CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

mtmd : add qwen2vl and qwen2.5vl (#13141)

Browse Source 
* llava : add clip_n_output_tokens, deprecate clip_n_patches

* mtmd : add qwen2vl and qwen2.5vl

* decode_embd_batch::set_position_...

* working version

* deprecate llama-qwen2vl-cli

* correct order W, H of clip_embd_nbytes_by_img

* edit existing line in hot topics
This commit is contained in:
Xuan-Son Nguyen
2025-04-29 11:47:04 +02:00
committed by GitHub
parent e98b3692be
commit 00e3e5a194
10 changed files with 196 additions and 79 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
README.md
View File

@@ -17,7 +17,7 @@ Inference of Meta's [LLaMA](https://arxiv.org/abs/2302.13971) model (and others)
## Hot topics
- **GGML developer experience survey (organized and reviewed by NVIDIA):** [link](https://forms.gle/Gasw3cRgyhNEnrwK9)
- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli` and `gemma3-cli` https://github.com/ggml-org/llama.cpp/pull/13012, `libllava` will be deprecated
- A new binary `llama-mtmd-cli` is introduced to replace `llava-cli`, `minicpmv-cli`, `gemma3-cli` ([#13012](https://github.com/ggml-org/llama.cpp/pull/13012)) and `qwen2vl-cli` ([#13141]((https://github.com/ggml-org/llama.cpp/pull/13141))), `libllava` will be deprecated
- VS Code extension for FIM completions: https://github.com/ggml-org/llama.vscode
- Universal [tool call support](./docs/function-calling.md) in `llama-server` https://github.com/ggml-org/llama.cpp/pull/9639
- Vim/Neovim plugin for FIM completions: https://github.com/ggml-org/llama.vim

8
examples/llava/CMakeLists.txt
View File

@@ -64,13 +64,7 @@ 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)
set(TARGET llama-qwen2vl-cli)
add_executable(${TARGET} qwen2vl-cli.cpp)
set_target_properties(${TARGET} PROPERTIES OUTPUT_NAME llama-qwen2vl-cli)
install(TARGETS ${TARGET} RUNTIME)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
target_compile_features(${TARGET} PRIVATE cxx_std_17)
add_executable(llama-qwen2vl-cli deprecation-warning.cpp)
set(TARGET llama-mtmd-cli)
add_executable(${TARGET} mtmd-cli.cpp)

34
examples/llava/clip.cpp
View File

@@ -2825,15 +2825,18 @@ void clip_free(clip_ctx * ctx) {
delete ctx;
}
// deprecated
size_t clip_embd_nbytes(const struct clip_ctx * ctx) {
return clip_n_patches(ctx) * clip_n_mmproj_embd(ctx) * sizeof(float);
const int32_t nx = ctx->vision_model.hparams.image_size;
const int32_t ny = ctx->vision_model.hparams.image_size;
return clip_embd_nbytes_by_img(ctx, nx, ny);
}
size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w) {
size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h) {
clip_image_f32 img;
img.nx = img_w;
img.ny = img_h;
return clip_n_patches_by_img(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float);
return clip_n_output_tokens(ctx, &img) * clip_n_mmproj_embd(ctx) * sizeof(float);
}
int32_t clip_get_image_size(const struct clip_ctx * ctx) {
@@ -2863,14 +2866,37 @@ size_t get_clip_image_grid_size(const struct clip_ctx * ctx) {
return ctx->vision_model.hparams.image_grid_pinpoints.size();
}
// deprecated
int clip_n_patches(const struct clip_ctx * ctx) {
clip_image_f32 img;
img.nx = ctx->vision_model.hparams.image_size;
img.ny = ctx->vision_model.hparams.image_size;
return clip_n_patches_by_img(ctx, &img);
return clip_n_output_tokens(ctx, &img);
}
// deprecated
int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
return clip_n_output_tokens(ctx, img);
}
int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
const auto & params = ctx->vision_model.hparams;
const int n_total = clip_n_output_tokens(ctx, img);
if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) {
return img->nx / (params.patch_size * 2) + (int)(img->nx % params.patch_size > 0);
}
return n_total;
}
int clip_n_output_tokens_y(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
const auto & params = ctx->vision_model.hparams;
if (ctx->proj_type == PROJECTOR_TYPE_QWEN2VL || ctx->proj_type == PROJECTOR_TYPE_QWEN25VL) {
return img->ny / (params.patch_size * 2) + (int)(img->ny % params.patch_size > 0);
}
return 1;
}
int clip_n_output_tokens(const struct clip_ctx * ctx, struct clip_image_f32 * img) {
const auto & params = ctx->vision_model.hparams;
int n_patches = (params.image_size / params.patch_size) * (params.image_size / params.patch_size);

