Port mdmd from mainline + Qwen2/2.5-VL support#798

Merged

ikawrakow merged 21 commits intomainfrom

Sep 27, 2025

Commits on Sep 26, 2025

Add mtmd: the beginning
Iwan Kawrakow
committed
Add mtmd: mtmd.cpp compiles
Iwan Kawrakow
committed
Add mtmd: clip initialization compiles
Iwan Kawrakow
committed
Add mtmd: clip.cpp compiles
Iwan Kawrakow
committed
Add mtmd: builds successfully
Iwan Kawrakow
committed
Add CPU implementation for GGML_OP_GLU
Iwan Kawrakow
committed
Add CUDA implementation for GGML_OP_GLU
Iwan Kawrakow
committed
Add CPU implementation for GGML_OP_CONV_2D and GGML_OP_CONV_2D_DW
Iwan Kawrakow
committed
Add CUDA implementation for GGML_OP_CONV_2D and GGML_OP_CONV_2D_DW
Iwan Kawrakow
committed
Add mtmd: refresh CPU rope
Iwan Kawrakow
committed
Add mtmd: refresh CUDA rope
Iwan Kawrakow
committed
Add mtmd: add Qwen2-VL
Iwan Kawrakow
committed
Add mtmd: Qwen2.5-VL text seems to work with this change
Iwan Kawrakow
committed
Add mtmd: fix swiglu
Iwan Kawrakow
committed
Add mtmd: use LOG_TEE so generated tokens show up in terminal
Iwan Kawrakow
committed
Add mtmd: do not attempt to load a GPU backend if none are available
Iwan Kawrakow
committed
GLU, not GPU
Iwan Kawrakow
committed
Fix typo
Iwan Kawrakow
committed
Fix new/free mismatch
Iwan Kawrakow
committed
LOG stuff
Iwan Kawrakow
committed
Add mtmd: this fixes gibberish on second image
Iwan Kawrakow
committed