mirror of
https://github.com/ggml-org/llama.cpp.git
synced 2025-06-27 20:05:20 +00:00
mtmd : merge llava, gemma3 and minicpmv CLI into single llama-mtmd-cli (#13012)
* mtmd : merge `llava-cli` and `gemma3-cli` into single `mtmd-cli` * support for minicpmv * remove cpp files of llava and minicpmv * update hot topics * mtmd : add not supported msg for qwen2vl * Update examples/llava/mtmd.cpp Co-authored-by: Georgi Gerganov <ggerganov@gmail.com> --------- Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
This commit is contained in:
Xuan-Son Nguyen
GitHub
@ -12,6 +12,15 @@
|
||||
#include <limits>
|
||||
#include <vector>
|
||||
|
||||
// slice template, used by some llava-uhd models to correctly place the special tokens around image embeddings
|
||||
// models not having it (llava-1.6) will process embeddings without any special tokens in-between
|
||||
enum mtmd_slice_tmpl {
|
||||
MTMD_SLICE_TMPL_NONE,
|
||||
MTMD_SLICE_TMPL_MINICPMV_2_5,
|
||||
MTMD_SLICE_TMPL_MINICPMV_2_6,
|
||||
// TODO @ngxson : add support for idefics (SmolVLM)
|
||||
};
|
||||
|
||||
struct mtmd_context {
|
||||
struct clip_ctx * ctx_clip;
|
||||
const struct llama_model * text_model;
|
||||
@ -21,6 +30,16 @@ struct mtmd_context {
|
||||
int n_threads;
|
||||
std::string image_marker;
|
||||
|
||||
// for minicpmv, we need special tokens in-between slices
|
||||
mtmd_slice_tmpl slice_tmpl = MTMD_SLICE_TMPL_NONE;
|
||||
llama_token tok_ov_img_start = LLAMA_TOKEN_NULL; // overview image
|
||||
llama_token tok_ov_img_end = LLAMA_TOKEN_NULL; // overview image
|
||||
llama_token tok_slices_start = LLAMA_TOKEN_NULL; // start of all slices
|
||||
llama_token tok_slices_end = LLAMA_TOKEN_NULL; // end of all slices
|
||||
llama_token tok_sli_img_start = LLAMA_TOKEN_NULL; // single slice
|
||||
llama_token tok_sli_img_end = LLAMA_TOKEN_NULL; // single slice
|
||||
llama_token tok_row_end = LLAMA_TOKEN_NULL; // end of row
|
||||
|
||||
// TODO @ngxson : add timings
|
||||
|
||||
mtmd_context(const char * mmproj_fname,
|
||||
@ -38,11 +57,66 @@ struct mtmd_context {
|
||||
throw std::runtime_error(string_format("Failed to load CLIP model from %s\n", mmproj_fname));
|
||||
}
|
||||
this->text_model = text_model;
|
||||
|
||||
GGML_ASSERT(!clip_is_qwen2vl(ctx_clip) && "Qwen2VL model is not supported yet, use llama-qwen2vl-cli instead");
|
||||
|
||||
int minicpmv_version = clip_is_minicpmv(ctx_clip);
|
||||
if (minicpmv_version == 2) {
|
||||
// minicpmv 2.5 format:
|
||||
// <image> (overview) </image><slice><image> (slice) </image><image> (slice) </image>\n ... </slice>
|
||||
slice_tmpl = MTMD_SLICE_TMPL_MINICPMV_2_5;
|
||||
tok_ov_img_start = lookup_token("<image>");
|
||||
tok_ov_img_end = lookup_token("</image>");
|
||||
tok_slices_start = lookup_token("<slice>");
|
||||
tok_slices_end = lookup_token("</slice>");
|
||||
tok_sli_img_start = tok_ov_img_start;
|
||||
tok_sli_img_end = tok_ov_img_end;
|
||||
tok_row_end = lookup_token("\n");
|
||||
|
||||
} else if (minicpmv_version == 3 || minicpmv_version == 4) {
|
||||
// minicpmv 2.6 format:
|
||||
// <image> (overview) </image><slice> (slice) </slice><slice> (slice) </slice>\n ...
