Jiewen Tan comments

Results 60 comments of


                                            Jiewen Tan

[Pallas] Introduce make_kernel_from_pallas

> Do you need this pr in 2.3? Yea, will also need a couple for the TODOs.

[Pallas] Introduce make_kernel_from_pallas

Can I get any reviews?

[Pallas] Introduce make_kernel_from_pallas

> I still think we should refactor `convert_torch_dtype_to_jax` and invesgate bf16(which I assume most people will use), approve to unblock. Yea, for sure. Let me follow up with that.

Exception on TPU when compiling gemma 2b

@JackCaoG No, I don't think there is anyone working on it at this moment.

Gemma finetuning on Kaggle TPU doesn't work

@windmaple You need to install the nightly torch-xla and torch.

Exception on TPU when compiling gemma 2b

@ManfeiBai Can you take a look?

Gemma finetuning on Kaggle TPU doesn't work

@windmaple Here is the instructions to install nightly: https://github.com/pytorch/xla#available-docker-images-and-wheels

Exception on TPU when compiling gemma 2b

@ManfeiBai Can you try reproducing it?

Gemma finetuning on Kaggle TPU doesn't work

@PawKanarek What's your libtpu version?

Gemma finetuning on Kaggle TPU doesn't work

@windmaple Yea, usually you just need nightly for both pytorch and pytorch/xla. pytorch/xla heavily depends on pytorch.