dtype} - shape = {data.shape}")
# print(f"Projector {name} tensors added\n")
state_dict = model.state_dict() # pyright: ignore[reportAttributeAccessIssue]
for name, data in state_dict.items():
if should_skip_tensor(name, has_text_encoder, has_vision_encoder, has_glm_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"siglip {name} - {data.dtype} - shape = {data.shape}")
# 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)

33
examples/mtmd/legacy-models/glmedge-surgery.py Normal file
View File

@@ -0,0 +1,33 @@
import argparse
import os
import torch
from transformers import AutoModel
ap = argparse.ArgumentParser()
ap.add_argument("-m", "--model", help="Path to GLM 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("vision.adapter.")]
# 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}/glm.projector")
clip_tensors = [k for k, v in checkpoint.items() if k.startswith("vision.vit.model.vision_model.")]
if len(clip_tensors) > 0:
clip = {name.replace("vision.vit.model.", ""): checkpoint[name].float() for name in clip_tensors}
torch.save(clip, f"{args.model}/glm.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")
print("Done!")
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
print(f"Also, use {args.model}glm.projector to prepare a glm-encoder.gguf file.")

38
examples/mtmd/legacy-models/llava_surgery.py Normal file
View File

@@ -0,0 +1,38 @@
import argparse
import glob
import os
import torch
ap = argparse.ArgumentParser()
ap.add_argument("-m", "--model", help="Path to LLaVA v1.5 model")
args = ap.parse_args()
# find the model part that includes the the multimodal projector weights
path = sorted(glob.glob(f"{args.model}/pytorch_model*.bin"))[-1]
checkpoint = torch.load(path)
# get a list of mm tensor names
mm_tensors = [k for k, v in checkpoint.items() if k.startswith("model.mm_projector")]
# 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}/llava.projector")
# BakLLaVA models contain CLIP tensors in it
clip_tensors = [k for k, v in checkpoint.items() if k.startswith("model.vision_tower")]
if len(clip_tensors) > 0:
clip = {name.replace("vision_tower.vision_tower.", ""): checkpoint[name].float() for name in clip_tensors}
torch.save(clip, f"{args.model}/llava.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")
print("Done!")
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

180
examples/mtmd/legacy-models/llava_surgery_v2.py Normal file
View File

@@ -0,0 +1,180 @@
import argparse
import glob
import os
import torch
from safetensors import safe_open
from safetensors.torch import save_file
from typing import Any, ContextManager, cast
# Function to determine if file is a SafeTensor file
def is_safetensor_file(file_path):
return file_path.endswith('.safetensors')
# Unified loading function
def load_model(file_path):
if is_safetensor_file(file_path):
tensors = {}
with cast(ContextManager[Any], safe_open(file_path, framework="pt", device="cpu")) as f:
for key in f.keys():
tensors[key] = f.get_tensor(key).clone()
# output shape
print(f"{key} : {tensors[key].shape}")
return tensors, 'safetensor'
else:
return torch.load(file_path, map_location=torch.device('cpu')), 'pytorch'
# Unified saving function
def save_model(model, file_path, file_type):
if file_type == 'safetensor':
# safe_save(model, file_path)
save_file(model, file_path)
else:
torch.save(model, file_path)
# Helpers to match weight names from specific components or
# determine if a saved shard contains that component
def is_vision_tower(weight_name):
return (
weight_name.startswith("model.vision_tower") or
weight_name.startswith("vit.") or
weight_name.startswith("vision_tower")
)
def is_newline(weight_name):
return (
weight_name.startswith("model.image_newline") or
weight_name.startswith("image_newline")
)
def is_mm_projector(weight_name):
return (
weight_name.startswith("model.mm_projector") or
weight_name.startswith("vision_proj.") or
weight_name.startswith("multi_modal_projector")
)
def newline_criteria(checkpoint):
return any(is_newline(k) for k in checkpoint.keys())
def proj_criteria(checkpoint):
return any(is_mm_projector(k) for k in checkpoint.keys())
# Adapted function to clean vision tower from checkpoint
def clean_vision_tower_from_checkpoint(checkpoint_path):
checkpoint, file_type = load_model(checkpoint_path)
# file_type = 'pytorch'
model_path = os.path.dirname(checkpoint_path)
print(f"Searching for vision tower tensors in {checkpoint_path}")
clip_tensors = [k for k, v in checkpoint.items() if is_vision_tower(k)]
if len(clip_tensors) > 0:
print(f"Found {len(clip_tensors)} tensors to extract from {checkpoint_path}")
# Adapted for file type
clip_path = os.path.join(model_path, "llava.clip")
if os.path.exists(clip_path):
print(f"Loading existing llava.clip from {clip_path}")
existing_clip, _ = load_model(clip_path)
else:
print(f"Creating new llava.clip at {clip_path}")
existing_clip = {}
# Update existing_clip with new tensors, avoid duplicates
for name in clip_tensors:
simple_name = name[name.index('vision_model.'):] if 'vision_model.' in name else name
print(f"Adding {simple_name} to llava.clip")
if simple_name not in existing_clip:
existing_clip[simple_name] = checkpoint[name]
# Save the updated clip tensors back to llava.clip
save_model(existing_clip, clip_path, 'pytorch')
# Remove the tensors from the original checkpoint
for name in clip_tensors:
del checkpoint[name]
checkpoint_path = checkpoint_path
return True
return False
def find_relevant_checkpoints(checkpoint_paths, newline_criteria, projector):
newline_checkpoint_path = None
projector_checkpoint_path = None
for path in checkpoint_paths:
checkpoint, _ = load_model(path)
if newline_criteria(checkpoint) and newline_checkpoint_path is None:
newline_checkpoint_path = path
if projector(checkpoint):
projector_checkpoint_path = path
return newline_checkpoint_path, projector_checkpoint_path
# Command-line interface setup
ap = argparse.ArgumentParser()
ap.add_argument("-m", "--model", required=True, help="Path to LLaVA v1.5+ model")
ap.add_argument("-C", "--clean-vision-tower", action="store_true", help="Remove any vision tower from the model files")
args = ap.parse_args()
if args.clean_vision_tower:
# Generalized to handle both PyTorch and SafeTensors models
model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
# checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and path.startswith('pytorch')) or (path.endswith('.safetensors') and path.startswith('model'))]
checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
for projector_checkpoint_path in checkpoint_paths:
print(f"Cleaning {projector_checkpoint_path}")
if not clean_vision_tower_from_checkpoint(projector_checkpoint_path):
print(f"No vision tower found in {projector_checkpoint_path}")
# we break once none is found, so far all models append them at the end
# break
print("Done! All vision tower tensors are removed from the model files and stored in llava.clip file.")
# Now we look for the projector in the last checkpoint
model_files = sorted(glob.glob(f"{args.model}/*"), key=os.path.getmtime, reverse=True)
checkpoint_paths = [path for path in model_files if (path.endswith('.bin') and 'pytorch' in path.split('/')[-1].split('\\')[-1]) or (path.endswith('.safetensors') and 'model' in path.split('/')[-1].split('\\')[-1])]
# last_checkpoint_path = checkpoint_paths[0]
# first_checkpoint_path = checkpoint_paths[-1]
newline_checkpoint_path, projector_checkpoint_path = find_relevant_checkpoints(checkpoint_paths, newline_criteria, proj_criteria)
print(f"Taking projector from {projector_checkpoint_path}")
first_mm_tensors = []
first_checkpoint = None
if newline_checkpoint_path is not None:
print(f"Taking newline from {newline_checkpoint_path}")
first_checkpoint, file_type = load_model(newline_checkpoint_path)
first_mm_tensors = [k for k, v in first_checkpoint.items() if is_newline(k)]
# Load the checkpoint
mm_tensors = []
last_checkpoint = None
if projector_checkpoint_path is not None:
last_checkpoint, file_type = load_model(projector_checkpoint_path)
mm_tensors = [k for k, v in last_checkpoint.items() if is_mm_projector(k)]
if len(mm_tensors) == 0:
if last_checkpoint is not None:
for k, v in last_checkpoint.items():
print(k)
print(f"Found {len(mm_tensors)} tensors to extract out of {len(last_checkpoint) if last_checkpoint is not None else 0} tensors.")
print("No tensors found. Is this a LLaVA model?")
exit()
print(f"Found {len(mm_tensors)} tensors to extract.")
print(f"Found additional {len(first_mm_tensors)} tensors to extract.")
# projector = {name: checkpoint.[name].float() for name in mm_tensors}
projector = {}
for name in mm_tensors:
assert last_checkpoint is not None
projector[name] = last_checkpoint[name].float()
for name in first_mm_tensors:
assert first_checkpoint is not None
projector[name] = first_checkpoint[name].float()
if len(projector) > 0:
save_model(projector, f"{args.model}/llava.projector", 'pytorch')
print("Done!")
print(f"Now you can convert {args.model} to a regular LLaMA GGUF file.")
print(f"Also, use {args.model}/llava.projector to prepare a llava-encoder.gguf file.")

885
examples/mtmd/legacy-models/minicpmv-convert-image-encoder-to-gguf.py Normal file
View File

@@ -0,0 +1,885 @@
# coding=utf-8
# Copyright 2024 Google AI and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" PyTorch Siglip model. """
# Copied from HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit and add tgt_sizes
import os
import math
import warnings
import numpy as np
import torch
import torch.nn.functional as F
from torch import nn
from torch.nn.init import _calculate_fan_in_and_fan_out
from transformers.activations import ACT2FN
from transformers.modeling_utils import PreTrainedModel
from transformers.configuration_utils import PretrainedConfig
from transformers.utils import (
logging,
)
from transformers.utils import logging
logger = logging.get_logger(__name__)
class SiglipVisionConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`SiglipVisionModel`]. It is used to instantiate a
Siglip vision encoder according to the specified arguments, defining the model architecture. Instantiating a
configuration with the defaults will yield a similar configuration to that of the vision encoder of the Siglip
[google/siglip-base-patch16-224](https://huggingface.co/google/siglip-base-patch16-224) architecture.
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
hidden_size (`int`, *optional*, defaults to 768):
Dimensionality of the encoder layers and the pooler layer.
intermediate_size (`int`, *optional*, defaults to 3072):
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
num_hidden_layers (`int`, *optional*, defaults to 12):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 12):
Number of attention heads for each attention layer in the Transformer encoder.
num_channels (`int`, *optional*, defaults to 3):
Number of channels in the input images.
image_size (`int`, *optional*, defaults to 224):
The size (resolution) of each image.
patch_size (`int`, *optional*, defaults to 16):
The size (resolution) of each patch.
hidden_act (`str` or `function`, *optional*, defaults to `"gelu_pytorch_tanh"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
layer_norm_eps (`float`, *optional*, defaults to 1e-06):
The epsilon used by the layer normalization layers.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
Example:
```python
>>> from transformers import SiglipVisionConfig, SiglipVisionModel
>>> # Initializing a SiglipVisionConfig with google/siglip-base-patch16-224 style configuration
>>> configuration = SiglipVisionConfig()
>>> # Initializing a SiglipVisionModel (with random weights) from the google/siglip-base-patch16-224 style configuration
>>> model = SiglipVisionModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "siglip_vision_model"
def __init__(
self,
hidden_size=768,
intermediate_size=3072,
num_hidden_layers=12,
num_attention_heads=12,
num_channels=3,
image_size=224,
patch_size=16,
hidden_act="gelu_pytorch_tanh",
layer_norm_eps=1e-6,
attention_dropout=0.0,
**kwargs,
):
super().__init__(**kwargs)
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.num_channels = num_channels
self.patch_size = patch_size
self.image_size = image_size
self.attention_dropout = attention_dropout
self.layer_norm_eps = layer_norm_eps
self.hidden_act = hidden_act
_CHECKPOINT_FOR_DOC = "google/siglip-base-patch16-224"
SIGLIP_PRETRAINED_MODEL_ARCHIVE_LIST = [
"google/siglip-base-patch16-224",
# See all SigLIP models at https://huggingface.co/models?filter=siglip
]
# Copied from transformers.models.llama.modeling_llama._get_unpad_data
def _get_unpad_data(attention_mask):
seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32)
indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten()
max_seqlen_in_batch = seqlens_in_batch.max().item()
cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.int32), (1, 0))
return (
indices,
cu_seqlens,
max_seqlen_in_batch,
)
def _trunc_normal_(tensor, mean, std, a, b):
# Cut & paste from PyTorch official master until it's in a few official releases - RW
# Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
def norm_cdf(x):
# Computes standard normal cumulative distribution function
return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
if (mean < a - 2 * std) or (mean > b + 2 * std):
warnings.warn(
"mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
"The distribution of values may be incorrect.",
stacklevel=2,
)
# Values are generated by using a truncated uniform distribution and
# then using the inverse CDF for the normal distribution.
# Get upper and lower cdf values
l = norm_cdf((a - mean) / std)
u = norm_cdf((b - mean) / std)
# Uniformly fill tensor with values from [l, u], then translate to
# [2l-1, 2u-1].
tensor.uniform_(2 * l - 1, 2 * u - 1)
# Use inverse cdf transform for normal distribution to get truncated
# standard normal
if tensor.dtype in [torch.float16, torch.bfloat16]:
# The `erfinv_` op is not (yet?) defined in float16+cpu, bfloat16+gpu
og_dtype = tensor.dtype
tensor = tensor.to(torch.float32)
tensor.erfinv_()
tensor = tensor.to(og_dtype)
else:
tensor.erfinv_()
# Transform to proper mean, std
tensor.mul_(std * math.sqrt(2.0))
tensor.add_(mean)
# Clamp to ensure it's in the proper range
if tensor.dtype == torch.float16:
# The `clamp_` op is not (yet?) defined in float16+cpu
tensor = tensor.to(torch.float32)
tensor.clamp_(min=a, max=b)
tensor = tensor.to(torch.float16)
else:
tensor.clamp_(min=a, max=b)
def trunc_normal_tf_(
tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
):
"""Fills the input Tensor with values drawn from a truncated
normal distribution. The values are effectively drawn from the
normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
with values outside :math:`[a, b]` redrawn until they are within
the bounds. The method used for generating the random values works
best when :math:`a \\leq \text{mean} \\leq b`.
NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
and the result is subsquently scaled and shifted by the mean and std args.
Args:
tensor: an n-dimensional `torch.Tensor`
mean: the mean of the normal distribution
std: the standard deviation of the normal distribution
a: the minimum cutoff value
b: the maximum cutoff value
"""
with torch.no_grad():
_trunc_normal_(tensor, 0, 1.0, a, b)
tensor.mul_(std).add_(mean)
def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
denom = fan_in
if mode == "fan_in":
denom = fan_in
elif mode == "fan_out":
denom = fan_out
elif mode == "fan_avg":
denom = (fan_in + fan_out) / 2
variance = scale / denom
if distribution == "truncated_normal":
# constant is stddev of standard normal truncated to (-2, 2)
trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
elif distribution == "normal":
with torch.no_grad():
tensor.normal_(std=math.sqrt(variance))
elif distribution == "uniform":
bound = math.sqrt(3 * variance)
with torch.no_grad():
tensor.uniform_(-bound, bound)
else:
raise ValueError(f"invalid distribution {distribution}")
def lecun_normal_(tensor):
variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
def default_flax_embed_init(tensor):
variance_scaling_(tensor, mode="fan_in", distribution="normal")
class SiglipVisionEmbeddings(nn.Module):
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.image_size = config.image_size
self.patch_size = config.patch_size
self.patch_embedding = nn.Conv2d(
in_channels=config.num_channels,
out_channels=self.embed_dim,
kernel_size=self.patch_size,
stride=self.patch_size,
padding="valid",
)
self.num_patches_per_side = self.image_size // self.patch_size
self.num_patches = self.num_patches_per_side**2
self.num_positions = self.num_patches
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
class SiglipAttention(nn.Module):
"""Multi-headed attention from 'Attention Is All You Need' paper"""
# Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
def __init__(self, config):
super().__init__()
self.config = config
self.embed_dim = config.hidden_size
self.num_heads = config.num_attention_heads
self.head_dim = self.embed_dim // self.num_heads
if self.head_dim * self.num_heads != self.embed_dim:
raise ValueError(
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
f" {self.num_heads})."
)
self.scale = self.head_dim**-0.5
self.dropout = config.attention_dropout
self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
class SiglipMLP(nn.Module):
def __init__(self, config):
super().__init__()
self.config = config
self.activation_fn = ACT2FN[config.hidden_act]
self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Siglip
class SiglipEncoderLayer(nn.Module):
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.embed_dim = config.hidden_size
self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
self.self_attn = (
SiglipAttention(config)
)
self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
self.mlp = SiglipMLP(config)
self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
class SiglipPreTrainedModel(PreTrainedModel):
"""
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
models.
"""
config_class = SiglipVisionConfig
base_model_prefix = "siglip"
supports_gradient_checkpointing = True
def _init_weights(self, module):
"""Initialize the weights"""
if isinstance(module, SiglipVisionEmbeddings):
width = self.config.hidden_size
nn.init.normal_(module.position_embedding.weight, std=1 / np.sqrt(width))
elif isinstance(module, nn.Embedding):
default_flax_embed_init(module.weight)
elif isinstance(module, SiglipAttention):
nn.init.normal_(module.q_proj.weight)
nn.init.normal_(module.k_proj.weight)
nn.init.normal_(module.v_proj.weight)
nn.init.normal_(module.out_proj.weight)
nn.init.zeros_(module.q_proj.bias)
nn.init.zeros_(module.k_proj.bias)
nn.init.zeros_(module.v_proj.bias)
nn.init.zeros_(module.out_proj.bias)
elif isinstance(module, SiglipMLP):
nn.init.normal_(module.fc1.weight)
nn.init.normal_(module.fc2.weight)
nn.init.normal_(module.fc1.bias, std=1e-6)
nn.init.normal_(module.fc2.bias, std=1e-6)
elif isinstance(module, (nn.Linear, nn.Conv2d)):
lecun_normal_(module.weight)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.LayerNorm):
module.bias.data.zero_()
module.weight.data.fill_(1.0)
SIGLIP_START_DOCSTRING = r"""
This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
etc.)
This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
and behavior.
Parameters:
config ([`SiglipVisionConfig`]): Model configuration class with all the parameters of the model.
Initializing with a config file does not load the weights associated with the model, only the
configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
"""
SIGLIP_VISION_INPUTS_DOCSTRING = r"""
Args:
pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
Pixel values. Padding will be ignored by default should you provide it. Pixel values can be obtained using
[`AutoImageProcessor`]. See [`CLIPImageProcessor.__call__`] for details.
output_attentions (`bool`, *optional*):
Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
tensors for more detail.
output_hidden_states (`bool`, *optional*):
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
more detail.
return_dict (`bool`, *optional*):
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
"""
# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Siglip
class SiglipEncoder(nn.Module):
"""
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
[`SiglipEncoderLayer`].
Args:
config: SiglipConfig
"""
def __init__(self, config: SiglipVisionConfig):
super().__init__()
self.config = config
self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
self.gradient_checkpointing = False
class SiglipVisionTransformer(SiglipPreTrainedModel):
config_class = SiglipVisionConfig
main_input_name = "pixel_values"
_supports_flash_attn_2 = True
def __init__(self, config: SiglipVisionConfig):
super().__init__(config)
self.config = config
embed_dim = config.hidden_size
self.embeddings = SiglipVisionEmbeddings(config)
self.encoder = SiglipEncoder(config)
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
self._use_flash_attention_2 = config._attn_implementation == "flash_attention_2"
# Initialize weights and apply final processing
self.post_init()
def get_input_embeddings(self) -> nn.Module:
return self.embeddings.patch_embedding
import argparse
import json
import re
import numpy as np
from gguf import *
from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer
from transformers.models.idefics2.configuration_idefics2 import 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 = [0.5, 0.5, 0.5]
default_image_std = [0.5, 0.5, 0.5]
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)
ap.add_argument('--minicpmv_version', type=int, help='minicpmv_version: MiniCPM-V-2 use 1; MiniCPM-V-2.5 use 2; MiniCPM-V-2.6 use 3; MiniCPM-o-2.6 use 4; MiniCPM-V 4.0 use 5; MiniCPM-o-4.0 use 6', default=2)
# 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
# Read config.json to get actual model configuration
config_path = os.path.join(dir_model, "config.json")
model_config = {}
if os.path.isfile(config_path):
with open(config_path, "r", encoding="utf-8") as f:
model_config = json.load(f)
print(f"Loaded config from {config_path}")
else:
print(f"Warning: config.json not found at {config_path}")
# If minicpmv_projector is not specified but the default path exists, use the default path
if args.minicpmv_projector is None:
default_projector_path = os.path.join(dir_model, "minicpmv.projector")
if os.path.isfile(default_projector_path):
args.minicpmv_projector = default_projector_path
print(f"Found default projector file: {default_projector_path}")
# If output_dir is not specified, use model_dir as the default value
if args.output_dir is None:
args.output_dir = dir_model
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)
minicpmv_version = args.minicpmv_version
# Use actual config values instead of hardcoded ones
if model_config:
# For the projector/resampler, use the main model's hidden_size
emb_dim = model_config.get("hidden_size", 1536)
# For the vision model, use vision_config values
vision_config_dict = model_config.get("vision_config", {})
default_vision_config = {
"hidden_size": vision_config_dict.get("hidden_size", 1152),
"image_size": vision_config_dict.get("image_size", 980),
"intermediate_size": vision_config_dict.get("intermediate_size", 4304),
"model_type": vision_config_dict.get("model_type", "siglip"),
"num_attention_heads": vision_config_dict.get("num_attention_heads", 16),
"num_hidden_layers": vision_config_dict.get("num_hidden_layers", 27),
"patch_size": vision_config_dict.get("patch_size", 14),
}
# Use vision model's num_hidden_layers for block_count
block_count = vision_config_dict.get("num_hidden_layers", 27)
print(f"Using config values: emb_dim={emb_dim}, block_count={block_count}")
print(f"Vision config: {default_vision_config}")
else:
# Fallback to original hardcoded logic if config.json not found
emb_dim = 4096
block_count = 26
if minicpmv_version == 1:
emb_dim = 2304
block_count = 26
elif minicpmv_version == 2:
emb_dim = 4096
block_count = 27
elif minicpmv_version == 3:
emb_dim = 3584
block_count = 27
elif minicpmv_version == 4:
emb_dim = 3584
block_count = 27
elif minicpmv_version == 5:
emb_dim = 2560
block_count = 27
elif minicpmv_version == 6:
emb_dim = 4096
block_count = 27
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)
if minicpmv_version == 3 or (model_config and model_config.get("vision_config", {}).get("model_type") == "siglip"):
vision_config = SiglipVisionConfig(**default_vision_config)
model = SiglipVisionTransformer(vision_config)
elif minicpmv_version == 4:
vision_config = SiglipVisionConfig(**default_vision_config)
model = SiglipVisionTransformer(vision_config)
elif minicpmv_version == 5:
default_vision_config["model_type"] = "siglip_vision_model"
vision_config = SiglipVisionConfig(**default_vision_config)
model = SiglipVisionTransformer(vision_config)
elif minicpmv_version == 6:
default_vision_config["model_type"] = "siglip_vision_model"
vision_config = SiglipVisionConfig(**default_vision_config)
model = SiglipVisionTransformer(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
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")
fout.add_int32("clip.minicpmv_version", minicpmv_version)
else:
fout.add_description("two-tower CLIP model")
if has_vision_encoder:
# vision_model hparams - use actual config values
vision_image_size = model_config.get("image_size", 448) if model_config else 448
vision_patch_size = default_vision_config.get("patch_size", 14)
vision_hidden_size = default_vision_config.get("hidden_size", 1152)
vision_intermediate_size = default_vision_config.get("intermediate_size", 4304)
vision_attention_heads = default_vision_config.get("num_attention_heads", 16)
fout.add_uint32("clip.vision.image_size", vision_image_size)
fout.add_uint32("clip.vision.patch_size", vision_patch_size)
fout.add_uint32(add_key_str(KEY_EMBEDDING_LENGTH, VISION), vision_hidden_size)
fout.add_uint32(add_key_str(KEY_FEED_FORWARD_LENGTH, VISION), vision_intermediate_size)
fout.add_uint32("clip.vision.projection_dim", 0)
fout.add_uint32(add_key_str(KEY_ATTENTION_HEAD_COUNT, VISION), vision_attention_heads)
fout.add_float32(add_key_str(KEY_ATTENTION_LAYERNORM_EPS, VISION), 1e-6)
fout.add_uint32(add_key_str(KEY_BLOCK_COUNT, VISION), block_count)
# Add MiniCPM-V specific parameters
query_num = model_config.get("query_num", 0) if model_config else 0
resampler_emb_dim = model_config.get("hidden_size", 0) if model_config else 0
fout.add_uint32("clip.minicpmv_query_num", query_num)
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(emb_dim, (70, 70))),
}
if re.match("resampler.proj", s):
return {
re.sub("proj", "pos_embed_k", s): torch.from_numpy(get_2d_sincos_pos_embed(emb_dim, (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/mtmd/legacy-models/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 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, torch_dtype=torch.bfloat16)
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}
if 'resampler.proj' in projector.keys() and hasattr(model.llm.config,'scale_emb') is True:
projector['resampler.proj'] = projector['resampler.proj'] / model.llm.config.scale_emb
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.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")
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.")

769
examples/mtmd/mtmd-audio.cpp Normal file
View File

@@ -0,0 +1,769 @@
#include "mtmd-audio.h"
#define _USE_MATH_DEFINES // for M_PI
#include <cmath>
#include <cstdint>
#include <cstring>
#include <thread>
#include <vector>
#include <fstream>
#include <algorithm>
// most of the code here is copied from whisper.cpp
// align x to upper multiple of n
#define _ALIGN(x, n) ((((x) + (n) - 1) / (n)) * (n))
namespace whisper_preprocessor {
#define SIN_COS_N_COUNT WHISPER_N_FFT
namespace {
struct whisper_global_cache {
// In FFT, we frequently use sine and cosine operations with the same values.
