RuntimeError: at::cuda::blas::gemm: not implemented for struct c10::BFloat16

[dhchenx]

RuntimeError: at::cuda::blas::gemm: not implemented for struct c10::BFloat16

[pengchongjin]

What's your GPU type? It might be the case that you are running it on a GPU that does not support bfloat16 type.

As a work-around, you can try to change here to

dtype: str = 'float16' # or 'float32'

and test to see if that works.

[dhchenx]

What's your GPU type? It might be the case that you are running it on a GPU that does not support bfloat16 type.

As a work-around, you can try to change here to

dtype: str = 'float16' # or 'float32'

and test to see if that works.

Thanks, I used NVIDIA Quadro P2200 (5 GB) .

I first tried to change dtype:str='float16', the error is below:

Then, I tried to change dtype:str='float32', the another error is shown below:

I wonder if the example code can run on low computing capacity computers. Any solution? Thanks!

[pengchongjin]

Which model variant are you trying? 2B or 7B or 7B-quant? In any case, 5GB might be insufficient.

[dhchenx]

Which model variant are you trying? 2B or 7B or 7B-quant?

I used 2B-it

[pengchongjin]

Yeah, 2B f16 may still be too large to fit 5GB card. Maybe try CPU first.

[dhchenx]

Yeah, 2B f16 may still be too large to fit 5GB card. Maybe try CPU first.

Thanks. I try CPU (32GB) and it works well.

[pengchongjin]

OK, yeah, to fit 5GB card, we probably would need a int8 quantized 2B variant, which is not included in the release.

You can try the free Colab T4 runtime using this Colab, and it should work.
https://ai.google.dev/gemma/docs/pytorch_gemma

[dhchenx]

OK, yeah, to fit 5GB card, we probably would need a int8 quantized 2B variant, which is not included in the release.

You can try the free Colab T4 runtime using this Colab, and it should work. https://ai.google.dev/gemma/docs/pytorch_gemma

Thanks!

[pengchongjin]

Closing the issue for now.