19
examples/llava/clip.h
View File

@@ -47,7 +47,7 @@ CLIP_API struct clip_ctx * clip_init(const char * fname, struct clip_context_par
CLIP_API void clip_free(struct clip_ctx * ctx);
CLIP_API size_t clip_embd_nbytes(const struct clip_ctx * ctx);
CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_h, int img_w);
CLIP_API size_t clip_embd_nbytes_by_img(const struct clip_ctx * ctx, int img_w, int img_h);
CLIP_API int32_t clip_get_image_size (const struct clip_ctx * ctx);
CLIP_API int32_t clip_get_patch_size (const struct clip_ctx * ctx);
@@ -59,9 +59,20 @@ CLIP_API const char * clip_patch_merge_type(const struct clip_ctx * ctx);
CLIP_API const int32_t * clip_image_grid(const struct clip_ctx * ctx);
CLIP_API size_t get_clip_image_grid_size(const struct clip_ctx * ctx);
CLIP_API int clip_n_patches (const struct clip_ctx * ctx);
CLIP_API int clip_n_patches_by_img (const struct clip_ctx * ctx, struct clip_image_f32 * img);
CLIP_API int clip_n_mmproj_embd (const struct clip_ctx * ctx);
GGML_DEPRECATED(CLIP_API int clip_n_patches(const struct clip_ctx * ctx),
"use clip_n_output_tokens instead");
GGML_DEPRECATED(CLIP_API int clip_n_patches_by_img(const struct clip_ctx * ctx, struct clip_image_f32 * img),
"use clip_n_output_tokens instead");
CLIP_API 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
CLIP_API int clip_n_output_tokens_x(const struct clip_ctx * ctx, struct clip_image_f32 * img);
CLIP_API 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
CLIP_API int clip_n_mmproj_embd(const struct clip_ctx * ctx);
CLIP_API int clip_uhd_num_image_embeds_col(struct clip_ctx * ctx_clip);
CLIP_API void clip_add_load_image_size(struct clip_ctx * ctx_clip, struct clip_image_size * load_image_size);

15
examples/llava/llava.cpp
View File

@@ -112,7 +112,7 @@ static struct clip_image_grid_shape get_anyres_image_grid_shape(const std::pair<
}
// Take the image segments in a grid configuration and return the embeddings and the number of embeddings into preallocated memory (image_embd_out)
static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out) {
static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *> & image_embd_v, struct clip_image_grid_shape grid_shape, float * image_embd_out, int * n_img_pos_out, clip_image_f32 * img_input) {
struct {
struct ggml_context * ctx;
} model;
@@ -175,7 +175,7 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
model.ctx = ggml_init(params);
struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_patches(ctx_clip), num_images - 1); // example: 4096 x 576 x 4
struct ggml_tensor * image_features = ggml_new_tensor_3d(model.ctx, GGML_TYPE_F32, clip_n_mmproj_embd(ctx_clip), clip_n_output_tokens(ctx_clip, img_input), num_images - 1); // example: 4096 x 576 x 4
// ggml_tensor_printf(image_features,"image_features",__LINE__,false,false);
// fill it with the image embeddings, ignoring the base
for (size_t i = 1; i < num_images; i++) {
@@ -214,8 +214,8 @@ static bool clip_llava_handle_patches(clip_ctx * ctx_clip, std::vector<float *>
memcpy(image_embd_out, image_embd_v[0], clip_embd_nbytes(ctx_clip)); // main image as global context
// append without newline tokens (default behavior in llava_arch when not using unpad ):
memcpy(image_embd_out + clip_n_patches(ctx_clip) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
*n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_patches(ctx_clip));
memcpy(image_embd_out + clip_n_output_tokens(ctx_clip, img_input) * clip_n_mmproj_embd(ctx_clip), (float*)result->data, clip_embd_nbytes(ctx_clip) * (num_images-1)); // grid patches
*n_img_pos_out = static_cast<int>(result->ne[1]+clip_n_output_tokens(ctx_clip, img_input));
// Debug: Test single segments
// Current findings: sending base image, sending a segment embedding all works similar to python
@@ -313,7 +313,7 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
image_embd_v[i],
clip_embd_nbytes_by_img(ctx_clip, nx, ny));
n_img_pos_out += clip_n_patches_by_img(ctx_clip, img_res);
n_img_pos_out += clip_n_output_tokens(ctx_clip, img_res);
}
*n_img_pos = n_img_pos_out;
for (size_t i = 0; i < image_embd_v.size(); i++) {
@@ -342,8 +342,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
}
else if (strcmp(mm_patch_merge_type, "spatial_unpad") != 0) {
// flat / default llava-1.5 type embedding
*n_img_pos = clip_n_patches(ctx_clip);
clip_image_f32 * img_res = clip_image_f32_get_img(img_res_v.get(), 0);
*n_img_pos = clip_n_output_tokens(ctx_clip, img_res);
bool encoded = clip_image_encode(ctx_clip, n_threads, img_res, image_embd); // image_embd shape is 576 x 4096
if (!encoded) {
LOG_ERR("Unable to encode image\n");
@@ -381,7 +381,8 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
struct clip_image_grid_shape grid_shape = get_anyres_image_grid_shape({img->nx,img->ny}, grid_pinpoints, image_size);
int n_img_pos_out;
clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out);
clip_image_f32 * img_input = clip_image_f32_get_img(img_res_v.get(), 0);
clip_llava_handle_patches(ctx_clip, image_embd_v, grid_shape, image_embd, &n_img_pos_out, img_input);
*n_img_pos = n_img_pos_out;
for (size_t i = 0; i < image_embd_v.size(); i++) {