// We can use precalculated values to speed up the process.
float sin_vals[SIN_COS_N_COUNT];
float cos_vals[SIN_COS_N_COUNT];
// Hann window (Use cosf to eliminate difference)
// ref: https://pytorch.org/docs/stable/generated/torch.hann_window.html
// ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L147
float hann_window[WHISPER_N_FFT];
whisper_global_cache() {
fill_sin_cos_table();
fill_hann_window(sizeof(hann_window)/sizeof(hann_window[0]), true, hann_window);
}
void fill_sin_cos_table() {
for (int i = 0; i < SIN_COS_N_COUNT; i++) {
double theta = (2 * M_PI * i) / SIN_COS_N_COUNT;
sin_vals[i] = sinf(theta);
cos_vals[i] = cosf(theta);
}
}
void fill_hann_window(int length, bool periodic, float * output) {
int offset = -1;
if (periodic) {
offset = 0;
}
for (int i = 0; i < length; i++) {
output[i] = 0.5 * (1.0 - cosf((2.0 * M_PI * i) / (length + offset)));
}
}
} global_cache;
}
// naive Discrete Fourier Transform
// input is real-valued
// output is complex-valued
static void dft(const float* in, int N, float* out) {
const int sin_cos_step = SIN_COS_N_COUNT / N;
for (int k = 0; k < N; k++) {
float re = 0;
float im = 0;
for (int n = 0; n < N; n++) {
int idx = (k * n * sin_cos_step) % (SIN_COS_N_COUNT); // t = 2*M_PI*k*n/N
re += in[n]*global_cache.cos_vals[idx]; // cos(t)
im -= in[n]*global_cache.sin_vals[idx]; // sin(t)
}
out[k*2 + 0] = re;
out[k*2 + 1] = im;
}
}
// Cooley-Tukey FFT
// poor man's implementation - use something better
// input is real-valued
// output is complex-valued
static void fft(float* in, int N, float* out) {
if (N == 1) {
out[0] = in[0];
out[1] = 0;
return;
}
const int half_N = N / 2;
if (N - half_N*2 == 1) {
dft(in, N, out);
return;
}
float* even = in + N;
for (int i = 0; i < half_N; ++i) {
even[i]= in[2*i];
}
float* even_fft = out + 2 * N;
fft(even, half_N, even_fft);
float* odd = even;
for (int i = 0; i < half_N; ++i) {
odd[i] = in[2*i + 1];
}
float* odd_fft = even_fft + N;
fft(odd, half_N, odd_fft);
const int sin_cos_step = SIN_COS_N_COUNT / N;
for (int k = 0; k < half_N; k++) {
int idx = k * sin_cos_step; // t = 2*M_PI*k/N
float re = global_cache.cos_vals[idx]; // cos(t)
float im = -global_cache.sin_vals[idx]; // sin(t)
float re_odd = odd_fft[2*k + 0];
float im_odd = odd_fft[2*k + 1];
out[2*k + 0] = even_fft[2*k + 0] + re*re_odd - im*im_odd;
out[2*k + 1] = even_fft[2*k + 1] + re*im_odd + im*re_odd;
out[2*(k + half_N) + 0] = even_fft[2*k + 0] - re*re_odd + im*im_odd;
out[2*(k + half_N) + 1] = even_fft[2*k + 1] - re*im_odd - im*re_odd;
}
}
static void log_mel_spectrogram_worker_thread(int ith, const float * hann, const std::vector<float> & samples,
int n_samples, int frame_size, int frame_step, int n_threads,
const whisper_filters & filters, whisper_mel & mel) {
std::vector<float> fft_in(frame_size * 2, 0.0);
std::vector<float> fft_out(frame_size * 2 * 2 * 2);
int n_fft = filters.n_fft;
int i = ith;
// make sure n_fft == 1 + (WHISPER_N_FFT / 2), bin_0 to bin_nyquist
WHISPER_ASSERT(n_fft == 1 + (frame_size / 2));
// calculate FFT only when fft_in are not all zero
for (; i < std::min(n_samples / frame_step + 1, mel.n_len); i += n_threads) {
const int offset = i * frame_step;
// apply Hann window (~10% faster)
for (int j = 0; j < std::min(frame_size, n_samples - offset); j++) {
fft_in[j] = hann[j] * samples[offset + j];
}
// fill the rest with zeros
if (n_samples - offset < frame_size) {
std::fill(fft_in.begin() + (n_samples - offset), fft_in.end(), 0.0);
}
// FFT
fft(fft_in.data(), frame_size, fft_out.data());
// Calculate modulus^2 of complex numbers
// Use pow(fft_out[2 * j + 0], 2) + pow(fft_out[2 * j + 1], 2) causes inference quality problem? Interesting.
for (int j = 0; j < n_fft; j++) {
fft_out[j] = (fft_out[2 * j + 0] * fft_out[2 * j + 0] + fft_out[2 * j + 1] * fft_out[2 * j + 1]);
}
// mel spectrogram
for (int j = 0; j < mel.n_mel; j++) {
double sum = 0.0;
// unroll loop (suggested by GH user @lunixbochs)
int k = 0;
for (k = 0; k < n_fft - 3; k += 4) {
sum +=
fft_out[k + 0] * filters.data[j * n_fft + k + 0] +
fft_out[k + 1] * filters.data[j * n_fft + k + 1] +
fft_out[k + 2] * filters.data[j * n_fft + k + 2] +
fft_out[k + 3] * filters.data[j * n_fft + k + 3];
}
// handle n_fft remainder
for (; k < n_fft; k++) {
sum += fft_out[k] * filters.data[j * n_fft + k];
}
sum = log10(std::max(sum, 1e-10));
mel.data[j * mel.n_len + i] = sum;
}
}
// Otherwise fft_out are all zero
double sum = log10(1e-10);
for (; i < mel.n_len; i += n_threads) {
for (int j = 0; j < mel.n_mel; j++) {
mel.data[j * mel.n_len + i] = sum;
}
}
}
// ref: https://github.com/openai/whisper/blob/main/whisper/audio.py#L110-L157
static bool log_mel_spectrogram(
const float * samples,
const int n_samples,
const int /*sample_rate*/,
const int frame_size,
const int frame_step,
const int n_mel,
const int n_threads,
const whisper_filters & filters,
const bool debug,
whisper_mel & mel) {
//const int64_t t_start_us = ggml_time_us();
// Hann window
WHISPER_ASSERT(frame_size == WHISPER_N_FFT && "Unsupported frame_size");
const float * hann = global_cache.hann_window;
// Calculate the length of padding
int64_t stage_1_pad = WHISPER_SAMPLE_RATE * 30;
int64_t stage_2_pad = frame_size / 2;
// Initialize a vector and copy data from C array to it.
std::vector<float> samples_padded;
samples_padded.resize(n_samples + stage_1_pad + stage_2_pad * 2);
std::copy(samples, samples + n_samples, samples_padded.begin() + stage_2_pad);
// pad 30 seconds of zeros at the end of audio (480,000 samples) + reflective pad 200 samples at the end of audio
std::fill(samples_padded.begin() + n_samples + stage_2_pad, samples_padded.begin() + n_samples + stage_1_pad + 2 * stage_2_pad, 0);
// reflective pad 200 samples at the beginning of audio
std::reverse_copy(samples + 1, samples + 1 + stage_2_pad, samples_padded.begin());
mel.n_mel = n_mel;
// https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/SpectralOps.cpp#L936
// Calculate number of frames + remove the last frame
mel.n_len = (samples_padded.size() - frame_size) / frame_step;
// Calculate semi-padded sample length to ensure compatibility
mel.n_len_org = 1 + (n_samples + stage_2_pad - frame_size) / frame_step;
mel.data.resize(mel.n_mel * mel.n_len);
{
std::vector<std::thread> workers(n_threads - 1);
for (int iw = 0; iw < n_threads - 1; ++iw) {
workers[iw] = std::thread(
log_mel_spectrogram_worker_thread, iw + 1, hann, std::cref(samples_padded),
n_samples + stage_2_pad, frame_size, frame_step, n_threads,
std::cref(filters), std::ref(mel));
}
// main thread
log_mel_spectrogram_worker_thread(0, hann, samples_padded, n_samples + stage_2_pad, frame_size, frame_step, n_threads, filters, mel);
for (int iw = 0; iw < n_threads - 1; ++iw) {
workers[iw].join();
}
}
// clamping and normalization
double mmax = -1e20;
for (int i = 0; i < mel.n_mel*mel.n_len; i++) {
if (mel.data[i] > mmax) {
mmax = mel.data[i];
}
}
mmax -= 8.0;
for (int i = 0; i < mel.n_mel*mel.n_len; i++) {
if (mel.data[i] < mmax) {
mel.data[i] = mmax;
}
mel.data[i] = (mel.data[i] + 4.0)/4.0;
}
// Dump log_mel_spectrogram
if (debug) {
std::ofstream outFile("log_mel_spectrogram.json");
outFile << "[";
for (uint64_t i = 0; i < mel.data.size() - 1; i++) {
outFile << mel.data[i] << ", ";
}
outFile << mel.data[mel.data.size() - 1] << "]";
outFile.close();
}
return true;
}
bool preprocess_audio(
const float * samples,
size_t n_samples,
const whisper_filters & filters,
std::vector<whisper_mel> & output) {
if (n_samples == 0) {
// empty audio
return false;
}
whisper_mel out_full;
bool ok = log_mel_spectrogram(
samples,
n_samples,
COMMON_SAMPLE_RATE,
WHISPER_N_FFT,
WHISPER_HOP_LENGTH,
filters.n_mel,
4, // n_threads
filters,
false, // debug
out_full);
if (!ok) {
return false;
}
// because the cgraph in clip.cpp only accepts 3000 frames each, we need to split the mel
// we always expect the mel to have 3000 silent frames at the end
// printf("n_len %d\n", out_full.n_len);
const size_t frames_per_chunk = 3000;
GGML_ASSERT((size_t)out_full.n_len > frames_per_chunk);
for (size_t off = 0; off < (size_t)out_full.n_len; off += frames_per_chunk) {
int n_len = std::min(frames_per_chunk, (size_t)out_full.n_len - off);
if ((size_t)n_len < frames_per_chunk) {
break; // last uncomplete chunk will always be a padded chunk, safe to ignore
}
whisper_mel out_chunk;
out_chunk.n_len = n_len;
out_chunk.n_mel = out_full.n_mel;
out_chunk.n_len_org = out_full.n_mel; // unused
out_chunk.data.reserve(out_chunk.n_mel * out_chunk.n_len);
for (int i = 0; i < out_full.n_mel; i++) {
auto src = out_full.data.begin() + i*out_full.n_len + off;
out_chunk.data.insert(out_chunk.data.end(), src, src + frames_per_chunk);
}
output.push_back(std::move(out_chunk));
}
return true;
}
} // namespace whisper_preprocessor
// precalculated mel filter banks
// values are multiplied by 1000.0 to save space, and will be divided by 1000.0 in the end of the function
//
// generated from python code:
//
// from numpy import load
// data = load('mel_filters.npz')
// lst = data.files
// for item in lst:
// print(item)
// print(data[item].shape)
// n_mel = data[item].shape[0]
// n_fft = data[item].shape[1]
// for i, row in enumerate(data[item]):
// for j, val in enumerate(row):
// val = val * 1000.0
// if val != 0:
// print(f"data[{i*n_fft + j}] = {val:.6f};")
namespace whisper_precalc_filters {
whisper_preprocessor::whisper_filters get_128_bins() {
whisper_preprocessor::whisper_filters filters;
filters.n_mel = 128;
filters.n_fft = 201;
std::vector data(filters.n_mel * filters.n_fft, 0.0f);
data[1] = 12.37398665;
data[202] = 30.39256483;
data[404] = 24.74797331;
data[605] = 18.01857911;
data[807] = 37.12195903;
data[1008] = 5.64459199;
data[1009] = 6.72939420;
data[1210] = 36.03715822;
data[1412] = 19.10337992;
data[1613] = 23.66316877;
data[1815] = 31.47736564;
data[2016] = 11.28918398;
data[2017] = 1.08480197;
data[2218] = 41.68175161;
data[2420] = 13.45878839;
data[2621] = 29.30776216;
data[2823] = 25.83277412;
data[3024] = 16.93377644;
data[3226] = 38.20675984;
data[3427] = 4.55979025;
data[3428] = 7.81419594;
data[3629] = 34.95235741;
data[3831] = 20.18818259;
data[4032] = 22.57836796;
data[4234] = 32.56217018;
data[4435] = 10.20438317;
data[4436] = 2.16960395;
data[4637] = 40.59694707;
data[4839] = 14.54358920;
data[5040] = 28.22295949;
data[5242] = 26.91757679;
data[5443] = 15.84897563;
data[5645] = 39.29156065;
data[5846] = 3.47498828;
data[5847] = 8.89899861;
data[6048] = 33.86755288;
data[6250] = 21.27298526;
data[6451] = 21.49356715;
data[6653] = 33.64697099;
data[6854] = 9.11958050;
data[6855] = 3.25440569;
data[7056] = 39.51214626;
data[7258] = 15.62839188;
data[7459] = 27.13815868;
data[7661] = 28.00237760;
data[7862] = 14.76417296;
data[8064] = 40.37636518;
data[8265] = 2.38068704;
data[8266] = 10.20263787;
