mtmd : add vision support for llama 4 #13282
ngxson
merged
32 commits into
ggml-org:master
from
ngxson:xsn/mtmd_llama4_new
wip llama 4 conversion
c912c674
rm redundant __init__
a67a1bed
fix conversion
10db58b5
Merge branch 'master' into xsn/mtmd_llama4_new
55ad3a63
fix conversion
c50e627b
test impl
8775bc4e
try this
7341e709
reshape patch_embeddings_0
893ad9c2
fix view
15605e46
rm ffn_post_norm
32a62d1f
cgraph ok
97a5cd13
f32 for pos embd
c6c2d66f
add image marker tokens
224cbbad
Llama4UnfoldConvolution
9d1a4d65
correct pixel shuffle
532c3321
Merge branch 'master' into xsn/mtmd_llama4_new
844a3449
fix merge conflicts
8caeed5c
correct
2ffafd5d
add debug_graph
7d9d4e34
logits matched, but it still preceives the image incorrectly
919318e3
fix style
5b81972e
add image_grid_pinpoints
b74122f0
handle llama 4 preprocessing
4217d424
rm load_image_size
3645fe0b
rm unused line
791f1069
fix
8646240a
small fix 2
53fb6223
add test & docs
f083c192
ngxson
changed the title
fix llava-1.6 test
b199e70d
test: add notion of huge models
e52481ba
ggerganov
commented
on 2025-05-19
ggerganov
commented
on 2025-05-19
ngxson
commented
on 2025-05-19
ggerganov
approved these changes
on 2025-05-19
add comment
186d7a88
add warn about degraded quality
d5e50aa0
Reviewers
Assignees
No one assigned
Labels
documentation
help wanted
examples
python
Milestone
No milestone
Login to write a write a comment.
Login via GitHub