36
examples/llava/mtmd-cli.cpp
View File

@@ -136,39 +136,6 @@ struct mtmd_cli_context {
}
};
struct decode_embd_batch {
std::vector<llama_pos> pos;
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, llama_pos pos_0, llama_seq_id seq_id) {
pos .resize(n_tokens);
n_seq_id.resize(n_tokens);
seq_ids .resize(n_tokens + 1);
logits .resize(n_tokens);
seq_id_0.resize(1);
seq_id_0[0] = seq_id;
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(),
};
for (int i = 0; i < 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;
}
}
};
static int generate_response(mtmd_cli_context & ctx, common_sampler * smpl, int n_predict) {
llama_tokens generated_tokens;
for (int i = 0; i < n_predict; i++) {
@@ -243,7 +210,7 @@ static int eval_message(mtmd_cli_context & ctx, common_chat_msg & msg, std::vect
return 1;
}
ctx.n_past += mtmd_helper_get_n_tokens(chunks);
ctx.n_past += mtmd_helper_get_n_pos(chunks);
return 0;
}
@@ -371,6 +338,7 @@ int main(int argc, char ** argv) {
}
}
if (g_is_interrupted) LOG("\nInterrupted by user\n");
LOG("\n\n");
llama_perf_context_print(ctx.lctx);
return g_is_interrupted ? 130 : 0;
}