data[8467] = 31.61146119;
data[8669] = 24.54700135;
data[8870] = 15.32919332;
data[8871] = 1.66583748;
data[9072] = 36.72905266;
data[9274] = 20.09709924;
data[9475] = 16.93102531;
data[9476] = 2.90265540;
data[9677] = 32.84499049;
data[9879] = 23.52004871;
data[10080] = 11.03894413;
data[10081] = 10.72582975;
data[10282] = 22.71829173;
data[10484] = 32.27872774;
data[10685] = 0.11626833;
data[10686] = 22.85348251;
data[10887] = 8.56344029;
data[10888] = 14.97978810;
data[11089] = 15.51398356;
data[11090] = 8.51490628;
data[11291] = 21.10680379;
data[11292] = 3.32652032;
data[11493] = 25.47064796;
data[11695] = 27.35907957;
data[11896] = 0.65853616;
data[11897] = 23.83812517;
data[12098] = 3.44359246;
data[12099] = 21.22455277;
data[12300] = 5.35842171;
data[12301] = 19.42555793;
data[12502] = 6.49324711;
data[12503] = 18.35542172;
data[12704] = 6.93138083;
data[12705] = 17.93504693;
data[12906] = 6.74968259;
data[12907] = 18.09151843;
data[13108] = 6.01899112;
data[13109] = 18.75767298;
data[13310] = 4.80452832;
data[13311] = 19.87172849;
data[13512] = 3.16627859;
data[13513] = 21.37690969;
data[13514] = 1.25317345;
data[13714] = 1.15934468;
data[13715] = 20.80361731;
data[13716] = 4.04486805;
data[13917] = 17.55363122;
data[13918] = 7.08320038;
data[14119] = 14.07538634;
data[14120] = 10.32655034;
data[14321] = 10.40921453;
data[14322] = 13.73696327;
data[14523] = 6.59187697;
data[14524] = 17.27988198;
data[14525] = 1.46804214;
data[14725] = 2.65681883;
data[14726] = 18.09193194;
data[14727] = 5.85655728;
data[14928] = 13.34277913;
data[14929] = 10.28267574;
data[15130] = 8.56800377;
data[15131] = 14.72230814;
data[15132] = 1.04039861;
data[15332] = 3.79085587;
data[15333] = 17.14678481;
data[15334] = 6.11609267;
data[15535] = 11.75929047;
data[15536] = 11.13393717;
data[15737] = 6.43857848;
data[15738] = 16.07806236;
data[15739] = 4.23917221;
data[15939] = 1.19989377;
data[15940] = 12.75671553;
data[15941] = 9.65298992;
data[16142] = 7.06935255;
data[16143] = 14.94054683;
data[16144] = 4.19024844;
data[16344] = 1.51483389;
data[16345] = 12.00899947;
data[16346] = 9.84823331;
data[16547] = 6.10224018;
data[16548] = 15.33857174;
data[16549] = 5.57676842;
data[16749] = 0.36827257;
data[16750] = 9.89749376;
data[16751] = 11.35340426;
data[16752] = 2.05122307;
data[16952] = 3.89297144;
data[16953] = 12.97352277;
data[16954] = 8.06631614;
data[17155] = 6.74493238;
data[17156] = 13.85874674;
data[17157] = 5.41190524;
data[17357] = 0.74220158;
data[17358] = 8.98779090;
data[17359] = 11.37871388;
data[17360] = 3.32958088;
data[17560] = 2.82313535;
data[17561] = 10.68049297;
data[17562] = 9.43340641;
data[17563] = 1.76325557;
data[17763] = 4.39018616;
data[17764] = 11.87758986;
data[17765] = 7.97005836;
data[17766] = 0.66104700;
data[17966] = 5.49466675;
data[17967] = 12.62953598;
data[17968] = 6.93987962;
data[18169] = 6.18401915;
data[18170] = 12.93473132;
data[18171] = 6.29778765;
data[18371] = 0.02325210;
data[18372] = 6.50206627;
data[18373] = 12.32661773;
data[18374] = 6.00216538;
data[18574] = 0.31548753;
data[18575] = 6.48925547;
data[18576] = 12.04130240;
data[18577] = 6.01462880;
data[18777] = 0.29979556;
data[18778] = 6.18288014;
data[18779] = 12.04272825;
data[18780] = 6.29981188;
data[18781] = 0.55689598;
data[18980] = 0.01120471;
data[18981] = 5.61729167;
data[18982] = 11.22337859;
data[18983] = 6.82516303;
data[18984] = 1.35264499;
data[19184] = 4.82410006;
data[19185] = 10.16623247;
data[19186] = 7.56075513;
data[19187] = 2.34590308;
data[19387] = 3.83235747;
data[19388] = 8.92296247;
data[19389] = 8.47910438;
data[19390] = 3.50978645;
data[19590] = 2.66873185;
data[19591] = 7.51965167;
data[19592] = 9.55500547;
data[19593] = 4.81966138;
data[19594] = 0.08431751;
data[19793] = 1.35767367;
data[19794] = 5.98019501;
data[19795] = 10.60271543;
data[19796] = 6.25298498;
data[19797] = 1.74059917;
data[19997] = 4.32644226;
data[19998] = 8.73131864;
data[19999] = 7.78916525;
data[20000] = 3.48923868;
data[20200] = 2.57835095;
data[20201] = 6.77582854;
data[20202] = 9.40941647;
data[20203] = 5.31194592;
data[20204] = 1.21447595;
data[20403] = 0.75411191;
data[20404] = 4.75395704;
data[20405] = 8.75380263;
data[20406] = 7.19209015;
data[20407] = 3.28754401;
data[20607] = 2.68179690;
data[20608] = 6.49331464;
data[20609] = 9.11457930;
data[20610] = 5.39387390;
data[20611] = 1.67316827;
data[20810] = 0.57394296;
data[20811] = 4.20600036;
data[20812] = 7.83805829;
data[20813] = 7.52023002;
data[20814] = 3.97470826;
data[20815] = 0.42918732;
data[21014] = 1.90464477;
data[21015] = 5.36569161;
data[21016] = 8.82673822;
data[21017] = 6.27609482;
data[21018] = 2.89750961;
data[21218] = 2.89885257;
data[21219] = 6.19694078;
data[21220] = 8.56699049;
data[21221] = 5.34748193;
data[21222] = 2.12797290;
data[21421] = 0.44750227;
data[21422] = 3.59030394;
data[21423] = 6.73310598;
data[21424] = 7.77023612;
data[21425] = 4.70231380;
data[21426] = 1.63439126;
data[21625] = 1.01536023;
data[21626] = 4.01018746;
data[21627] = 7.00501446;
data[21628] = 7.23442994;
data[21629] = 4.31095669;
data[21630] = 1.38748321;
data[21829] = 1.33348850;
data[21830] = 4.18730825;
data[21831] = 7.04112789;
data[21832] = 6.93188375;
data[21833] = 4.14605811;
data[21834] = 1.36023236;
data[22033] = 1.42879714;
data[22034] = 4.14824858;
data[22035] = 6.86769979;
data[22036] = 6.83705276;
data[22037] = 4.18239459;
data[22038] = 1.52773573;
data[22237] = 1.32610439;
data[22238] = 3.91751388;
data[22239] = 6.50892360;
data[22240] = 6.92639686;
data[22241] = 4.39672917;
data[22242] = 1.86706171;
data[22441] = 1.04827771;
data[22442] = 3.51767405;
data[22443] = 5.98707050;
data[22444] = 7.17824046;
data[22445] = 4.76767914;
data[22446] = 2.35711760;
data[22645] = 0.61636406;
data[22646] = 2.96949223;
data[22647] = 5.32262027;
data[22648] = 7.57265091;
data[22649] = 5.27558755;
data[22650] = 2.97852419;
data[22651] = 0.68146095;
data[22849] = 0.04971400;
data[22850] = 2.29204819;
data[22851] = 4.53438237;
data[22852] = 6.77671656;
data[22853] = 5.90240723;
data[22854] = 3.71349836;
data[22855] = 1.52458926;
data[23054] = 1.50285335;
data[23055] = 3.63961048;
data[23056] = 5.77636715;
data[23057] = 6.63159089;
data[23058] = 4.54574358;
data[23059] = 2.45989650;
data[23060] = 0.37404924;
data[23258] = 0.61795861;
data[23259] = 2.65410915;
data[23260] = 4.69025923;
data[23261] = 6.72641024;
data[23262] = 5.46034705;
data[23263] = 3.47270933;
data[23264] = 1.48507138;
data[23463] = 1.59233576;
data[23464] = 3.53261665;
data[23465] = 5.47289755;
data[23466] = 6.44368259;
data[23467] = 4.54962999;
data[23468] = 2.65557761;
data[23469] = 0.76152512;
data[23667] = 0.46749352;
data[23668] = 2.31641904;
data[23669] = 4.16534441;
data[23670] = 6.01426978;
data[23671] = 5.67844696;
data[23672] = 3.87357362;
data[23673] = 2.06870004;
data[23674] = 0.26382666;
data[23872] = 1.05349103;
data[23873] = 2.81536230;
data[23874] = 4.57723346;
data[23875] = 6.33910485;
data[23876] = 5.12815686;
data[23877] = 3.40826320;
data[23878] = 1.68837002;
data[24077] = 1.43350090;
data[24078] = 3.11241671;
data[24079] = 4.79133241;
data[24080] = 6.40943693;
data[24081] = 4.77052201;
data[24082] = 3.13160778;
data[24083] = 1.49269309;
data[24281] = 0.02932359;
data[24282] = 1.62918994;
data[24283] = 3.22905602;
data[24284] = 4.82892245;
data[24285] = 6.14671456;
data[24286] = 4.58496623;
data[24287] = 3.02321767;
data[24288] = 1.46146910;
data[24486] = 0.13601698;
data[24487] = 1.66055572;
data[24488] = 3.18509457;
data[24489] = 4.70963307;
data[24490] = 6.04072399;
data[24491] = 4.55250870;
data[24492] = 3.06429295;
data[24493] = 1.57607743;
data[24494] = 0.08786193;
data[24691] = 0.09328097;
data[24692] = 1.54603878;
data[24693] = 2.99879676;
data[24694] = 4.45155473;
data[24695] = 5.90431225;
data[24696] = 4.65566106;
data[24697] = 3.23751615;
data[24698] = 1.81937125;
data[24699] = 0.40122634;
data[24897] = 1.30262633;
data[24898] = 2.68698297;
data[24899] = 4.07133950;
data[24900] = 5.45569602;
data[24901] = 4.87832492;
data[24902] = 3.52695142;
data[24903] = 2.17557792;
data[24904] = 0.82420459;
data[25102] = 0.94595028;
data[25103] = 2.26512621;
data[25104] = 3.58430226;
data[25105] = 4.90347855;
data[25106] = 5.20569785;
data[25107] = 3.91795207;
data[25108] = 2.63020652;
data[25109] = 1.34246063;
data[25110] = 0.05471494;
data[25307] = 0.49037894;
data[25308] = 1.74744334;
data[25309] = 3.00450763;
data[25310] = 4.26157191;
data[25311] = 5.51863620;
data[25312] = 4.39707236;
data[25313] = 3.16995848;
data[25314] = 1.94284460;
data[25315] = 0.71573065;
data[25513] = 1.14698056;
data[25514] = 2.34485767;
data[25515] = 3.54273478;
data[25516] = 4.74061165;
data[25517] = 4.95198462;
data[25518] = 3.78264743;
data[25519] = 2.61331047;
data[25520] = 1.44397374;
data[25521] = 0.27463681;
data[25718] = 0.47569509;
data[25719] = 1.61717169;
data[25720] = 2.75864848;
data[25721] = 3.90012516;
data[25722] = 5.04160160;
data[25723] = 4.45712078;
data[25724] = 3.34284059;
data[25725] = 2.22856039;
data[25726] = 1.11428020;
for (auto & val : data) {
val /= 1000.0f;
}
filters.data = std::move(data);
return filters;
}
} // namespace whisper_precalc_filters

47
examples/mtmd/mtmd-audio.h Normal file
View File

@@ -0,0 +1,47 @@
#pragma once
#include "ggml.h"
#include <cstdint>
#include <vector>
#include <string>
#define WHISPER_ASSERT GGML_ASSERT
#define WHISPER_SAMPLE_RATE 16000
#define WHISPER_N_FFT 400
#define WHISPER_HOP_LENGTH 160
#define WHISPER_CHUNK_SIZE 30
#define COMMON_SAMPLE_RATE 16000
namespace whisper_preprocessor {
struct whisper_mel {
int n_len;
int n_len_org;
int n_mel;
std::vector<float> data;
};
struct whisper_filters {
int32_t n_mel;
int32_t n_fft;
std::vector<float> data;
};
bool preprocess_audio(
const float * samples,
size_t n_samples,
const whisper_filters & filters,
std::vector<whisper_mel> & output);
} // namespace whisper_preprocessor
namespace whisper_precalc_filters {
whisper_preprocessor::whisper_filters get_128_bins();
} // namespace whisper_precalc_filters

445
examples/mtmd/mtmd-cli.cpp Normal file
View File

@@ -0,0 +1,445 @@
//#include "arg.h"
#include "log.h"
#include "common.h"
#include "sampling.h"
#include "llama.h"
#include "ggml.h"
#include "console.h"
#include "chat.h"
#include "mtmd.h"
#include "mtmd-helper.h"
#include <vector>
#include <limits.h>
#include <cinttypes>
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
#include <signal.h>
#include <unistd.h>
#elif defined (_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#include <signal.h>
#endif
// volatile, because of signal being an interrupt
static volatile bool g_is_generating = false;
static volatile bool g_is_interrupted = false;
/**
* Please note that this is NOT a production-ready stuff.
