vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-07-19 05:43:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

gemma4uv vision still a little buggy

Browse source
This commit is contained in:
Concedo 2026-06-05 00:53:23 +08:00
parent f997d23660
commit 2cfcd40fd4
1 changed files with 20 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

22
gpttype_adapter.cpp
View file

@ -127,6 +127,7 @@ static std::string media_composite_image_signature = ""; //for identifying when
static int current_media_identifier = MEDIA_TOKEN_IDENTIFIER_A;
static int vision_max_res = 2048;
static bool use_mrope = false;
static bool vision_use_non_casual = false;
static kcpp_params * kcpp_data = nullptr;
static int max_context_limit_at_load = 0;
@ -2007,7 +2008,10 @@ static bool kcpp_eval_media(llama_context * ctx_llama, const media_chunk & media
const int image_n_past = *n_past;
kcpp_embd_batch media_batch = kcpp_embd_batch(img_embd, num_img_tokens, image_n_past, use_mrope, is2d, img_nx, img_ny);
if(vision_use_non_casual)
{
llama_set_causal_attn(llama_ctx_v4, false);
}
for (int i = 0; i < num_img_tokens; i += n_batch) {
const int n_eval = std::min(n_batch, num_img_tokens - i);
llama_batch batch_embd_view = media_batch.get_view(i, n_eval, n_embd_mmproj);
@ -2016,6 +2020,10 @@ static bool kcpp_eval_media(llama_context * ctx_llama, const media_chunk & media
return false;
}
}
if(vision_use_non_casual)
{
llama_set_causal_attn(llama_ctx_v4, true);
}
*n_past += num_img_tokens;
return true;
}
@ -3022,6 +3030,16 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
if(clp_ctx_v)
{
vision_multimodal_supported = true;
int proj_type = clip_get_projector_type_ext(clp_ctx_v);
switch (proj_type) {
case PROJECTOR_TYPE_GEMMA3:
case PROJECTOR_TYPE_GEMMA4V:
case PROJECTOR_TYPE_GEMMA4UV:
vision_use_non_casual = true;
break;
default:
break;
}
}
clp_img_data = clip_image_u8_init();
if(clp_ctx_a) //init audio
@ -4833,7 +4851,7 @@ generation_outputs gpttype_generate(const generation_inputs inputs)
img_start = "<|begin_of_image|>";
img_end = "<|end_of_image|>\n\n";
}
else if(ptype==PROJECTOR_TYPE_GEMMA4V)
else if(ptype==PROJECTOR_TYPE_GEMMA4V || ptype==PROJECTOR_TYPE_GEMMA4UV)
{
img_start = "<|image>";
img_end = "<image|>\n\n";