Baptiste Jamin
Baptiste Jamin
Sorry about the typo. Int8 indeed. So here at the tests with Flan T5 XXL: `pip install ctranslate2-3.13.0-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl` With `ct2-transformers-converter --model google/flan-t5-xxl --output_dir flan-xxl-c2` `Who is Barrack Obama?:` int8: `us...
I can confirm it works fine with Flan XL int8 with the latest build However, Flan XXL int8 it still returns non-sensical text, and it for the same prompt, the...
What’s your prompt and parameters ?Sent from my iPhoneOn 15 Jun 2023, at 03:55, Bradley Fox ***@***.***> wrote: I've been working with XXL in int8 with 3.15.1 and it appears...
Are there plans to a real multicast method? IMO, SendMulticast and SendEachForMulticast are not doing the same thing at all. SendEachForMulticast does 1 request per registration_ids, where SendMulticast sends 1...
> The `ChatActivity` is completely transparent/translucent but a dim is applied to show the calling activity behind to mimic a BottomSheet behavior as you could have implemented in your own...
Flan T5 has some weight on FP32 for large variants (XL and XXL). Transformers lib has a patch for this: https://github.com/huggingface/transformers/commit/b9b70b0e66694cec4a1f4429335f335592688189 This could be related. Also, anyone here tried to...
So far only this project works successfully: https://bellard.org/ts_server/ts_server.html
There is a very high chance it is related to this problem: https://github.com/huggingface/transformers/issues/20287#issuecomment-1342219429 The `wo` module needs to stay in fp32
Hey there! Just merged a combined version of different PRs for this case under version `v1.6.1`
Can you provide a bigger extract of this email? I see two things here: - Adding a regex for Sent from my (.*), (.*) smartphone - Updating the Original message...