* It is a playground for trying multimodal support in llama.cpp.
* For contributors: please keep this code simple and easy to understand.
*/
static void show_additional_info(int /*argc*/, char ** argv) {
LOG_TEE(
"Experimental CLI for multimodal\n\n"
"Usage: %s [options] -m <model> --mmproj <mmproj> --image <image> --audio <audio> -p <prompt>\n\n"
" -m and --mmproj are required\n"
" -hf user/repo can replace both -m and --mmproj in most cases\n"
" --image, --audio and -p are optional, if NOT provided, the CLI will run in chat mode\n"
" to disable using GPU for mmproj model, add --no-mmproj-offload\n",
argv[0]
);
}
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__)) || defined (_WIN32)
static void sigint_handler(int signo) {
if (signo == SIGINT) {
if (g_is_generating) {
g_is_generating = false;
} else {
console::cleanup();
if (g_is_interrupted) {
_exit(1);
}
g_is_interrupted = true;
}
}
}
#endif
// ======================= compat ================================
using common_init_result = llama_init_result;
using common_sampler = llama_sampling_context;
using llama_tokens = std::vector<llama_token>;
using common_params = gpt_params;
inline common_init_result common_init_from_params(gpt_params & params) {
return llama_init_from_gpt_params(params);
}
inline llama_sampling_context * common_sampler_init(const llama_model * model, const llama_sampling_params & sparams) {
return llama_sampling_init(llama_get_model_vocab(model), sparams);
}
inline std::vector<llama_token> common_tokenize(const llama_context * ctx, const std::string & text, bool add_special, bool parse_special = false) {
return llama_tokenize(ctx, text, add_special, parse_special);
}
inline void common_sampler_free(common_sampler * smpl) {
llama_sampling_free(smpl);
}
inline llama_token common_sampler_sample(common_sampler * gsmpl, llama_context * ctx, int idx, [[maybe_unused]] bool grammar_first = false) {
return llama_sampling_sample(gsmpl, ctx, nullptr, idx);
}
inline void common_sampler_accept(common_sampler * gsmpl, llama_context * ctx, llama_token token, bool accept_grammar) {
llama_sampling_accept(gsmpl, ctx, token, accept_grammar);
}
inline std::string common_token_to_piece(const llama_context * ctx, llama_token token, bool special = true) {
return llama_token_to_piece(ctx, token, special);
}
inline void common_batch_clear(llama_batch & batch) {
llama_batch_clear(batch);
}
inline void common_batch_add(llama_batch & batch, llama_token id, llama_pos pos, const std::vector<llama_seq_id> & seq_ids, bool logits) {
llama_batch_add(batch, id, pos, seq_ids, logits);
}
void common_init() {
#ifdef NDEBUG
const char * build_type = "";
#else
const char * build_type = " (debug)";
#endif
LOG_TEE("build: %d (%s) with %s for %s%s\n", LLAMA_BUILD_NUMBER, LLAMA_COMMIT, LLAMA_COMPILER, LLAMA_BUILD_TARGET, build_type);
}
#ifndef LOG_ERR
#define LOG_ERR LOG_TEE
#endif
#ifndef LOG_INF
#define LOG_INF LOG_TEE
#endif
#ifndef LOG_DBG
#define LOG_DBG LOG
#endif
// ======================= end compat ================================
struct mtmd_cli_context {
mtmd::context_ptr ctx_vision;
common_init_result llama_init;
llama_model * model;
llama_context * lctx;
const llama_vocab * vocab;
common_sampler * smpl;
llama_batch batch;
int n_batch;
mtmd::bitmaps bitmaps;
// note: we know that gemma3 template is "linear", meaning each turn is completely separated to another
// so here we don't need to keep track of chat history
common_chat_templates_ptr tmpls;
// support for legacy templates (models not having EOT token)
llama_tokens antiprompt_tokens;
int n_threads = 1;
llama_pos n_past = 0;
mtmd_cli_context(common_params & params) : llama_init(common_init_from_params(params)) {
model = llama_init.model; //.get();
lctx = llama_init.context; //.get();
vocab = llama_model_get_vocab(model);
smpl = common_sampler_init(model, params.sparams); //sampling);
n_threads = params.n_threads;
batch = llama_batch_init(1, 0, 1); // batch for next token generation
n_batch = params.n_batch;
if (!model || !lctx) {
exit(1);
}
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");
LOG_ERR(" For Mistral Small 3.1, use '--chat-template mistral-v7'\n");
exit(1);
}
tmpls = common_chat_templates_init(model, params.chat_template);
LOG_TEE("%s: chat template example:\n%s\n", __func__, common_chat_format_example(tmpls.get(), params.use_jinja, params.default_template_kwargs).c_str());
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);
}
}
~mtmd_cli_context() {
llama_batch_free(batch);
common_sampler_free(smpl);
}
void init_vision_context(common_params & params) {
const char * clip_path = params.mmproj.path.c_str();
mtmd_context_params mparams = mtmd_context_params_default();
mparams.use_gpu = params.mmproj_use_gpu;
mparams.print_timings = true;
mparams.n_threads = params.n_threads;
mparams.verbosity = params.verbosity > 0 ? GGML_LOG_LEVEL_DEBUG : GGML_LOG_LEVEL_INFO;
ctx_vision.reset(mtmd_init_from_file(clip_path, model, mparams));
if (!ctx_vision.get()) {
LOG_ERR("Failed to load vision model from %s\n", clip_path);
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()
);
}
bool load_media(const std::string & fname) {
mtmd::bitmap bmp(mtmd_helper_bitmap_init_from_file(ctx_vision.get(), fname.c_str()));
if (!bmp.ptr) {
return false;
}
bitmaps.entries.push_back(std::move(bmp));
return true;
}
};
static int generate_response(mtmd_cli_context & ctx, int n_predict) {
llama_tokens generated_tokens;
for (int i = 0; i < n_predict; i++) {
if (i > n_predict || !g_is_generating || g_is_interrupted) {
LOG_TEE("\n");
break;
}
llama_token token_id = common_sampler_sample(ctx.smpl, ctx.lctx, -1);
generated_tokens.push_back(token_id);
common_sampler_accept(ctx.smpl, ctx.lctx, token_id, true);
if (llama_vocab_is_eog(ctx.vocab, token_id) || ctx.check_antiprompt(generated_tokens)) {
LOG_TEE("\n");
break; // end of generation
}
LOG_TEE("%s", common_token_to_piece(ctx.lctx, token_id).c_str());
fflush(stdout);
if (g_is_interrupted) {
LOG_TEE("\n");
break;
}
// eval the token
common_batch_clear(ctx.batch);
common_batch_add(ctx.batch, token_id, ctx.n_past++, {0}, true);
if (llama_decode(ctx.lctx, ctx.batch)) {
LOG_ERR("failed to decode token\n");
return 1;
}
}
return 0;
}
static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, bool add_bos = false) {
common_chat_templates_inputs tmpl_inputs;
tmpl_inputs.messages = {msg};
tmpl_inputs.add_generation_prompt = true;
tmpl_inputs.use_jinja = false; // jinja is buggy here
auto formatted_chat = common_chat_templates_apply(ctx.tmpls.get(), tmpl_inputs);
LOG_DBG("formatted_chat.prompt: %s\n", formatted_chat.prompt.c_str());
mtmd_input_text text;
text.text = formatted_chat.prompt.c_str();
text.add_special = add_bos;
text.parse_special = true;
if (g_is_interrupted) return 0;
mtmd::input_chunks chunks(mtmd_input_chunks_init());
auto bitmaps_c_ptr = ctx.bitmaps.c_ptr();
int32_t res = mtmd_tokenize(ctx.ctx_vision.get(),
chunks.ptr.get(), // output
&text, // text
bitmaps_c_ptr.data(),
bitmaps_c_ptr.size());
if (res != 0) {
LOG_ERR("Unable to tokenize prompt, res = %d\n", res);
return 1;
}
ctx.bitmaps.entries.clear();
llama_pos new_n_past;
if (mtmd_helper_eval_chunks(ctx.ctx_vision.get(),
ctx.lctx, // lctx
chunks.ptr.get(), // chunks
ctx.n_past, // n_past
0, // seq_id
ctx.n_batch, // n_batch
true, // logits_last
&new_n_past)) {
LOG_ERR("Unable to eval prompt\n");
return 1;
}
ctx.n_past = new_n_past;
LOG("\n");
return 0;
}
int main(int argc, char ** argv) {
ggml_time_init();
common_params params;
params.sparams.temp = 0.2; // lower temp by default for better quality
if (!gpt_params_parse(argc, argv, params)) {
return 1;
}
//if (!common_params_parse(argc, argv, params, LLAMA_EXAMPLE_MTMD, show_additional_info)) {
// return 1;
//}
common_init();
if (params.mmproj.path.empty()) {
show_additional_info(argc, argv);
LOG_ERR("ERR: Missing --mmproj argument\n");
return 1;
}
mtmd_cli_context ctx(params);
LOG("%s: loading model: %s\n", __func__, params.model.c_str());
bool is_single_turn = !params.prompt.empty() && !params.image.empty();
int n_predict = params.n_predict < 0 ? INT_MAX : params.n_predict;
// Ctrl+C handling
{
#if defined (__unix__) || (defined (__APPLE__) && defined (__MACH__))
struct sigaction sigint_action;
sigint_action.sa_handler = sigint_handler;
sigemptyset (&sigint_action.sa_mask);
sigint_action.sa_flags = 0;
sigaction(SIGINT, &sigint_action, NULL);
#elif defined (_WIN32)
auto console_ctrl_handler = +[](DWORD ctrl_type) -> BOOL {
return (ctrl_type == CTRL_C_EVENT) ? (sigint_handler(SIGINT), true) : false;
};
SetConsoleCtrlHandler(reinterpret_cast<PHANDLER_ROUTINE>(console_ctrl_handler), true);
#endif
}
if (g_is_interrupted) return 130;
if (is_single_turn) {
g_is_generating = true;
if (params.prompt.find(mtmd_default_marker()) == std::string::npos) {
for (size_t i = 0; i < params.image.size(); i++) {
params.prompt += mtmd_default_marker();
}
}
common_chat_msg msg;
msg.role = "user";
msg.content = params.prompt;
for (const auto & image : params.image) {
if (!ctx.load_media(image)) {
return 1; // error is already printed by libmtmd
}
}
if (eval_message(ctx, msg, true)) {
return 1;
}
if (!g_is_interrupted && generate_response(ctx, n_predict)) {
return 1;
}
} else {
LOG_TEE("\n Running in chat mode, available commands:");
if (mtmd_support_vision(ctx.ctx_vision.get())) {
LOG_TEE("\n /image <path> load an image");
}
if (mtmd_support_audio(ctx.ctx_vision.get())) {
LOG_TEE("\n /audio <path> load an audio");
}
LOG_TEE("\n /clear clear the chat history");
LOG_TEE("\n /quit or /exit exit the program");
LOG_TEE("\n");
bool is_first_msg = true;
std::string content;
while (!g_is_interrupted) {
g_is_generating = false;
LOG_TEE("\n> ");
console::set_display(console::user_input);
std::string line;
console::readline(line, false);
if (g_is_interrupted) break;
console::set_display(console::reset);
line = string_strip(line);
if (line.empty()) {
continue;
}
if (line == "/quit" || line == "/exit") {
break;
}
if (line == "/clear") {
ctx.n_past = 0;
llama_kv_cache_seq_rm(ctx.lctx, 0, 1, -1);
//llama_memory_seq_rm(llama_get_memory(ctx.lctx), 0, 1, -1); // keep BOS
LOG_TEE("Chat history cleared\n\n");
continue;
}
g_is_generating = true;
bool is_image = line == "/image" || line.find("/image ") == 0;
bool is_audio = line == "/audio" || line.find("/audio ") == 0;
if (is_image || is_audio) {
if (line.size() < 8) {
LOG_ERR("ERR: Missing media filename\n");
continue;
}
std::string media_path = line.substr(7);
if (ctx.load_media(media_path)) {
LOG_TEE("%s %s loaded\n", media_path.c_str(), is_image ? "image" : "audio");
content += mtmd_default_marker();
}
// else, error is already printed by libmtmd
continue;
} else {
content += line;
}
common_chat_msg msg;
msg.role = "user";
msg.content = content;
int ret = eval_message(ctx, msg, is_first_msg);
if (ret) {
return 1;
}
if (g_is_interrupted) break;
if (generate_response(ctx, n_predict)) {
return 1;
}
content.clear();
is_first_msg = false;
}
}
if (g_is_interrupted) LOG_TEE("\nInterrupted by user\n");
LOG_TEE("\n\n");
llama_print_timings(ctx.lctx);
//llama_perf_context_print(ctx.lctx);
return g_is_interrupted ? 130 : 0;
}

460
examples/mtmd/mtmd-helper.cpp Normal file
View File

@@ -0,0 +1,460 @@
// fix problem with std::min and std::max
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#ifndef NOMINMAX
# define NOMINMAX
#endif
#include <windows.h>
#endif
#include "mtmd.h"
#include "mtmd-helper.h"
#include "llama.h"
#include <algorithm>
#include <cinttypes>
#include <vector>
//#define MTMD_AUDIO_DEBUG
#define MINIAUDIO_IMPLEMENTATION
#ifndef MTMD_AUDIO_DEBUG
# define MA_NO_ENCODING
#endif
#define MA_NO_DEVICE_IO
#define MA_NO_RESOURCE_MANAGER
#define MA_NO_NODE_GRAPH
#define MA_NO_ENGINE
#define MA_NO_GENERATION
#define MA_API static
#include "miniaudio/miniaudio.h"
#define STB_IMAGE_IMPLEMENTATION
#include "stb/stb_image.h"
#define LOG_INF(...) fprintf(stdout, __VA_ARGS__)
#define LOG_ERR(...) fprintf(stderr, __VA_ARGS__)
size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks) {
size_t n_tokens = 0;
for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
auto chunk = mtmd_input_chunks_get(chunks, i);
