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/mtmd-image.h
+16
View File
@@ -129,6 +129,22 @@ struct mtmd_image_preprocessor_longest_edge : mtmd_image_preprocessor {
mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
};
// ref: inference/image_processor.py in the HF repo (DeepSeek-V4-Flash-Vision)
struct mtmd_image_preprocessor_deepseek4v : mtmd_image_preprocessor {
mtmd_image_preprocessor_deepseek4v(const clip_ctx * ctx) : mtmd_image_preprocessor(ctx) {}
mtmd_image_preproc_out preprocess(const clip_image_u8 & img) override;
private:
struct grid_info {
int n_llm_h;
int n_llm_w;
int n_tokens; // token count of the block (incl. newline/pad rows and start/end, excl. lead pads)
};
static grid_info grid_tokens(int best_height, int best_width, int patch_size, int r);
static void solve_resize_ratio(int height, int width, int p, int r, int max_n_token, int & best_height, int & best_width);
static void safe_resize(int height, int width, int & best_height, int & best_width, int p, int r, int max_n_token);
};
// custom llava-uhd slicing logic for MiniCPM-V
struct mtmd_image_preprocessor_minicpmv : mtmd_image_preprocessor_llava_uhd {
using mtmd_image_preprocessor_llava_uhd::mtmd_image_preprocessor_llava_uhd;