Kimi-K3 text runs on master (#26185), but the model is natively multimodal and unsloth ships
mmproj-{BF16,F16,F32}.gguf declaring clip.projector_type = kimik3 (MoonViT-3d: image_size 896,
patch 14, 27 blocks, projection_dim 7168, scale_factor 2, head_dim 128, 3-D position embeddings).
master's CLIP loader rejects it (Failed to load CLIP model).
unslothai/llama.cpp PR #70, commit 27fd568 (by @danielhanchen), adds the projector:
tools/mtmd/models/kimik3.cpp + clip-impl.h / clip.cpp / mtmd.cpp entries, +199/-0, purely
additive. It applies cleanly to current master and works end-to-end for me on top of 749f688
(numbers in my comment on that PR: unslothai#70 (comment)).
Would you take it upstream? Happy to test.
Kimi-K3 text runs on master (#26185), but the model is natively multimodal and unsloth ships
mmproj-{BF16,F16,F32}.gguf declaring
clip.projector_type = kimik3(MoonViT-3d: image_size 896,patch 14, 27 blocks, projection_dim 7168, scale_factor 2, head_dim 128, 3-D position embeddings).
master's CLIP loader rejects it (
Failed to load CLIP model).unslothai/llama.cpp PR #70, commit 27fd568 (by @danielhanchen), adds the projector:
tools/mtmd/models/kimik3.cpp + clip-impl.h / clip.cpp / mtmd.cpp entries, +199/-0, purely
additive. It applies cleanly to current master and works end-to-end for me on top of 749f688
(numbers in my comment on that PR: unslothai#70 (comment)).
Would you take it upstream? Happy to test.