llama.cpp
mtmd: add granite-speech support (ibm-granite/granite-4.0-1b-speech)
#22101
Merged
Merged
mtmd: add granite-speech support (ibm-granite/granite-4.0-1b-speech) #22101
CISC
merged
21 commits into
ggml-org:master
from
ReinforcedKnowledge:add-granite-speech-support
gabe-l-hart
approved these changes
on 2026-04-27
ngxson
approved these changes
on 2026-05-02
CISC
approved these changes
on 2026-05-03
mtmd: add granite-speech support (ibm-granite/granite-4.0-1b-speech)
2131fa19
mtmd: rename gs_ prefixed tensors to generic/architecture names
2d390ce6
mtmd: use tensor_mapping.py for all granite_speech tensors
189d7347
convert: fold GraniteSpeechTextModel into GraniteModel
c249df50
mtmd: replace n_layer hack with explicit has_standard_layers flag
38ce1efe
mtmd: replace hardcoded magic numbers with GGUF hparams for granite s…
8778a066
mtmd: align KEY_A_ define spacing
b8244e4a
convert: register GraniteModel for GraniteSpeechForConditionalGeneration
cd63cc85
convert: fix ty type-check for GraniteSpeechMmprojModel registration
a3a3d195
mtmd: align TN_ define spacing
5faaaeff
mtmd: use generic layer loop for granite speech tensor loading
0735b46b
mtmd: merge qformer_proj_layer into clip_layer
c1c261c1
mtmd: granite_speech remove redundant ggml_build_forward_expand on in…
11e5a13e
mtmd: granite_speech add comment explaining why build_attn is not used
b241b4e8
mtmd: granite_speech hard-code eps in cpp, remove from GGUF metadata
0665d9ea
gguf: add spacing between granite_speech tensor mapping blocks
f33e45b9
mtmd: make generic audio layer_norm_eps read optional
4936492f
mtmd: granite_speech keep encoder eps in GGUF, only hard-code project…
54d32353
mtmd: align defines and struct fields in clip-impl.h and clip-model.h
13aeeb5a
mtmd: fix alignment and ordering issues across granite speech files
8b9c41c7
convert: granite_speech use filter_tensors instead of modify_tensors …
94606514
CISC
approved these changes
on 2026-05-06
Reviewers
Assignees
No one assigned
Labels
examples
python
Milestone
No milestone
Login to write a write a comment.
Login via GitHub