Add Gemma3 architecture (text-only) #711
Merged
Conversation
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
…nario (#689) Signed-off-by: Liu, Kaixuan <[email protected]>
Signed-off-by: Liu, Kaixuan <[email protected]>
Signed-off-by: Liu, Kaixuan <[email protected]>
Some models contain Dense layers after the Transformer + Pooling, which
means that the ONNX export might already contain the sentence embeddings
given that the export needs to come with Transformer + Pooling + Dense.
This commit adds the `outputs.contains_key("sentence_embedding")`
handling to overcome that.
Otherwise it will default to `DType::Float16` when `--features candle-cuda`
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
What does this PR do?
This PR adds support for the Gemma3 architecture only for text, running on FP32 precision on CPU, MPS and CUDA devices.
Note
Neither FP16 nor Flash Attention are supported for Gemma3, given that FP16 is not supported due to weight instability leading to
nanvalues in the hidden states.Additionally, this PR also adds support for the ONNX export, which contains the embeddings at
sentence_embeddingsrather than thelast_hidden_statesfrom the Transformer layer, as some models might contain Transformer + Pooling + Dense modules, meaning that the ONNX export comes with all those packed together, unlike standard Transformer + Dense ONNX exports that output thelast_hidden_statesonly so that the pooling is applied separately afterwards.Before submitting
Who can review?
@Narsil