|
||||
slice_tmpl = MTMD_SLICE_TMPL_MINICPMV_2_6;
|
||||
tok_ov_img_start = lookup_token("<image>");
|
||||
tok_ov_img_end = lookup_token("</image>");
|
||||
tok_sli_img_start = lookup_token("<slice>");
|
||||
tok_sli_img_end = lookup_token("</slice>");
|
||||
tok_row_end = lookup_token("\n");
|
||||
|
||||
} else if (minicpmv_version != 0) {
|
||||
GGML_ASSERT(false && "unsupported minicpmv version");
|
||||
}
|
||||
}
|
||||
|
||||
~mtmd_context() {
|
||||
clip_free(ctx_clip);
|
||||
}
|
||||
|
||||
private:
|
||||
llama_token lookup_token(const std::string & token_text) {
|
||||
const llama_vocab * vocab = llama_model_get_vocab(text_model);
|
||||
const int n_vocab = llama_vocab_n_tokens(vocab);
|
||||
for (int i = 0; i < n_vocab; i++) {
|
||||
if (token_to_piece(vocab, i, true) == token_text) {
|
||||
return i;
|
||||
}
|
||||
}
|
||||
return LLAMA_TOKEN_NULL;
|
||||
}
|
||||
|
||||
std::string token_to_piece(const llama_vocab * vocab, llama_token token, bool special) {
|
||||
std::string piece;
|
||||
piece.resize(piece.capacity()); // using string internal cache, 15 bytes + '\n'
|
||||
const int n_chars = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
|
||||
if (n_chars < 0) {
|
||||
piece.resize(-n_chars);
|
||||
int check = llama_token_to_piece(vocab, token, &piece[0], piece.size(), 0, special);
|
||||
GGML_ASSERT(check == -n_chars);
|
||||
} else {
|
||||
piece.resize(n_chars);
|
||||
}
|
||||
return piece;
|
||||
}
|
||||
};
|
||||
|
||||
struct mtmd_image_tokens_data {
|
||||
@ -102,21 +176,58 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
|
||||
|
||||
std::string prompt_modified(text.text);
|
||||
std::string marker_modified(ctx->image_marker);
|
||||
projector_type proj_type = clip_get_projector_type(ctx->ctx_clip);
|
||||
// a bit hacky here, but works for now
|
||||
// for some models, we need to add prefix and suffix to the image embeddings
|
||||
if (proj_type == PROJECTOR_TYPE_GEMMA3) {
|
||||
if (clip_is_gemma3(ctx->ctx_clip)) {
|
||||
// gemma 3
|
||||
// <start_of_image> ... (image embeddings) ... <end_of_image>
|
||||
marker_modified = "<start_of_image>" + ctx->image_marker + "<end_of_image>";
|
||||
string_replace_all(prompt_modified, ctx->image_marker, marker_modified);
|
||||
}
|
||||
|
||||
// llava-1.5, llava-1.6, Yi-VL, Yi-34B, granite: don't need to add prefix and suffix
|
||||
// for glm-edge, we don't need to add because the tokens are already in the returned embeddings
|
||||
|
||||
// TODO @ngxson : glm-edge : remove BOI / EOI tokens embeddings, decode them as normal tokens
|
||||
|
||||
std::vector<std::string> parts = string_split_str(prompt_modified, ctx->image_marker);
|
||||
output.clear();
|
||||
output.reserve(parts.size());
|
||||
|
||||
size_t i_img = 0;
|
||||
|
||||
// utility for adding raw tokens
|
||||
auto add_text_chunk = [&output](std::vector<llama_token> && tokens) {
|
||||
mtmd_input_chunk chunk{
|
||||
MTMD_INPUT_CHUNK_TYPE_TEXT,
|
||||
std::move(tokens),
|
||||
{},
|
||||
};
|
||||
output.emplace_back(std::move(chunk));
|
||||
};
|
||||
|
||||
// utility for splitting batch of multiple images into chunks of batch having single images
|
||||
auto split_batch_to_chunk = [&ctx](clip_image_f32_batch && batch_f32, const std::string & id) {
|
||||
std::vector<mtmd_input_chunk> chunks;
|
||||
|
||||
