vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

llama : add Qwen2VL support + multimodal RoPE (#10361)

Browse source 
* Barebone Qwen2VL LLM convertor

* Add Qwen2VL cli entrypoint

* [WIP] add qwen2vl arch

* Verify m-rope output

* Add vl-rope/2d-rope support for qwen2vl ViT

* update qwen2vl cli tool

* update 5D tensor op workaround

* [WIP] qwen2vl vision model

* make batch and clip utils compatible with qwen2vl

* [WIP] create inference workflow, gguf convert script but fix

* correcting vision-rope behavior, add the missing last layer back to ViT

* add arg parser to qwen2vl_surgery

* replace variable size array with vector

* cuda-gdb cmake preset

* add fp32 mrope, vision rope kernel

* add fp16 support for qwen2vl and m-rope

* add `GGML_ROPE_TYPE_MROPE`, `GGML_ROPE_TYPE_VISION`

* fix rope op mode switching, out dated func args

* update `llama_hparams`

* update to keep up stream changes

* resolve linter, test errors

* add makefile entry, update speical image padding token

* add mrope unit test, fix few compiler warnings

* rename `mrope` related function, params

* minor updates on debug util, bug fixs

* add `m-rope` testcase to `test-backend-ops`

* Apply suggestions from code review

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

* fix traililng whitespce

* store `llama_hparams.rope_sections` with fixed size array

* update position id tensor size check in GGML_OP_ROPE

* minor updates

* update `ggml_backend_*_supports_op` of unsupported backends

* remote old `rope_section` compare operator

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
HimariO 2024-12-14 20:43:46 +08:00 • committed by GitHub
parent 56eea0781c
commit ba1cb19cdd
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
24 changed files with 1873 additions and 114 deletions
Download patch file Download diff file Expand all files Collapse all files

37
examples/llava/llava.cpp
View file

@ -259,25 +259,33 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
const char * mm_patch_merge_type = clip_patch_merge_type(ctx_clip);
if (clip_is_minicpmv(ctx_clip)) {
if (clip_is_minicpmv(ctx_clip) || clip_is_qwen2vl(ctx_clip)) {
std::vector<float *> image_embd_v;
image_embd_v.resize(img_res_v.size);
struct clip_image_size * load_image_size = clip_image_size_init();
for (size_t i = 0; i < img_res_v.size; i++) {
const int64_t t_img_enc_step_start_us = ggml_time_us();
image_embd_v[i] = (float *)malloc(clip_embd_nbytes(ctx_clip));
image_embd_v[i] = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
int patch_size=14;
load_image_size->width = img_res_v.data[i].nx;
load_image_size->height = img_res_v.data[i].ny;
clip_add_load_image_size(ctx_clip, load_image_size);
bool encoded = false;
int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
if (has_minicpmv_projector == 2) {
encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
}
else if (has_minicpmv_projector == 3) {
if (clip_is_qwen2vl(ctx_clip)) {
encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
}
else {
int has_minicpmv_projector = clip_is_minicpmv(ctx_clip);
if (has_minicpmv_projector == 2) {
encoded = clip_image_encode(ctx_clip, n_threads, only_v2_5_reshape_by_patch(&img_res_v.data[i], patch_size), image_embd_v[i]);
}
else if (has_minicpmv_projector == 3) {
encoded = clip_image_encode(ctx_clip, n_threads, &img_res_v.data[i], image_embd_v[i]);
}
}
if (!encoded) {
LOG_ERR("Unable to encode image - spatial_unpad - subimage %d of %d\n", (int) i+1, (int) img_res_v.size);
return false;
@ -290,8 +298,11 @@ static bool encode_image_with_clip(clip_ctx * ctx_clip, int n_threads, const cli
int n_img_pos_out = 0;
for (size_t i = 0; i < image_embd_v.size(); i++) {
std::memcpy(image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip), image_embd_v[i], clip_embd_nbytes(ctx_clip));
n_img_pos_out += clip_n_patches(ctx_clip);
std::memcpy(
image_embd + n_img_pos_out * clip_n_mmproj_embd(ctx_clip),
image_embd_v[i],
clip_embd_nbytes_by_img(ctx_clip, img_res_v.data[i].nx, img_res_v.data[i].ny));
n_img_pos_out += clip_n_patches_by_img(ctx_clip, &img_res_v.data[i]);
}
*n_img_pos = n_img_pos_out;
for (size_t i = 0; i < image_embd_v.size(); i++) {
@ -387,7 +398,13 @@ bool llava_image_embed_make_with_clip_img(clip_ctx * ctx_clip, int n_threads, co
if (clip_is_minicpmv(ctx_clip)) {
num_max_patches = 10;
}
float * image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
float * image_embd;
if (clip_is_qwen2vl(ctx_clip)) {
// qwen2vl don't split image into chunks, so `num_max_patches` is not needed.
image_embd = (float *)malloc(clip_embd_nbytes_by_img(ctx_clip, img->nx, img->ny));
} else {
image_embd = (float *)malloc(clip_embd_nbytes(ctx_clip)*num_max_patches); // TODO: base on gridsize/llava model
}
if (!image_embd) {
LOG_ERR("Unable to allocate memory for image embeddings\n");
return false;