146
examples/llava/mtmd.cpp
View File

@@ -40,11 +40,14 @@ struct mtmd_context {
llama_token tok_sli_img_end = LLAMA_TOKEN_NULL; // single slice
llama_token tok_row_end = LLAMA_TOKEN_NULL; // end of row
bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE
// TODO @ngxson : add timings
mtmd_context(const char * mmproj_fname,
const llama_model * text_model,
const mtmd_context_params & ctx_params) :
text_model (text_model),
print_timings(ctx_params.print_timings),
n_threads (ctx_params.n_threads),
image_marker (ctx_params.image_marker)
@@ -56,9 +59,8 @@ struct mtmd_context {
if (!ctx_clip) {
throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
}
this->text_model = text_model;
GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead");
use_mrope = clip_is_qwen2vl(ctx_clip);
int minicpmv_version = clip_is_minicpmv(ctx_clip);
if (minicpmv_version == 2) {
@@ -126,6 +128,7 @@ struct mtmd_image_tokens_data {
struct mtmd_image_tokens {
uint32_t nx; // number of tokens in x direction
uint32_t ny; // number of tokens in y direction
bool use_mrope_pos = false; // use M-RoPE position counting (the whole image is 1 temporal position)
uint32_t n_tokens() const { return nx * ny; }
clip_image_f32_batch batch_f32; // preprocessed image patches
std::string id; // optional user-defined ID, useful for KV cache tracking
@@ -202,6 +205,13 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
}
else if (proj_type == PROJECTOR_TYPE_QWEN2VL || proj_type == PROJECTOR_TYPE_QWEN25VL) {
// <|vision_start|> ... (image embeddings) ... <|vision_end|>
marker_modified = "<|vision_start|>" + ctx->image_marker + "<|vision_end|>";
string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
}
// llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
@@ -226,7 +236,7 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
for (auto & entry : batch_f32.entries) {
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
image_tokens->nx = clip_n_patches_by_img(ctx->ctx_clip, entry.get());
image_tokens->nx = clip_n_output_tokens(ctx->ctx_clip, entry.get());
image_tokens->ny = 1;
image_tokens->batch_f32.entries.push_back(std::move(entry));
image_tokens->id = id;
@@ -322,12 +332,20 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
} else {
size_t n_tokens = 0;
for (const auto & entry : batch_f32.entries) {
n_tokens += clip_n_patches_by_img(ctx->ctx_clip, entry.get());
n_tokens += clip_n_output_tokens(ctx->ctx_clip, entry.get());
}
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
image_tokens->nx = n_tokens;
image_tokens->ny = 1; // TODO
if (ctx->use_mrope) {
// for Qwen2VL, we need this information for M-RoPE decoding positions
image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_clip, batch_f32.entries[0].get());
image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_clip, batch_f32.entries[0].get());
image_tokens->use_mrope_pos = true;
} else {
// other models, we only need the total number of tokens
image_tokens->nx = n_tokens;
image_tokens->ny = 1;
}
image_tokens->batch_f32 = std::move(batch_f32);
image_tokens->id = bitmaps[i_img].id; // optional
@@ -372,6 +390,13 @@ std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
return image_tokens->id;
}
llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens) {
if (image_tokens->use_mrope_pos) {
return 1; // for M-RoPE, the whole image is 1 in temporal dimension
}
return image_tokens->n_tokens();
}
int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
@@ -389,7 +414,7 @@ int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens)
// TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
const auto & entries = image_tokens->batch_f32.entries;
for (size_t i = 0; i < entries.size(); i++) {
int n_tokens_per_image = clip_n_patches_by_img(ctx->ctx_clip, entries[i].get());
int n_tokens_per_image = clip_n_output_tokens(ctx->ctx_clip, entries[i].get());
ok = clip_image_encode(
ctx->ctx_clip,
ctx->n_threads,
@@ -417,7 +442,7 @@ size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
n_tokens += chunk.tokens_text.size();
} else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
n_tokens += chunk.tokens_image->n_tokens();
n_tokens += mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
} else {
GGML_ASSERT(false && "chunk type not supported");
}
@@ -425,22 +450,38 @@ size_t mtmd_helper_get_n_tokens(mtmd_input_chunks & chunks) {
return n_tokens;
}
llama_pos mtmd_helper_get_n_pos(mtmd_input_chunks & chunks) {
llama_pos n_pos = 0;
for (auto & chunk : chunks) {
if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
n_pos += chunk.tokens_text.size();
} else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
n_pos += mtmd_image_tokens_get_n_pos(chunk.tokens_image.get());
} else {
GGML_ASSERT(false && "chunk type not supported");
}
}
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, llama_pos pos_0, llama_seq_id seq_id) {
pos .resize(n_tokens);
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_id_0[0] = seq_id;
seq_ids [n_tokens] = nullptr;
batch = {
/*n_tokens =*/ n_tokens,
@@ -451,13 +492,64 @@ struct decode_embd_batch {
/*seq_id =*/ seq_ids.data(),
/*logits =*/ logits.data(),
};
for (int i = 0; i < n_tokens; i++) {
}
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;
}
}
void set_position_mrope(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;
}
}
llama_batch get_view(int offset, int n_tokens) {
llama_pos * pos_ptr;
pos_view.clear();
pos_view.resize(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++) {
auto src = pos.begin() + i * batch.n_tokens + offset;
pos_view.insert(pos_view.end(), src, src + 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,
};
}
};
int32_t mtmd_helper_eval(mtmd_context * ctx,
@@ -470,6 +562,7 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
llama_pos n_past = pos0;
llama_batch text_batch = llama_batch_init(n_batch, 0, 1);
int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
int n_pos_per_embd = mtmd_decode_use_mrope(ctx) ? 4 : 1;
for (auto & chunk : chunks) {
bool is_last = &chunk == &chunks.back();
@@ -517,6 +610,16 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
int32_t i_batch = 0;
int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
float * embd = mtmd_get_output_embd(ctx);
decode_embd_batch batch_embd(embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
const int nx = mtmd_image_tokens_get_nx(chunk.tokens_image.get());
const int ny = mtmd_image_tokens_get_ny(chunk.tokens_image.get());
if (mtmd_decode_use_mrope(ctx)) {
batch_embd.set_position_mrope(n_past, nx, ny, seq_id);
} else {
batch_embd.set_position_normal(n_past, seq_id);
}
if (mtmd_decode_use_non_causal(ctx)) {
llama_set_causal_attn(lctx, false);
@@ -524,15 +627,14 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
}
while (i_batch < n_img_batches) { // split into batches
int32_t pos_offset = i_batch*n_batch;
int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
float * embd_batch = embd + pos_offset*n_mmproj_embd;
decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0);
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);
printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
LOG_INF("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
int64_t t1 = ggml_time_ms();
ret = llama_decode(lctx, batch_img.batch);
ret = llama_decode(lctx, batch_embd_view);
if (ret != 0) {
LOG_ERR("failed to decode image\n");
llama_set_causal_attn(lctx, true); // restore causal attn
@@ -545,9 +647,11 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
}
i_batch++;
n_past += n_tokens_batch;
}
// for mrope, one image is one single **temporal** position
n_past += mtmd_decode_use_mrope(ctx) ? 1 : n_tokens;
if (mtmd_decode_use_non_causal(ctx)) {
llama_set_causal_attn(lctx, true);
}
@@ -595,6 +699,10 @@ bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
return false;
}
bool mtmd_decode_use_mrope(mtmd_context * ctx) {
return ctx->use_mrope;
}
void mtmd_image_tokens_deleter::operator()(mtmd_image_tokens * val) {
mtmd_image_tokens_free(val);
}

