Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

mtmd: support DeepSeek-V4-Flash-Vision-Exp (#28133)

Browse Source 
* mtmd: support DeepSeek-V4-Flash-Vision-Exp

* handle min/max token counts from CLI

* rm debugging

* use GGML_ROPE_TYPE_VISION

* nits

* apply review comments

* correct token count
This commit is contained in:
Xuan-Son Nguyen
2026-09-02 16:43:43 +02:00
committed by GitHub
parent 8e93a9773b
commit 7798007a29
13 changed files with 504 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/mtmd/clip-model.h
+9
View File
@@ -100,6 +100,10 @@ struct clip_hparams {
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
// deepseek4v: resize solver caps the LLM token count of the aligner grid
int32_t dsv4_max_n_token = 0;
int32_t dsv4_max_wh_ratio = 0;
// deepseek-ocr (sam)
int32_t sam_n_layer = 0;
int32_t sam_n_head = 0;
@@ -724,6 +728,11 @@ struct clip_model {
// pixtral, glm4v
ggml_tensor * token_embd_img_break = nullptr;
// deepseek4v sentinel embeddings (image_newline is reused for IMAGE_NEW_LINE)
ggml_tensor * token_embd_img_start = nullptr;
ggml_tensor * token_embd_img_end = nullptr;
ggml_tensor * token_embd_img_pad = nullptr;
ggml_tensor * mm_patch_merger_w = nullptr;
ggml_tensor * mm_patch_merger_b = nullptr;