llama.cpp
Add Granite Vision Support
#11794
Merged

Add Granite Vision Support #11794

alex-jw-brooks
github-actions github-actions added examples
github-actions github-actions added python
github-actions github-actions added ggml
alex-jw-brooks
alex-jw-brooks commented on 2025-02-10
gabe-l-hart
gabe-l-hart commented on 2025-02-10
gabe-l-hart
gabe-l-hart commented on 2025-02-12
alex-jw-brooks alex-jw-brooks marked this pull request as ready for review 307 days ago
alex-jw-brooks alex-jw-brooks requested a review from ngxson ngxson 307 days ago
alex-jw-brooks alex-jw-brooks requested a review from gabe-l-hart gabe-l-hart 307 days ago
alex-jw-brooks
gabe-l-hart
ggerganov
ggerganov commented on 2025-02-20
ggerganov
danbev
alex-jw-brooks
ngxson
ngxson commented on 2025-02-20
ngxson
alex-jw-brooks
danbev
alex-jw-brooks alex-jw-brooks force pushed 300 days ago
alex-jw-brooks alex-jw-brooks requested a review from ngxson ngxson 299 days ago
alex-jw-brooks alex-jw-brooks requested a review from ggerganov ggerganov 299 days ago
alex-jw-brooks
danbev
ngxson
ngxson commented on 2025-02-23
alex-jw-brooks Add super wip scripts for multimodal granite gguf
7db1c510
alex-jw-brooks Add example for converting mmgranite to gguf
f6fbfc25
alex-jw-brooks remove hardcoded path
8b601070
alex-jw-brooks Add vision feature layer to gguf params
72c53dd8
alex-jw-brooks Clean up llava surgery and remove name substitution hacks
4212f373
alex-jw-brooks Add transformers llava next tensor name mapping
77ce6f2d
alex-jw-brooks Make siglip / openclip mutuall exclusive
3fc67a9f
alex-jw-brooks Fix projector linear substitution
e3c791c6
alex-jw-brooks Fix linear 2 substitution index
2d45e0d0
alex-jw-brooks Increase max flattened gridpoints to 64
3e3eebce
alex-jw-brooks Fix hardcoded concat for multiple feature layers
1ceef1a7
alex-jw-brooks Pull vision feature layers out of gguf keys
c788a455
alex-jw-brooks fix num gridpoints and use all layers
6725d6c2
alex-jw-brooks Avoid dropping last image encoder layer in llava models
403575cc
alex-jw-brooks Use 10 for max number of patches
095b8362
alex-jw-brooks Standardize vision feature layers
ff00515c
alex-jw-brooks Cleanup logs
eceee7fe
alex-jw-brooks Update comment for vision feature layer init
be204f0f
alex-jw-brooks Update notes for alternative to legacy llm conversion script
4588b901
alex-jw-brooks Fix notes rendering
264c2ca2
alex-jw-brooks Add v prefix to vision feature layer log
86b43da3
alex-jw-brooks Use current defaults for feature layer
eb545402
alex-jw-brooks Use constant for max gridpoints / feat layers, style fixes
243a8991
alex-jw-brooks clarify non-negative feature layers
ee2cf62b
alex-jw-brooks Remove CLIP_API from func signature
b5735ba3
alex-jw-brooks USE MAX_IMAGE_FEATURE_LAYERS const in layer calc
3670d0ed
alex-jw-brooks Clarify feature layers are non negative ints and not uint
b973f37b
alex-jw-brooks Fix condition for reading feature layers
82e88527
alex-jw-brooks pop last llava layer when feature layers are unset
ee6fb4de
alex-jw-brooks Fix unset vision layer 0
07e77167
alex-jw-brooks Update examples/llava/clip.cpp
5f69fdbd
alex-jw-brooks Reenable assertion for out of bounds get_rows
ab522d79
alex-jw-brooks Use std vector for gridpoints and feature layers
bb3e03a8
alex-jw-brooks Caculate max feature layer at load time
7bab3057
alex-jw-brooks Include base patch for granite vision allocation
65575237
alex-jw-brooks Fix trailing whitespace
cce01b89
alex-jw-brooks Add max num patches = 10 back for minicpmv
188bfb0e
alex-jw-brooks Use unordered set to store feature layers
86763161
alex-jw-brooks alex-jw-brooks force pushed to 86763161 296 days ago
alex-jw-brooks Use max feature layer for postnorm
49c08636
alex-jw-brooks alex-jw-brooks requested a review from ngxson ngxson 296 days ago
alex-jw-brooks
ngxson
ngxson approved these changes on 2025-02-24
ngxson Apply suggestions from code review
bec9ef1f
ngxson
alex-jw-brooks
ngxson ngxson merged 7a2c913e into master 296 days ago

Login to write a write a comment.

Login via GitHub

Assignees
No one assigned
Labels
Milestone