n_tokens += mtmd_input_chunk_get_n_tokens(chunk);
}
return n_tokens;
}
llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks) {
llama_pos n_pos = 0;
for (size_t i = 0; i < mtmd_input_chunks_size(chunks); i++) {
auto chunk = mtmd_input_chunks_get(chunks, i);
n_pos += mtmd_input_chunk_get_n_pos(chunk);
}
return n_pos;
}
// helper struct to make working with embd batch easier
// note: this will be removed after llama_batch_ext refactoring
struct decode_embd_batch {
int n_pos_per_embd;
int n_mmproj_embd;
std::vector<llama_pos> pos;
std::vector<llama_pos> pos_view; // used by mrope
std::vector<int32_t> n_seq_id;
std::vector<llama_seq_id> seq_id_0;
std::vector<llama_seq_id *> seq_ids;
std::vector<int8_t> logits;
llama_batch batch;
decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
pos .resize(n_tokens * n_pos_per_embd);
n_seq_id.resize(n_tokens);
seq_ids .resize(n_tokens + 1);
logits .resize(n_tokens);
seq_id_0.resize(1);
seq_ids [n_tokens] = nullptr;
batch = {
/*n_tokens =*/ n_tokens,
/*tokens =*/ nullptr,
/*embd =*/ embd,
/*pos =*/ pos.data(),
/*n_seq_id =*/ n_seq_id.data(),
/*seq_id =*/ seq_ids.data(),
/*logits =*/ logits.data(),
};
}
void set_position_normal(llama_pos pos_0, llama_seq_id seq_id) {
seq_id_0[0] = seq_id;
for (int i = 0; i < batch.n_tokens; i++) {
batch.pos [i] = pos_0 + i;
batch.n_seq_id[i] = 1;
batch.seq_id [i] = seq_id_0.data();
batch.logits [i] = false;
}
}
// M-RoPE for image
void set_position_mrope_2d(llama_pos pos_0, int nx, int ny, llama_seq_id seq_id) {
GGML_ASSERT(n_pos_per_embd == 4);
seq_id_0[0] = seq_id;
for (int y = 0; y < ny; y++) {
for (int x = 0; x < nx; x++) {
int i = y * nx + x;
pos[i ] = pos_0;
pos[i + batch.n_tokens ] = pos_0 + y;
pos[i + batch.n_tokens * 2] = pos_0 + x;
pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
}
}
for (int i = 0; i < batch.n_tokens; i++) {
batch.n_seq_id[i] = 1;
batch.seq_id [i] = seq_id_0.data();
batch.logits [i] = false;
}
}
// M-RoPE for audio
void set_position_mrope_1d(llama_pos pos_0, llama_seq_id seq_id) {
GGML_ASSERT(n_pos_per_embd == 4);
seq_id_0[0] = seq_id;
for (int i = 0; i < batch.n_tokens; i++) {
pos[i ] = pos_0 + i;
pos[i + batch.n_tokens ] = pos_0 + i;
pos[i + batch.n_tokens * 2] = pos_0 + i;
pos[i + batch.n_tokens * 3] = 0; // last pos dim is unused
}
for (int i = 0; i < batch.n_tokens; i++) {
batch.n_seq_id[i] = 1;
batch.seq_id [i] = seq_id_0.data();
batch.logits [i] = false;
}
}
llama_batch get_view(int offset, int n_tokens) {
llama_pos * pos_ptr;
pos_view.clear();
pos_view.reserve(n_tokens * n_pos_per_embd);
if (n_pos_per_embd > 1) {
// mrope
// for example, with layout of src: 1234...1234...1234...1234...
// offset 2 will give us dst: 34...34...34...34...
for (int i = 0; i < n_pos_per_embd; i++) {
// assume n_tokens is less than or equal to batch.n_tokens
// batch.n_tokens is number of **total** tokens
// n_tokens is number of viewed token
size_t src_idx = i * batch.n_tokens + offset;
pos_view.insert(pos_view.end(),
pos.data() + src_idx,
pos.data() + src_idx + n_tokens);
}
pos_ptr = pos_view.data();
} else {
// normal
pos_ptr = pos.data() + offset;
}
return {
/*n_tokens =*/ n_tokens,
/*tokens =*/ nullptr,
/*embd =*/ batch.embd + offset * n_mmproj_embd,
/*pos =*/ pos_ptr,
/*n_seq_id =*/ batch.n_seq_id + offset,
/*seq_id =*/ batch.seq_id + offset,
/*logits =*/ batch.logits + offset,
};
}
};
// Helper function for decoding an image whose embeddings have already been calculated
int32_t mtmd_helper_decode_image_chunk(
mtmd_context * ctx,
struct llama_context * lctx,
const mtmd_input_chunk * chunk,
float * encoded_embd,
llama_pos n_past,
llama_seq_id seq_id,
int32_t n_batch,
llama_pos * new_n_past) {
auto chunk_type = mtmd_input_chunk_get_type(chunk);
const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
LOG_ERR("failed to decode chunk: input chunk not of image/audio type\n");
return -1;
}
const llama_model * model = llama_get_model(lctx);
int n_mmproj_embd = llama_model_n_embd(model);
int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
int32_t n_tokens = mtmd_input_chunk_get_n_tokens(chunk);
int32_t i_batch = 0;
int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
decode_embd_batch batch_embd(encoded_embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
if (mtmd_decode_use_mrope(ctx)) {
if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
const auto image_tokens = mtmd_input_chunk_get_tokens_image(chunk);
if (!image_tokens) {
LOG_ERR("failed to decode chunk: image tokens are null\n");
return -1;
}
const int nx = mtmd_image_tokens_get_nx(image_tokens);
const int ny = mtmd_image_tokens_get_ny(image_tokens);
batch_embd.set_position_mrope_2d(n_past, nx, ny, seq_id);
} else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
batch_embd.set_position_mrope_1d(n_past, seq_id);
} else {
GGML_ABORT("invalid chunk type for M-RoPE");
}
} else {
batch_embd.set_position_normal(n_past, seq_id);
}
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
int pos_offset = i_batch*n_batch;
int n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
llama_batch batch_embd_view = batch_embd.get_view(pos_offset, n_tokens_batch);
LOG_INF("decoding %s batch %d/%d, n_tokens_batch = %d\n", name, i_batch+1, n_img_batches, n_tokens_batch);
int64_t t1 = ggml_time_ms();
int32_t ret = llama_decode(lctx, batch_embd_view);
if (ret != 0) {
LOG_ERR("failed to decode %s\n", name);
llama_set_causal_attn(lctx, true); // restore causal attn
return ret;
}
LOG_INF("%s decoded (batch %d/%d) in %" PRId64 " ms\n", name, i_batch+1, n_img_batches, ggml_time_ms() - t1);
i_batch++;
}
n_past += mtmd_input_chunk_get_n_pos(chunk);
*new_n_past = n_past;
if (mtmd_decode_use_non_causal(ctx)) {
llama_set_causal_attn(lctx, true);
}
return 0;
}
int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
struct llama_context * lctx,
const mtmd_input_chunk * chunk,
llama_pos n_past,
llama_seq_id seq_id,
int32_t n_batch,
bool logits_last,
llama_pos * new_n_past) {
int32_t ret;
llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
auto chunk_type = mtmd_input_chunk_get_type(chunk);
if (chunk_type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
size_t n_tokens;
const auto tokens = mtmd_input_chunk_get_tokens_text(chunk, &n_tokens);
// LOG_INF("decoding text chunk, n_tokens = %zu\n", n_tokens);
size_t i = 0;
while (i < n_tokens) { // split into batches
text_batch.n_tokens = 0; // clear the batch
for (; i < n_tokens && text_batch.n_tokens < n_batch; i++) {
int32_t j = text_batch.n_tokens;
text_batch.token [j] = tokens[i];
text_batch.pos [j] = n_past++;
text_batch.n_seq_id[j] = 1;
text_batch.seq_id [j][0] = seq_id;
text_batch.logits [j] = false;
text_batch.n_tokens++;
}
bool is_last_token = (i == n_tokens);
if (logits_last && is_last_token) {
text_batch.logits[text_batch.n_tokens - 1] = true;
}
ret = llama_decode(lctx, text_batch);
if (ret != 0) {
LOG_ERR("failed to decode text\n");
llama_batch_free(text_batch);
return ret;
}
*new_n_past += text_batch.n_tokens;
}
} else if (chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE || chunk_type == MTMD_INPUT_CHUNK_TYPE_AUDIO) {
const char * name = chunk_type == MTMD_INPUT_CHUNK_TYPE_IMAGE ? "image" : "audio";
int64_t t0 = ggml_time_ms();
LOG_INF("encoding %s slice...\n", name);
ret = mtmd_encode_chunk(ctx, chunk);
if (ret != 0) {
LOG_ERR("failed to encode %s slice\n", name);
llama_batch_free(text_batch);
return ret;
}
LOG_INF("%s slice encoded in %" PRId64 " ms\n", name, ggml_time_ms() - t0);
float * embd = mtmd_get_output_embd(ctx);
ret = mtmd_helper_decode_image_chunk(ctx, lctx, chunk, embd, n_past, seq_id, n_batch, new_n_past);
if (ret != 0) {
LOG_ERR("failed to decode %s\n", name);
llama_batch_free(text_batch);
return ret;
}
} else {
GGML_ABORT("chunk type not supported");
}
llama_batch_free(text_batch);
return 0;
}
int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
struct llama_context * lctx,
const mtmd_input_chunks * chunks,
llama_pos n_past,
llama_seq_id seq_id,
int32_t n_batch,
bool logits_last,
llama_pos * new_n_past) {
size_t n_chunks = mtmd_input_chunks_size(chunks);
if (n_chunks == 0) {
LOG_ERR("no chunks to eval\n");
return 0;
}
for (size_t i = 0; i < n_chunks; i++) {
bool chunk_logits_last = (i == n_chunks - 1) && logits_last;
auto chunk = mtmd_input_chunks_get(chunks, i);
int32_t res = mtmd_helper_eval_chunk_single(ctx, lctx, chunk, n_past, seq_id, n_batch, chunk_logits_last, &n_past);
if (res != 0) {
LOG_ERR("failed to eval chunk %zu\n", i);
return res;
}
*new_n_past = n_past;
}
return 0;
}
namespace audio_helpers {
static bool is_audio_file(const char * buf, size_t len) {
if (len < 12) {
return false;
}
// RIFF ref: https://en.wikipedia.org/wiki/Resource_Interchange_File_Format
// WAV ref: https://www.mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
bool is_wav = memcmp(buf, "RIFF", 4) == 0 && memcmp(buf + 8, "WAVE", 4) == 0;
bool is_mp3 = len >= 3 && (
memcmp(buf, "ID3", 3) == 0 ||
// Check for MPEG sync word (simplified check)
((unsigned char)buf[0] == 0xFF && ((unsigned char)buf[1] & 0xE0) == 0xE0)
);
bool is_flac = memcmp(buf, "fLaC", 4) == 0;
return is_wav || is_mp3 || is_flac;
}
// returns true if the buffer is a valid audio file
static bool decode_audio_from_buf(const unsigned char * buf_in, size_t len, int target_sampler_rate, std::vector<float> & pcmf32_mono) {
ma_result result;
const int channels = 1;
ma_decoder_config decoder_config = ma_decoder_config_init(ma_format_f32, channels, target_sampler_rate);
ma_decoder decoder;
result = ma_decoder_init_memory(buf_in, len, &decoder_config, &decoder);
if (result != MA_SUCCESS) {
return false;
}
ma_uint64 frame_count;
ma_uint64 frames_read;
result = ma_decoder_get_length_in_pcm_frames(&decoder, &frame_count);
if (result != MA_SUCCESS) {
ma_decoder_uninit(&decoder);
return false;
}
pcmf32_mono.resize(frame_count);
result = ma_decoder_read_pcm_frames(&decoder, pcmf32_mono.data(), frame_count, &frames_read);
if (result != MA_SUCCESS) {
ma_decoder_uninit(&decoder);
return false;
}
#ifdef MTMD_AUDIO_DEBUG
// save audio to wav file
ma_encoder_config config = ma_encoder_config_init(ma_encoding_format_wav, ma_format_f32, 1, target_sampler_rate);
ma_encoder encoder;
ma_encoder_init_file("output.wav", &config, &encoder);
ma_encoder_write_pcm_frames(&encoder, pcmf32_mono.data(), pcmf32_mono.size(), &frames_read);
ma_encoder_uninit(&encoder);
#endif
ma_decoder_uninit(&decoder);
return true;
}
} // namespace audio_helpers
mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len) {
if (audio_helpers::is_audio_file((const char *)buf, len)) {
std::vector<float> pcmf32;
int bitrate = mtmd_get_audio_bitrate(ctx);
if (bitrate < 0) {
LOG_ERR("This model does not support audio input\n");
return nullptr;
}
if (!audio_helpers::decode_audio_from_buf(buf, len, bitrate, pcmf32)) {
LOG_ERR("Unable to read WAV audio file from buffer\n");
return nullptr;
}
return mtmd_bitmap_init_from_audio(pcmf32.size(), pcmf32.data());
}
// otherwise, we assume it's an image
mtmd_bitmap * result = nullptr;
{
int nx, ny, nc;
auto * data = stbi_load_from_memory(buf, len, &nx, &ny, &nc, 3);
if (!data) {
LOG_ERR("%s: failed to decode image bytes\n", __func__);
return nullptr;
}
result = mtmd_bitmap_init(nx, ny, data);
stbi_image_free(data);
}
return result;
}
mtmd_bitmap * mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname) {
std::vector<unsigned char> buf;
FILE * f = fopen(fname, "rb");
if (!f) {
LOG_ERR("Unable to open file %s: %s\n", fname, strerror(errno));
return nullptr;
}
fseek(f, 0, SEEK_END);
long file_size = ftell(f);
fseek(f, 0, SEEK_SET);
buf.resize(file_size);
size_t n_read = fread(buf.data(), 1, file_size, f);
fclose(f);
if (n_read != (size_t)file_size) {
LOG_ERR("Failed to read entire file %s", fname);
return nullptr;
}
return mtmd_helper_bitmap_init_from_buf(ctx, buf.data(), buf.size());
}

91
examples/mtmd/mtmd-helper.h Normal file
View File

@@ -0,0 +1,91 @@
#ifndef MTMD_HELPER_H
#define MTMD_HELPER_H
#include "ggml.h"
#include "llama.h"
#include "mtmd.h"
#include <stddef.h>
#include <stdint.h>
#include <stdbool.h>
#ifdef __cplusplus
extern "C" {
#endif
//
// libmtmd helper functions
//
// Please note that these helpers are not guaranteed to be stable.
