- The feedforward hidden size is 16x the d_model, unlike most models which are typically 4x;
- The vocabulary size is 10x (256K vs. Mistral’s 32K);
- The training token count is tripled (6T vs. Llama2's 2T)
Apart from that, it uses the classic transformer variations: MQA, RoPE, RMSNorm.
How big was the batch size that it could be trained so fast?
https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.2/bl...
Damn, 6T? That's a lot!
Given that this model seems to roughly match Mistral (according to the numbers from Google), this makes me think we have saturated the 7B parameter space, and couldn't possibly make it much better unless new techniques are discovered.
[0]:https://huggingface.co/google/gemma-7b-it/blob/main/config.j...
[1]: https://storage.googleapis.com/deepmind-media/gemma/gemma-re...
EDIT: I didn't read the comment correctly, you have noticed the same thing.
In the report, "hidden dim" matches the number of GEGLU inputs. In the config, "intermediate_size" matches the number of GEGLU outputs. Most *GLU models so far have used intermediate_size=8/3*d_model as this makes have the same number of matmul FLOPS & parameters as a 4x-expanded non-GLU model, and PaLM vaguely showed that 4x is better than a smaller expansion factor.
If one considers Llama-2-7B's FFN expansion factor to be ~5.33x, Gemma's expansion factor is 16x.
You can use this playground to test it out: https://huggingface.co/spaces/Xenova/the-tokenizer-playgroun...
Doesn't seem like a well made tokenizer at first glance or it's heavily biased towards languages the model can't even generate coherently, lol. If they really wanted it to be SOTA at something they could've at least made it the first open source truly multilingual model, but that's apparently more effort than the lame skin colour oriented virtue signalling Google wants to do.