Repository navigation
model: support embeddinggemma2 (text+vision+audio) - #30054
Merged
Merged
Conversation
github-actions
Bot
added
model
danbev
approved these changes
Oct 6, 2026
CISC
approved these changes
Oct 6, 2026
CISC
added
the
highlight
imaami
pushed a commit
to imaami/llama.cpp
that referenced
this pull request
Oct 6, 2026
(cherry picked from commit 4fbc76d)
4 of 5 tasks
This was referenced Oct 7, 2026
6 tasks done
gagallo7
pushed a commit
to gagallo7/qvac-fabric-llm.cpp
that referenced
this pull request
Oct 8, 2026
[qvac-b11259: scale tokens after build_inp_embd, this base has no tok_scale parameter (ggml-org#29622)] Assisted-by: Claude Code (cherry picked from commit 4fbc76d)
edwardyoon
pushed a commit
to edwardyoon/focus-llama
that referenced
this pull request
Oct 8, 2026
(cherry picked from commit 4fbc76d)
CVeniamin
added a commit
to CVeniamin/ollama
that referenced
this pull request
Oct 9, 2026
The pinned llama.cpp (b11351) predates ggml-org/llama.cpp#30054, which adds the gemma-embedding2 architecture for EmbeddingGemma 2. Carry its graph as an Ollama-owned source plus a small registration patch, the way Laguna was carried. The only change from upstream is the input embedding scale, written with the pinned build_inp_embd API as gemma4.cpp does. Remove both files once LLAMA_CPP_VERSION includes that change.
aranlucas
pushed a commit
to aranlucas/LokalBot
that referenced
this pull request
Oct 9, 2026
Pins the source build that adds EmbeddingGemma 2 support (ggml-org/llama.cpp#30054). No v-tag after v0.6.0 contains it yet. The source archive has no .git, so pass the build number explicitly and --version reports build 11474.
brianreborn
added a commit
to brianreborn/familia
that referenced
this pull request
Oct 10, 2026
graph.yaml declares runtimes (bin, tag, commit, load-tested archs). Each node names its runtime; the validator checks the binary reports the declared commit and fails if the node's GGUF arch was not verified on that runtime. Coder stays on b11374; embeddinggemma-2 (gemma-embedding2, unknown to b11374) runs on official b11539 (includes ggml-org/llama.cpp#30054), port 9951, --embeddings. Tested on miryam: 768-dim embeddings from the graph-rendered command while coder ran (RAM used 5.7/7.1 GiB).
3 tasks
46 tasks done
thelittlefireman
pushed a commit
to thelittlefireman/llama.cpp
that referenced
this pull request
Oct 12, 2026
28 tasks done
5 tasks
9 tasks
Leo00703
added a commit
to Leo00703/Local-LLM
that referenced
this pull request
Oct 19, 2026
Submodule (app/src/main/cpp/llama.cpp): - dbe4c3ed4 (b11393) -> abeada335 (b11455), the first tag that contains ggml-org/llama.cpp#30054 "model: support embeddinggemma2 (text+vision+audio)" (merged 2026-10-06, commit 4fbc76de). Found by comparing the merge commit against tags with the GitHub compare API; b11450 does not contain it, b11455 does. The newest tag (b11736) is 343 commits ahead and changes chat.h/common.h/sampling.h/llama.h; b11455 is 62 commits ahead and was chosen to keep the first bump small. Public API diff b11393 -> b11455 (headers the app includes), read in full: - include/llama.h: LLAMA_SESSION_VERSION 10 -> 11, LLAMA_STATE_SEQ_VERSION 3 -> 4, new LLAMA_VOCAB_TYPE_PLAMO3. Additive. - common/common.h: tty_enable_ansi(), COMMON_DECISION_TYPE_PPLX_DECIDER, a filename overload of common_get_decision_type. common/log.h: common_log_get_colors. Additive. - ggml-backend.h: ggml_backend_sched_set_copy_callback. Additive. - chat.h, sampling.h, speculative.h, mtmd.h, mtmd-helper.h: unchanged. No native glue change was needed apart from the skip_audio parameter below. - Behaviour: the persisted preamble KV files (llama_state_seq_save_file) change format with STATE_SEQ_VERSION 4; LlamaGenerationSession already treats a failed load as a miss, clears the KV and regenerates, so old cache files are rebuilt once. - Qwen3.5 MTP graph now goes through crop_before/after_nextn and ASSERT_EMBD_OR_TOKEN (#29622, #30020); same tag, same speculative.cpp, but MTP was not exercised on a device here. Patches: - 0001 and 0002 verified to apply to b11455 (git apply --cached --check against a temporary index, then the real apply and reverse in the submodule). - New 0003-mtmd-add-skip_audio-context-param: mtmd_context_params.skip_audio and clip_context_params.skip_audio (default false = upstream), the Gemma 3n special case kept. Replaces the CI sed that forced skip_audio = true for every caller. That sed was written for b9496, where the Gemma 4 audio projector types were unknown and clip hit GGML_ABORT("Unknown projector type"); PROJECTOR_TYPE_GEMMA4A / GEMMA4UA are handled in b11455 (clip.cpp), so the abort it avoided is gone. It also would have blocked the audio encoder of EmbeddingGemma 2 for the upcoming indexing feature. - LlamaModel.cpp sets mparams.skip_audio = true, so chat behaviour is unchanged. - The "Patch mtmd - skip audio encoder" step is removed from build-debug-apk.yml, pull-request-check.yml and deploy-internal.yml; the generic "Apply Android patches" loop already picks up 0003. Docs: AGENTS.md (pin, patch list, how to check patches without a checkout, trailer now Sonnet 5.5) and patches/README.md. Changelog 1.9.133 entry. Not verified: compiles only in CI (no NDK locally); not run on a device. Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Overview
Support https://huggingface.co/google/embeddinggemma-2
GGUF: https://huggingface.co/ggml-org/embeddinggemma-2-GGUF
Requirements