// BREAKING CHANGES are expected.
//
// helper function to construct a mtmd_bitmap from a file
// it calls mtmd_helper_bitmap_init_from_buf() internally
// returns nullptr on failure
// this function is thread-safe
MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_file(mtmd_context * ctx, const char * fname);
// helper function to construct a mtmd_bitmap from a buffer containing a file
// supported formats:
// image: formats supported by stb_image: jpg, png, bmp, gif, etc.
// audio: formats supported by miniaudio: wav, mp3, flac
// note: audio files will be auto-detected based on magic bytes
// returns nullptr on failure
// this function is thread-safe
MTMD_API mtmd_bitmap * mtmd_helper_bitmap_init_from_buf(mtmd_context * ctx, const unsigned char * buf, size_t len);
// helper to count the total number of tokens from a list of chunks, useful to keep track of KV cache
MTMD_API size_t mtmd_helper_get_n_tokens(const mtmd_input_chunks * chunks);
// helper to count the total position of tokens from a list of chunks, useful to keep track of n_past
// normally, n_pos is equal to n_tokens, but for M-RoPE it is different
MTMD_API llama_pos mtmd_helper_get_n_pos(const mtmd_input_chunks * chunks);
// helper function that automatically:
// 1. run llama_decode() on text chunks
// 2. run mtmd_encode() on image chunks, then mtmd_get_output_embd() and then llama_decode()
// if any of the mtmd_encode() or llama_decode() calls return non-zero, stop and forward the error
// otherwise, returns 0 on success
// this function is NOT thread-safe
MTMD_API int32_t mtmd_helper_eval_chunks(mtmd_context * ctx,
struct llama_context * lctx,
const mtmd_input_chunks * chunks,
llama_pos n_past,
llama_seq_id seq_id,
int32_t n_batch,
bool logits_last,
llama_pos * new_n_past);
// works like mtmd_helper_eval_chunks(), but only for a single chunk
// this function is NOT thread-safe
MTMD_API int32_t mtmd_helper_eval_chunk_single(mtmd_context * ctx,
struct llama_context * lctx,
const mtmd_input_chunk * chunk,
llama_pos n_past,
llama_seq_id seq_id,
int32_t n_batch,
bool logits_last,
llama_pos * new_n_past);
// helper function to decode an image whose embeddings have already been calculated
// this helper will handle batching and pre/post decoding setup (for ex. gemma 3 requires non-causal attention)
// ret 0 on success, -1 on chunk not being a valid image chunk, 1 on decode failure
MTMD_API int32_t mtmd_helper_decode_image_chunk(mtmd_context * ctx,
struct llama_context * lctx,
const mtmd_input_chunk * chunk,
float * encoded_embd,
llama_pos n_past,
llama_seq_id seq_id,
int32_t n_batch,
llama_pos * new_n_past);
#ifdef __cplusplus
} // extern "C"
#endif
//
// C++ wrappers
//
#endif

1066
examples/mtmd/mtmd.cpp Normal file
View File

File diff suppressed because it is too large Load Diff

298
examples/mtmd/mtmd.h Normal file
View File

@@ -0,0 +1,298 @@
#ifndef MTMD_H
#define MTMD_H
#include "ggml.h"
#include "llama.h"
#include <stddef.h>
#include <stdint.h>
#include <stdbool.h>
#ifdef __cplusplus
#include <string>
#include <vector>
#include <cinttypes>
#include <memory>
#endif
/**
* libmtmd: A library for multimodal support in llama.cpp.
*
* WARNING: This API is experimental and subject to many BREAKING CHANGES.
* Issues related to API usage may receive lower priority support.
*
* For the usage, see an example in mtmd-cli.cpp
*/
#ifdef LLAMA_SHARED
# if defined(_WIN32) && !defined(__MINGW32__)
# ifdef LLAMA_BUILD
# define MTMD_API __declspec(dllexport)
# else
# define MTMD_API __declspec(dllimport)
# endif
# else
# define MTMD_API __attribute__ ((visibility ("default")))
# endif
#else
# define MTMD_API
#endif
// deprecated marker, use mtmd_default_marker() instead
#define MTMD_DEFAULT_IMAGE_MARKER "<__image__>"
#ifdef __cplusplus
extern "C" {
#endif
enum mtmd_input_chunk_type {
MTMD_INPUT_CHUNK_TYPE_TEXT,
MTMD_INPUT_CHUNK_TYPE_IMAGE,
MTMD_INPUT_CHUNK_TYPE_AUDIO,
};
// opaque types
struct mtmd_context;
struct mtmd_bitmap;
struct mtmd_image_tokens;
struct mtmd_input_chunk;
struct mtmd_input_chunks;
struct mtmd_input_text {
const char * text;
bool add_special;
bool parse_special;
};
//
// C API
//
typedef struct mtmd_context mtmd_context;
typedef struct mtmd_bitmap mtmd_bitmap;
typedef struct mtmd_image_tokens mtmd_image_tokens;
typedef struct mtmd_input_chunk mtmd_input_chunk;
typedef struct mtmd_input_chunks mtmd_input_chunks;
typedef struct mtmd_input_text mtmd_input_text;
struct mtmd_context_params {
bool use_gpu;
bool print_timings;
int n_threads;
enum ggml_log_level verbosity;
const char * image_marker; // deprecated, use media_marker instead
const char * media_marker;
};
MTMD_API const char * mtmd_default_marker(void);
MTMD_API struct mtmd_context_params mtmd_context_params_default(void);
// initialize the mtmd context
// return nullptr on failure
MTMD_API mtmd_context * mtmd_init_from_file(const char * mmproj_fname,
const struct llama_model * text_model,
const struct mtmd_context_params ctx_params);
MTMD_API void mtmd_free(mtmd_context * ctx);
// whether we need to set non-causal mask before llama_decode
MTMD_API bool mtmd_decode_use_non_causal(mtmd_context * ctx);
// whether the current model use M-RoPE for llama_decode
MTMD_API bool mtmd_decode_use_mrope(mtmd_context * ctx);
// whether the current model supports vision input
MTMD_API bool mtmd_support_vision(mtmd_context * ctx);
// whether the current model supports audio input
MTMD_API bool mtmd_support_audio(mtmd_context * ctx);
// get audio bitrate in Hz, for example 16000 for Whisper
// return -1 if audio is not supported
MTMD_API int mtmd_get_audio_bitrate(mtmd_context * ctx);
// mtmd_bitmap
//
// if bitmap is image:
// length of data must be nx * ny * 3
// the data is in RGBRGBRGB... format
// if bitmap is audio:
// length of data must be n_samples * sizeof(float)
// the data is in float format (PCM F32)
MTMD_API mtmd_bitmap * mtmd_bitmap_init (uint32_t nx, uint32_t ny, const unsigned char * data);
MTMD_API mtmd_bitmap * mtmd_bitmap_init_from_audio(size_t n_samples, const float * data);
MTMD_API uint32_t mtmd_bitmap_get_nx (const mtmd_bitmap * bitmap);
MTMD_API uint32_t mtmd_bitmap_get_ny (const mtmd_bitmap * bitmap);
MTMD_API const unsigned char * mtmd_bitmap_get_data (const mtmd_bitmap * bitmap);
MTMD_API size_t mtmd_bitmap_get_n_bytes(const mtmd_bitmap * bitmap);
MTMD_API bool mtmd_bitmap_is_audio (const mtmd_bitmap * bitmap);
MTMD_API void mtmd_bitmap_free (mtmd_bitmap * bitmap);
// bitmap ID is optional, but useful for KV cache tracking
// these getters/setters are dedicated functions, so you can for example calculate the hash of the image based on mtmd_bitmap_get_data()
MTMD_API const char * mtmd_bitmap_get_id(const mtmd_bitmap * bitmap);
MTMD_API void mtmd_bitmap_set_id(mtmd_bitmap * bitmap, const char * id);
// mtmd_input_chunks
//
// this is simply a list of mtmd_input_chunk
// the elements can only be populated via mtmd_tokenize()
MTMD_API mtmd_input_chunks * mtmd_input_chunks_init(void);
MTMD_API size_t mtmd_input_chunks_size(const mtmd_input_chunks * chunks);
MTMD_API const mtmd_input_chunk * mtmd_input_chunks_get (const mtmd_input_chunks * chunks, size_t idx);
MTMD_API void mtmd_input_chunks_free(mtmd_input_chunks * chunks);
// mtmd_input_chunk
//
// the instance will be constructed via mtmd_tokenize()
// it will be freed along with mtmd_input_chunks
MTMD_API enum mtmd_input_chunk_type mtmd_input_chunk_get_type (const mtmd_input_chunk * chunk);
MTMD_API const llama_token * mtmd_input_chunk_get_tokens_text (const mtmd_input_chunk * chunk, size_t * n_tokens_output);
MTMD_API const mtmd_image_tokens * mtmd_input_chunk_get_tokens_image(const mtmd_input_chunk * chunk);
MTMD_API size_t mtmd_input_chunk_get_n_tokens (const mtmd_input_chunk * chunk);
// returns nullptr for ID on text chunk
MTMD_API const char * mtmd_input_chunk_get_id (const mtmd_input_chunk * chunk);
// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
MTMD_API llama_pos mtmd_input_chunk_get_n_pos (const mtmd_input_chunk * chunk);
// in case you want to use custom logic to handle the chunk (i.e. KV cache management)
// you can move the chunk ownership to your own code by copying it
// remember to free the chunk when you are done with it
MTMD_API mtmd_input_chunk * mtmd_input_chunk_copy(const mtmd_input_chunk * chunk);
MTMD_API void mtmd_input_chunk_free(mtmd_input_chunk * chunk);
// mtmd_image_tokens
//
// the instance will be constructed via mtmd_tokenize()
// it will be freed along with mtmd_input_chunk
MTMD_API size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * image_tokens); // TODO: deprecate
MTMD_API size_t mtmd_image_tokens_get_nx (const mtmd_image_tokens * image_tokens);
MTMD_API size_t mtmd_image_tokens_get_ny (const mtmd_image_tokens * image_tokens);
MTMD_API const char * mtmd_image_tokens_get_id (const mtmd_image_tokens * image_tokens); // TODO: deprecate
// number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
MTMD_API llama_pos mtmd_image_tokens_get_n_pos (const mtmd_image_tokens * image_tokens); // TODO: deprecate
// tokenize an input text prompt and a list of bitmaps (images/audio)
// the prompt must have the input image marker (default: "<__media__>") in it
// the default marker is defined by mtmd_default_marker()
// the marker will be replaced with the image/audio chunk
// for example:
// "here is an image: <__media__>\ndescribe it in detail."