for (auto & entry : batch_f32.entries) {
|
||||
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
|
||||
image_tokens->nx = clip_n_patches(ctx->ctx_clip);
|
||||
image_tokens->ny = 1;
|
||||
image_tokens->batch_f32.entries.push_back(std::move(entry));
|
||||
image_tokens->id = id;
|
||||
|
||||
mtmd_input_chunk chunk{
|
||||
MTMD_INPUT_CHUNK_TYPE_IMAGE,
|
||||
{},
|
||||
std::move(image_tokens),
|
||||
};
|
||||
chunks.emplace_back(std::move(chunk));
|
||||
}
|
||||
|
||||
return chunks;
|
||||
};
|
||||
|
||||
for (const auto & part : parts) {
|
||||
//printf("tokenizing part: %s\n", part.c_str());
|
||||
bool add_bos = &parts.front() == ∂
|
||||
@ -139,12 +250,13 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
|
||||
return 1;
|
||||
}
|
||||
|
||||
// shim layer
|
||||
// convert mtmd_bitmap to clip_image_u8
|
||||
clip_image_u8_ptr img_u8(clip_image_u8_init());
|
||||
img_u8->nx = bitmaps[i_img].nx;
|
||||
img_u8->ny = bitmaps[i_img].ny;
|
||||
img_u8->buf.resize(bitmaps[i_img].data.size());
|
||||
std::memcpy(img_u8->buf.data(), bitmaps[i_img].data.data(), img_u8->nx * img_u8->ny * 3);
|
||||
clip_image_size img_u8_size{img_u8->nx, img_u8->ny};
|
||||
|
||||
// preprocess image
|
||||
clip_image_f32_batch batch_f32;
|
||||
@ -154,19 +266,70 @@ int32_t mtmd_tokenize(mtmd_context * ctx,
|
||||
return 2;
|
||||
}
|
||||
|
||||
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
|
||||
image_tokens->nx = clip_n_patches(ctx->ctx_clip); // TODO @ngxson : use clip_n_patches_by_image
|
||||
image_tokens->ny = 1; // TODO
|
||||
image_tokens->batch_f32 = std::move(batch_f32);
|
||||
image_tokens->id = bitmaps[i_img].id; // optional
|
||||
if (ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_5 || ctx->slice_tmpl == MTMD_SLICE_TMPL_MINICPMV_2_6) {
|
||||
// split batch into chunks of single images
|
||||
auto chunks = split_batch_to_chunk(std::move(batch_f32), bitmaps[i_img].id);
|
||||
GGML_ASSERT(chunks.size() > 0);
|
||||
|
||||
mtmd_input_chunk chunk{
|
||||
MTMD_INPUT_CHUNK_TYPE_IMAGE,
|
||||
{},
|
||||
std::move(image_tokens),
|
||||
};
|
||||
output.emplace_back(std::move(chunk));
|
||||
i_img++;
|
||||
// add overview image
|
||||
add_text_chunk({ctx->tok_ov_img_start});
|
||||
output.emplace_back(std::move(chunks.front()));
|
||||
chunks.erase(chunks.begin());
|
||||
add_text_chunk({ctx->tok_ov_img_end});
|
||||
|
||||
// add slices
|
||||
if (!chunks.empty()) {
|
||||
clip_add_load_image_size(ctx->ctx_clip, &img_u8_size);
|
||||
int n_col = clip_uhd_num_image_embeds_col(ctx->ctx_clip);
|
||||
int n_row = (int)chunks.size() / n_col;
|
||||
GGML_ASSERT(n_row * n_col == (int)chunks.size());
|
||||
if (ctx->tok_slices_start != LLAMA_TOKEN_NULL) {
|
||||
add_text_chunk({ctx->tok_slices_start});
|
||||
}
|
||||
for (int y = 0; y < n_row; y++) {
|
||||
for (int x = 0; x < n_col; x++) {
|
||||
if (ctx->tok_sli_img_start != LLAMA_TOKEN_NULL) {
|
||||
add_text_chunk({ctx->tok_sli_img_start});
|
||||
}
|
||||
output.emplace_back(std::move(chunks[y * n_col + x]));
|
||||
if (ctx->tok_sli_img_end != LLAMA_TOKEN_NULL) {
|
||||
add_text_chunk({ctx->tok_sli_img_end});
|
||||
}
|
||||
}
|
||||
if (ctx->tok_row_end != LLAMA_TOKEN_NULL && y != n_row - 1) {
|
||||
add_text_chunk({ctx->tok_row_end});
|
||||
}
|
||||
}
|
||||
if (ctx->tok_slices_end != LLAMA_TOKEN_NULL) {
