nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 18 Packages Projects Releases Wiki Activity

mtmd: add MiMo v2.5 vision (#22883)

Browse Source 
* mimo-v2.5: vision support

* mimo-v2.5: use fused qkv for vision

* mimi-v2.5: fix f16 vision overflow

* mimo-v2.5: comment cleanups

* mimo-v2.5: Flash doesn't have mmproj
more cleanup
remember to use filter_tensors

* mimo-v2.5: fix trailing whitespace
This commit is contained in:
AesSedai
2026-05-12 11:11:14 +02:00
committed by GitHub
parent 78fbbc2c07
commit 4178259130
12 changed files with 460 additions and 23 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/clip-model.h
+4
View File
@@ -42,6 +42,7 @@ struct clip_hparams {
int32_t n_ff = 0;
int32_t projection_dim = 0;
int32_t n_head = 0;
int32_t n_head_kv = 0;
int32_t n_layer = 0;
// idefics3
int32_t n_merge = 0; // number of patch merges **per-side**
@@ -83,6 +84,7 @@ struct clip_hparams {
int32_t attn_window_size = 0;
int32_t n_wa_pattern = 0;
std::unordered_set<int32_t> wa_layer_indexes; // explicit layer indexes that use full attention (for irregular patterns like YoutuVL)
std::vector<int32_t> wa_pattern_mode; // mimovl: per-layer window-attention mode
// deepseek-ocr (sam)
int32_t sam_n_layer = 0;
@@ -166,6 +168,8 @@ struct clip_layer {
ggml_tensor * o_w = nullptr;
ggml_tensor * o_b = nullptr;
ggml_tensor * attn_sinks = nullptr;
ggml_tensor * k_norm = nullptr;
ggml_tensor * q_norm = nullptr;