// this will gives 3 chunks:
// 1. "here is an image: <start_of_image>"
// 2. (image/audio tokens)
// 3. "<end_of_image>\ndescribe it in detail."
// number of bitmaps must be equal to the number of markers in the prompt
// this function is thread-safe (shared ctx)
// return values:
// 0 on success
// 1 on number of bitmaps not matching the number of markers
// 2 on image preprocessing error
MTMD_API int32_t mtmd_tokenize(mtmd_context * ctx,
mtmd_input_chunks * output,
const mtmd_input_text * text,
const mtmd_bitmap ** bitmaps,
size_t n_bitmaps);
// returns 0 on success
// TODO: deprecate
MTMD_API int32_t mtmd_encode(mtmd_context * ctx,
const mtmd_image_tokens * image_tokens);
// returns 0 on success
MTMD_API int32_t mtmd_encode_chunk(mtmd_context * ctx,
const mtmd_input_chunk * chunk);
// get output embeddings from the last encode pass
// the reading size (in bytes) is equal to:
// llama_model_n_embd(model) * mtmd_input_chunk_get_n_tokens(chunk) * sizeof(float)
MTMD_API float * mtmd_get_output_embd(mtmd_context * ctx);
/////////////////////////////////////////
// test function, to be used in test-mtmd-c-api.c
MTMD_API mtmd_input_chunks * mtmd_test_create_input_chunks(void);
#ifdef __cplusplus
} // extern "C"
#endif
//
// C++ wrappers
//
#ifdef __cplusplus
namespace mtmd {
struct mtmd_context_deleter {
void operator()(mtmd_context * val) { mtmd_free(val); }
};
using context_ptr = std::unique_ptr<mtmd_context, mtmd_context_deleter>;
struct mtmd_bitmap_deleter {
void operator()(mtmd_bitmap * val) { mtmd_bitmap_free(val); }
};
using bitmap_ptr = std::unique_ptr<mtmd_bitmap, mtmd_bitmap_deleter>;
struct mtmd_input_chunks_deleter {
void operator()(mtmd_input_chunks * val) { mtmd_input_chunks_free(val); }
};
using input_chunks_ptr = std::unique_ptr<mtmd_input_chunks, mtmd_input_chunks_deleter>;
struct mtmd_input_chunk_deleter {
void operator()(mtmd_input_chunk * val) { mtmd_input_chunk_free(val); }
};
using input_chunk_ptr = std::unique_ptr<mtmd_input_chunk, mtmd_input_chunk_deleter>;
struct bitmap {
bitmap_ptr ptr;
bitmap() : ptr(nullptr) {}
bitmap(mtmd_bitmap * bitmap) : ptr(bitmap) {}
bitmap(bitmap && other) noexcept : ptr(std::move(other.ptr)) {}
bitmap(uint32_t nx, uint32_t ny, const unsigned char * data) {
ptr.reset(mtmd_bitmap_init(nx, ny, data));
}
~bitmap() = default;
uint32_t nx() { return mtmd_bitmap_get_nx(ptr.get()); }
uint32_t ny() { return mtmd_bitmap_get_ny(ptr.get()); }
const unsigned char * data() { return mtmd_bitmap_get_data(ptr.get()); }
size_t n_bytes() { return mtmd_bitmap_get_n_bytes(ptr.get()); }
std::string id() { return mtmd_bitmap_get_id(ptr.get()); }
void set_id(const char * id) { mtmd_bitmap_set_id(ptr.get(), id); }
};
struct bitmaps {
std::vector<bitmap> entries;
~bitmaps() = default;
// return list of pointers to mtmd_bitmap
// example:
// auto bitmaps_c_ptr = bitmaps.c_ptr();
// int32_t res = mtmd_tokenize(... bitmaps_c_ptr.data(), bitmaps_c_ptr.size());
std::vector<const mtmd_bitmap *> c_ptr() {
std::vector<const mtmd_bitmap *> res(entries.size());
for (size_t i = 0; i < entries.size(); i++) {
res[i] = entries[i].ptr.get();
}
return res;
}
};
struct input_chunks {
input_chunks_ptr ptr;
input_chunks() = default;
input_chunks(mtmd_input_chunks * chunks) : ptr(chunks) {}
~input_chunks() = default;
size_t size() { return mtmd_input_chunks_size(ptr.get()); }
const mtmd_input_chunk * operator[](size_t idx) {
return mtmd_input_chunks_get(ptr.get(), idx);
}
};
} // namespace mtmd
#endif
#endif

5
examples/mtmd/requirements.txt Normal file
View File

@@ -0,0 +1,5 @@
-r ../../requirements/requirements-convert_legacy_llama.txt
--extra-index-url https://download.pytorch.org/whl/cpu
pillow~=11.3.0
torch~=2.6.0
torchvision~=0.21.0

BIN
examples/mtmd/test-1.jpeg Normal file
View File

Binary file not shown.
Side by Side

After

Width:  |  Height:  |  Size: 121 KiB

BIN
examples/mtmd/test-2.mp3 Normal file
View File

Binary file not shown.

163
examples/mtmd/tests.sh Executable file
View File

@@ -0,0 +1,163 @@
#!/usr/bin/env bash
# make sure we are in the right directory
SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
cd $SCRIPT_DIR
#export LLAMA_CACHE="$SCRIPT_DIR/tmp"
set -eux
mkdir -p $SCRIPT_DIR/output
PROJ_ROOT="$SCRIPT_DIR/../.."
cd $PROJ_ROOT
# Check if the first argument is "big", then run test with big models
# This is useful if we're running the script on a larger machine, so we can test the big models
RUN_BIG_TESTS=false
if [ "${1:-}" = "big" ]; then
RUN_BIG_TESTS=true
echo "Include BIG models..."
fi
RUN_HUGE_TESTS=false
if [ "${1:-}" = "huge" ]; then
RUN_HUGE_TESTS=true
RUN_BIG_TESTS=true
echo "Include BIG and HUGE models..."
fi
###############
arr_prefix=()
arr_hf=()
arr_tmpl=() # chat template
arr_file=()
add_test_vision() {
local hf=$1
local tmpl=${2:-""} # default to empty string if not provided
arr_prefix+=("[vision]")
arr_hf+=("$hf")
arr_tmpl+=("$tmpl")
arr_file+=("test-1.jpeg")
}
add_test_audio() {
local hf=$1
arr_prefix+=("[audio] ")
arr_hf+=("$hf")
arr_tmpl+=("") # no need for chat tmpl
arr_file+=("test-2.mp3")
}
add_test_vision "ggml-org/SmolVLM-500M-Instruct-GGUF:Q8_0"
add_test_vision "ggml-org/SmolVLM2-2.2B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/SmolVLM2-500M-Video-Instruct-GGUF:Q8_0"
add_test_vision "ggml-org/gemma-3-4b-it-GGUF:Q4_K_M"
add_test_vision "THUDM/glm-edge-v-5b-gguf:Q4_K_M"
add_test_vision "second-state/Llava-v1.5-7B-GGUF:Q2_K" "vicuna"
add_test_vision "cjpais/llava-1.6-mistral-7b-gguf:Q3_K_M" "vicuna"
add_test_vision "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
add_test_vision "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted
add_test_vision "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
add_test_vision "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
add_test_vision "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/InternVL2_5-1B-GGUF:Q8_0"
add_test_vision "ggml-org/InternVL3-1B-Instruct-GGUF:Q8_0"
add_test_vision "ggml-org/Qwen2.5-Omni-3B-GGUF:Q4_K_M"
add_test_vision "ggml-org/LFM2-VL-450M-GGUF:Q8_0"
add_test_audio "ggml-org/ultravox-v0_5-llama-3_2-1b-GGUF:Q8_0"
add_test_audio "ggml-org/Qwen2.5-Omni-3B-GGUF:Q4_K_M"
add_test_audio "ggml-org/Voxtral-Mini-3B-2507-GGUF:Q4_K_M"
# to test the big models, run: ./tests.sh big
if [ "$RUN_BIG_TESTS" = true ]; then
add_test_vision "ggml-org/pixtral-12b-GGUF:Q4_K_M"
add_test_vision "ggml-org/Mistral-Small-3.1-24B-Instruct-2503-GGUF" "mistral-v7"
add_test_vision "ggml-org/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/Qwen2-VL-7B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/Qwen2.5-VL-7B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/InternVL3-8B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/InternVL3-14B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/Qwen2.5-Omni-7B-GGUF:Q4_K_M"
# add_test_vision "ggml-org/Qwen2.5-VL-32B-Instruct-GGUF:Q4_K_M" # does not work on my mac M3 Ultra
add_test_vision "ggml-org/Kimi-VL-A3B-Thinking-2506-GGUF:Q4_K_M"
add_test_audio "ggml-org/ultravox-v0_5-llama-3_1-8b-GGUF:Q4_K_M"
add_test_audio "ggml-org/Qwen2.5-Omni-7B-GGUF:Q4_K_M"
fi
# to test the huge models, run: ./tests.sh huge
# this will run both the big and huge models
# huge models are > 32B parameters
if [ "$RUN_HUGE_TESTS" = true ]; then
add_test_vision "ggml-org/Qwen2.5-VL-72B-Instruct-GGUF:Q4_K_M"
add_test_vision "ggml-org/Llama-4-Scout-17B-16E-Instruct-GGUF:IQ1_S"
fi
# these models always give the wrong answer, not sure why
# add_test_vision "ggml-org/SmolVLM-Instruct-GGUF:Q4_K_M"
# add_test_vision "ggml-org/SmolVLM-256M-Instruct-GGUF:Q8_0"
# add_test_vision "ggml-org/SmolVLM2-256M-Video-Instruct-GGUF:Q8_0"
# this model has broken chat template, not usable
# add_test_vision "cmp-nct/Yi-VL-6B-GGUF:Q5_K"
# add_test_vision "guinmoon/MobileVLM-3B-GGUF:Q4_K_M" "deepseek"
###############
cmake --build build -j --target llama-mtmd-cli
arr_res=()
for i in "${!arr_hf[@]}"; do
bin="llama-mtmd-cli"
prefix="${arr_prefix[$i]}"
hf="${arr_hf[$i]}"
tmpl="${arr_tmpl[$i]}"
inp_file="${arr_file[$i]}"
echo "Running test with binary: $bin and HF model: $hf"
echo ""
echo ""
output=$(\
"$PROJ_ROOT/build/bin/$bin" \
-hf "$hf" \
--image $SCRIPT_DIR/$inp_file \
-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
if echo "$output" | grep -iq "new york"; then
result="$prefix \033[32mOK\033[0m: $bin $hf"
else
result="$prefix \033[31mFAIL\033[0m: $bin $hf"
fi
echo -e "$result"
arr_res+=("$result")
echo ""
echo ""
echo ""
echo "#################################################"
echo "#################################################"
echo ""
echo ""
done
set +x
for i in "${!arr_res[@]}"; do
echo -e "${arr_res[$i]}"
done
echo ""
echo "Output logs are saved in $SCRIPT_DIR/output"

4
examples/server/server.cpp
View File

@@ -1235,7 +1235,7 @@ struct server_context {
chat_templates = common_chat_templates_init(model, params.chat_template);
try {
common_chat_format_example(chat_templates.get(), params.use_jinja);
common_chat_format_example(chat_templates.get(), params.use_jinja, {});
}
catch (const std::exception& e) {
LOG_WARNING("%s: The chat template that comes with this model is not yet supported, falling back to chatml. This may cause the model to output suboptimal responses\n", __func__);
@@ -3778,7 +3778,7 @@ int main(int argc, char ** argv) {
});
LOG_INFO("chat template", {
{"chat_example", common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params.use_jinja).c_str()
{"chat_example", common_chat_format_example(ctx_server.chat_templates.get(), ctx_server.params.use_jinja, {}).c_str()
},
{"built_in", params.chat_template.empty()},
});