|
||||
add_text_chunk({ctx->tok_slices_end});
|
||||
}
|
||||
}
|
||||
|
||||
} else {
|
||||
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
|
||||
image_tokens->nx = clip_n_patches(ctx->ctx_clip) * batch_f32.entries.size(); // TODO @ngxson : use clip_n_patches_by_image
|
||||
image_tokens->ny = 1; // TODO
|
||||
image_tokens->batch_f32 = std::move(batch_f32);
|
||||
image_tokens->id = bitmaps[i_img].id; // optional
|
||||
|
||||
LOG_DBG("image_tokens->nx = %d\n", image_tokens->nx);
|
||||
LOG_DBG("image_tokens->ny = %d\n", image_tokens->ny);
|
||||
LOG_DBG("batch_f32 size = %d\n", (int)image_tokens->batch_f32.entries.size());
|
||||
|
||||
if (clip_is_glm(ctx->ctx_clip)) {
|
||||
// glm-edge
|
||||
image_tokens->nx += 2; // add 2 for the begin_of_image and end_of_image token embeddings
|
||||
}
|
||||
|
||||
mtmd_input_chunk chunk{
|
||||
MTMD_INPUT_CHUNK_TYPE_IMAGE,
|
||||
{},
|
||||
std::move(image_tokens),
|
||||
};
|
||||
output.emplace_back(std::move(chunk));
|
||||
}
|
||||
|
||||
i_img++; // move to next image
|
||||
}
|
||||
}
|
||||
|
||||
@ -198,11 +361,35 @@ std::string mtmd_image_tokens_get_id(const mtmd_image_tokens * image_tokens) {
|
||||
int32_t mtmd_encode(mtmd_context * ctx, const mtmd_image_tokens * image_tokens) {
|
||||
int n_mmproj_embd = clip_n_mmproj_embd(ctx->ctx_clip);
|
||||
ctx->image_embd_v.resize(image_tokens->n_tokens() * n_mmproj_embd);
|
||||
bool ok = clip_image_batch_encode(
|
||||
ctx->ctx_clip,
|
||||
ctx->n_threads,
|
||||
&image_tokens->batch_f32,
|
||||
ctx->image_embd_v.data());
|
||||
bool ok = false;
|
||||
|
||||
// only effective for minicpmv and qwen2vl, other models will ignore load_image_size
|
||||
{
|
||||
clip_image_size slice_size{
|
||||
image_tokens->batch_f32.entries[0]->nx,
|
||||
image_tokens->batch_f32.entries[0]->ny};
|
||||
clip_add_load_image_size(ctx->ctx_clip, &slice_size);
|
||||
}
|
||||
|
||||
if (clip_is_llava(ctx->ctx_clip) || clip_is_minicpmv(ctx->ctx_clip) || clip_is_glm(ctx->ctx_clip)) {
|
||||
// TODO @ngxson : llava does not support batched encoding ; this should be fixed inside clip_image_batch_encode()
|
||||
const auto & entries = image_tokens->batch_f32.entries;
|
||||
for (size_t i = 0; i < entries.size(); i++) {
|
||||
int n_tokens_per_image = clip_n_patches(ctx->ctx_clip);
|
||||
ok = clip_image_encode(
|
||||
ctx->ctx_clip,
|
||||
ctx->n_threads,
|
||||
entries[i].get(),
|
||||
ctx->image_embd_v.data() + i*n_mmproj_embd*n_tokens_per_image);
|
||||
}
|
||||
} else {
|
||||
ok = clip_image_batch_encode(
|
||||
ctx->ctx_clip,
|
||||
ctx->n_threads,
|
||||
&image_tokens->batch_f32,
|
||||
ctx->image_embd_v.data());
|
||||
}
|
||||
|
||||
return ok ? 0 : 1;
|
||||
}
|
||||
|
||||
@ -268,28 +455,31 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
|
||||
int32_t ret;
|
||||
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);
|
||||
|
||||
for (auto & chunk : chunks) {
|
||||
bool is_last = &chunk == &chunks.back();
|
||||
if (chunk.type == MTMD_INPUT_CHUNK_TYPE_TEXT) {
|
||||
// TODO @ngxson : may need to split into smaller batches
|
||||
text_batch.n_tokens = chunk.tokens_text.size();
|
||||
for (size_t i = 0; i < chunk.tokens_text.size(); i++) {
|
||||
text_batch.token [i] = chunk.tokens_text[i];
|
||||
text_batch.pos [i] = n_past++;
|
||||
text_batch.n_seq_id[i] = 1;
