]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
mtmd: deepseek-ocr v1 multi-tile (#24717)
authorXuan-Son Nguyen <redacted>
Fri, 10 Jul 2026 14:05:49 +0000 (16:05 +0200)
committerGitHub <redacted>
Fri, 10 Jul 2026 14:05:49 +0000 (16:05 +0200)
commit3e706dd55f56f59126c053c3db9cfc219a0a7074
tree7919a986b6baa9dac9ee79e9fd6365fe15923392
parent07d937828636e305bc0cfe738b288f9ab05ff748
mtmd: deepseek-ocr v1 multi-tile (#24717)

* mtmd: deepseek-ocr v1 multi-tile dynamic resolution + unified image-preprocessors for both versions (ds-ocr v1 and v2)

* remove hacky API

* fuse row into a long image

* almost working

* adapt to new preprocessor api

* rm debugging printf

* improve

* mtmd: dsocr-tiles fixes (#25481)

* ds-ocr img-preproc fuse_row tile-drop fix for multi rows and columns images

* mtmd drop the duplicate redundant img_end

* deepseekocr graph simplify CLS broadcast cleanup

* test-deepseek-ocr: relax v1 single-view tolerance; drop trailing prompt space; make DRY opt-in and n_predict model-specific (#25486)

---------

Co-authored-by: Saba Fallah <redacted>
Co-authored-by: Saba Fallah <redacted>
tools/mtmd/clip-graph.h
tools/mtmd/clip-model.h
tools/mtmd/clip.cpp
tools/mtmd/models/deepseekocr.cpp
tools/mtmd/models/models.h
tools/mtmd/mtmd-image.cpp
tools/mtmd/mtmd-image.h
tools/mtmd/mtmd.cpp
tools/mtmd/tests/test-1-positive.png [new file with mode: 0644]
tools/mtmd/tests/test-deepseek-ocr.py