starred/ollama-ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-04-18 08:13:29 +02:00
Code Issues Packages Projects Releases Wiki Activity

Update vendor ggml code to a5bb8ba4 (#13832)

Browse Source 
Co-authored-by: Daniel Hiltgen <daniel@ollama.com>
Co-authored-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Shalini Salomi Bodapati <Shalini.Salomi.Bodapati@ibm.com>
This commit is contained in:
Jeffrey Morgan
2026-02-02 17:31:59 -08:00
committed by GitHub
parent 8f4a008139
commit ef00199fb4
241 changed files with 21271 additions and 5074 deletions
Download Patch File Download Diff File Expand all files Collapse all files

39
llama/llama.cpp/tools/mtmd/mtmd.cpp vendored
View File

@@ -121,6 +121,8 @@ mtmd_context_params mtmd_context_params_default() {
/* warmup */ true,
/* image_min_tokens */ -1,
/* image_max_tokens */ -1,
/* cb_eval */ nullptr,
/* cb_eval_user_data */ nullptr,
};
return params;
}
@@ -156,8 +158,6 @@ struct mtmd_context {
bool tok_row_end_trail = false;
bool ov_img_first = false;
bool use_mrope = false; // for Qwen2VL, we need to use M-RoPE
// string template for slice image delimiters with row/col (idefics3)
std::string sli_img_start_tmpl;
@@ -188,6 +188,8 @@ struct mtmd_context {
/* image_min_tokens */ ctx_params.image_min_tokens,
/* image_max_tokens */ ctx_params.image_max_tokens,
/* warmup */ ctx_params.warmup,
/* cb_eval */ ctx_params.cb_eval,
/* cb_eval_user_data */ ctx_params.cb_eval_user_data,
};
auto res = clip_init(mmproj_fname, ctx_clip_params);
@@ -227,7 +229,6 @@ struct mtmd_context {
void init_vision() {
GGML_ASSERT(ctx_v != nullptr);
use_mrope = clip_is_mrope(ctx_v);
projector_type proj = clip_get_projector_type(ctx_v);
int minicpmv_version = clip_is_minicpmv(ctx_v);
@@ -276,7 +277,7 @@ struct mtmd_context {
}
// set boi/eoi
if (proj == PROJECTOR_TYPE_GEMMA3) {
if (proj == PROJECTOR_TYPE_GEMMA3 || proj == PROJECTOR_TYPE_GEMMA3NV) {
// <start_of_image> ... (image embeddings) ... <end_of_image>
img_beg = "<start_of_image>";
img_end = "<end_of_image>";
@@ -293,7 +294,7 @@ struct mtmd_context {
// https://github.com/huggingface/transformers/blob/1cd110c6cb6a6237614130c470e9a902dbc1a4bd/docs/source/en/model_doc/pixtral.md
img_end = "[IMG_END]";
} else if (proj == PROJECTOR_TYPE_QWEN2VL || proj == PROJECTOR_TYPE_QWEN25VL || proj == PROJECTOR_TYPE_QWEN3VL) {
} else if (proj == PROJECTOR_TYPE_QWEN2VL || proj == PROJECTOR_TYPE_QWEN25VL || proj == PROJECTOR_TYPE_QWEN3VL || proj == PROJECTOR_TYPE_YOUTUVL) {
// <|vision_start|> ... (image embeddings) ... <|vision_end|>
img_beg = "<|vision_start|>";
img_end = "<|vision_end|>";
@@ -339,8 +340,13 @@ struct mtmd_context {
case PROJECTOR_TYPE_QWEN25O:
case PROJECTOR_TYPE_ULTRAVOX:
case PROJECTOR_TYPE_VOXTRAL:
case PROJECTOR_TYPE_GLMA:
case PROJECTOR_TYPE_MUSIC_FLAMINGO:
audio_preproc = std::make_unique<mtmd_audio_preprocessor_whisper>(ctx_a);
break;
case PROJECTOR_TYPE_LFM2A:
audio_preproc = std::make_unique<mtmd_audio_preprocessor_conformer>(ctx_a);
break;
default:
GGML_ABORT("unsupported audio projector type");
}
@@ -358,6 +364,9 @@ struct mtmd_context {
// [BEGIN_AUDIO] ... (embeddings) ...
aud_beg = "[BEGIN_AUDIO]";
} else if (proj == PROJECTOR_TYPE_MUSIC_FLAMINGO) {
// <sound> ... (embeddings) ...
aud_beg = "<sound>";
}
}
@@ -629,7 +638,7 @@ struct mtmd_tokenizer {
}
mtmd_image_tokens_ptr image_tokens(new mtmd_image_tokens);
if (ctx->use_mrope) {
if (mtmd_decode_use_mrope(ctx)) {
// for Qwen2VL, we need this information for M-RoPE decoding positions
image_tokens->nx = clip_n_output_tokens_x(ctx->ctx_v, batch_f32.entries[0].get());
image_tokens->ny = clip_n_output_tokens_y(ctx->ctx_v, batch_f32.entries[0].get());
@@ -864,14 +873,24 @@ float * mtmd_get_output_embd(mtmd_context * ctx) {
}
bool mtmd_decode_use_non_causal(mtmd_context * ctx) {
if (ctx->ctx_v && clip_get_projector_type(ctx->ctx_v) == PROJECTOR_TYPE_GEMMA3) {
return true;
switch (ctx->proj_type_v()) {
case PROJECTOR_TYPE_GEMMA3:
return true;
default:
return false;
}
return false;
}
bool mtmd_decode_use_mrope(mtmd_context * ctx) {
return ctx->use_mrope;
switch (ctx->proj_type_v()) {
case PROJECTOR_TYPE_QWEN2VL:
case PROJECTOR_TYPE_QWEN25VL:
case PROJECTOR_TYPE_QWEN3VL:
case PROJECTOR_TYPE_GLM4V:
return true;
default:
return false;
}
}
bool mtmd_support_vision(mtmd_context * ctx) {