|
||||
text_batch.seq_id [i][0] = seq_id;
|
||||
text_batch.logits [i] = false;
|
||||
}
|
||||
if (is_last) {
|
||||
// always get logits for last input chunk
|
||||
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;
|
||||
size_t i = 0;
|
||||
while (i < chunk.tokens_text.size()) { // split into batches
|
||||
for (; i < chunk.tokens_text.size() && text_batch.n_tokens < n_batch; i++) {
|
||||
text_batch.token [i] = chunk.tokens_text[i];
|
||||
text_batch.pos [i] = n_past++;
|
||||
text_batch.n_seq_id[i] = 1;
|
||||
text_batch.seq_id [i][0] = seq_id;
|
||||
text_batch.logits [i] = false;
|
||||
}
|
||||
if (is_last) {
|
||||
// always get logits for last input chunk
|
||||
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;
|
||||
}
|
||||
}
|
||||
|
||||
} else if (chunk.type == MTMD_INPUT_CHUNK_TYPE_IMAGE) {
|
||||
@ -297,7 +487,7 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
|
||||
GGML_ASSERT(chunk.tokens_image != nullptr);
|
||||
int64_t t0 = ggml_time_ms();
|
||||
if (ctx->print_timings) {
|
||||
LOG_INF("encoding image...\n");
|
||||
LOG_INF("encoding image or slice...\n");
|
||||
}
|
||||
ret = mtmd_encode(ctx, chunk.tokens_image.get());
|
||||
if (ret != 0) {
|
||||
@ -306,24 +496,47 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
|
||||
return ret;
|
||||
}
|
||||
if (ctx->print_timings) {
|
||||
LOG_INF("image encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
|
||||
LOG_INF("image/slice encoded in %" PRId64 " ms\n", ggml_time_ms() - t0);
|
||||
}
|
||||
|
||||
int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
|
||||
int32_t 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_img(embd, n_tokens, n_past, 0);
|
||||
int64_t t1 = ggml_time_ms();
|
||||
ret = llama_decode(lctx, batch_img.batch);
|
||||
if (ret != 0) {
|
||||
LOG_ERR("failed to decode image\n");
|
||||
llama_batch_free(text_batch);
|
||||
return ret;
|
||||
}
|
||||
if (ctx->print_timings) {
|
||||
LOG_INF("image decoded in %" PRId64 " ms\n", ggml_time_ms() - t1);
|
||||
|
||||
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
|
||||
}
|
||||
|
||||
n_past += n_tokens;
|
||||
while (i_batch < n_img_batches) { // split into batches
|
||||
int32_t pos_offset = i_batch*n_batch;
|
||||
int32_t n_tokens_batch = std::min(n_batch, n_tokens - pos_offset);
|
||||
float * embd_batch = embd + pos_offset*n_mmproj_embd;
|
||||
decode_embd_batch batch_img(embd_batch, n_tokens_batch, n_past, 0);
|
||||
|
||||
printf("decoding image batch %d/%d, n_tokens_batch = %d\n", i_batch+1, n_img_batches, n_tokens_batch);
|
||||
|
||||
int64_t t1 = ggml_time_ms();
|
||||
ret = llama_decode(lctx, batch_img.batch);
|
||||
if (ret != 0) {
|
||||
LOG_ERR("failed to decode image\n");
|
||||
llama_set_causal_attn(lctx, true); // restore causal attn
|
||||
llama_batch_free(text_batch);
|
||||
return ret;
|
||||
}
|
||||
|
||||
if (ctx->print_timings) {
|
||||
LOG_INF("image decoded (batch %d/%d) in %" PRId64 " ms\n", i_batch+1, n_img_batches, ggml_time_ms() - t1);
|
||||
}
|
||||
|
||||
i_batch++;
|
||||
n_past += n_tokens_batch;
|
||||
}
|
||||
|
||||
if (mtmd_decode_use_non_causal(ctx)) {
|
||||
llama_set_causal_attn(lctx, true);
|
||||
}
|
||||
|
||||
} else {
|
||||
GGML_ASSERT(false && "chunk type not supported");
|
||||
|
||||
Reference in New Issue
Block a user