9
examples/llava/mtmd.h
View File

@@ -102,6 +102,7 @@ MTMD_API size_t mtmd_image_tokens_get_n_tokens(const mtmd_image_tokens * im
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 std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens);
MTMD_API llama_pos mtmd_image_tokens_get_n_pos(const mtmd_image_tokens * image_tokens); // number of temporal positions (always 1 for M-RoPE, n_tokens otherwise)
MTMD_API void mtmd_image_tokens_free(mtmd_image_tokens * image_tokens);
// returns 0 on success
@@ -114,15 +115,21 @@ MTMD_API float * mtmd_get_output_embd(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);
//
// helper functions (can be implemented based on other functions)
//
// helper to count the total number of tokens from a list of chunks, useful to keep track of n_past
// 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(mtmd_input_chunks & chunks);
// helper to count the total position of tokens from a list of chunks, useful to keep track of n_past
MTMD_API llama_pos mtmd_helper_get_n_pos(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()

2
examples/llava/qwen2vl-cli.cpp → examples/llava/qwen2vl-test.cpp
View File

@@ -27,6 +27,8 @@
#include <cassert>
#include <cmath>
// THIS FILE IS ONLY USED FOR TESTING THE QWEN2VL MODEL
// IT IS NOT A PRODUCTION CODE
static bool qwen2vl_eval_image_embed(llama_context * ctx_llama, const struct llava_image_embed * image_embed,
int n_batch, int * n_past, int * st_pos_id, struct clip_image_size * image_size) {

4
examples/llava/tests.sh
View File

@@ -54,8 +54,8 @@ add_test "llama-mtmd-cli" "ibm-research/granite-vision-3.2-2b-GGUF:Q4_K_M"
add_test "llama-mtmd-cli" "second-state/MiniCPM-Llama3-V-2_5-GGUF:Q2_K" # model from openbmb is corrupted
add_test "llama-mtmd-cli" "openbmb/MiniCPM-V-2_6-gguf:Q2_K"
add_test "llama-mtmd-cli" "openbmb/MiniCPM-o-2_6-gguf:Q4_0"
add_test "llama-qwen2vl-cli" "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
add_test "llama-qwen2vl-cli" "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
add_test "llama-mtmd-cli" "bartowski/Qwen2-VL-2B-Instruct-GGUF:Q4_K_M"
add_test "llama-mtmd-cli" "ggml-org/Qwen2.5-VL-3B-Instruct-GGUF:Q4_K_M"
# to test the big models, run: ./tests.sh big
add_test_big "llama-mtmd-cli" "ggml-org/pixtral-12b-GGUF:Q4_K_M"