[gemma_cpp] Add support for Qwen3 models (0.6B, 1.7B and 4B). - #964
Merged
Conversation
copybara-service
Bot
force-pushed
the
test_952434363
branch
3 times, most recently
from
July 27, 2026 23:55
7474b01 to
93d09d3
Compare
Qwen3 shares a very similar architecture to Gemma3 ([comparison](https://sebastianraschka.com/llm-architecture-gallery/?compare=qwen3-0-6b%2Cgemma-3-270m#architecture-diff-tool)). Differences include: - *Tokenizer* : Qwen3 uses bytelevel BPE while Gemma3 uses `Sentencepiece`; - *Embedding* : Qwen3 0.6B and 1.7B do not share the `input_embedding` and `lm_head` weights, and there is no Embedding Scaling for Qwen3; - *Activation* : Qwen3 uses SiLU while gemma3 uses GeLU; - *Norm* : Qwen3 only has two pre-norms, and they are not Zero-centered (we substract 1.0 while converting the weights to keep the kernel untouched); - *Attention* : Qwen3 uses full-attention for all the layers. - *Prompt Wrapping* : Qwen3 uses different special tokens and does not require <BOS>. PiperOrigin-RevId: 954918486
copybara-service
Bot
force-pushed
the
test_952434363
branch
from
July 28, 2026 00:22
93d09d3 to
891e43d
Compare
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
[gemma_cpp] Add support for Qwen3 models (0.6B, 1.7B and 4B).
Qwen3 shares a very similar architecture to Gemma3 (comparison).
Differences include:
Sentencepiece;input_embeddingandlm_headweights, and there is no Embedding Scaling for Qwen3;