Peter Hawkins comments

Results 291 comments of


                                            Peter Hawkins

ValueError: DenseElementsAttr could not be constructed from the given buffer. This may mean that the Python buffer layout does not match that MLIR expected layout and is a bug.

(I hope this is fixed in JAX head at this point. Please let me know if it isn't...)

ValueError: DenseElementsAttr could not be constructed from the given buffer. This may mean that the Python buffer layout does not match that MLIR expected layout and is a bug.

Note that change has already been made at JAX head.

`hal.command_buffer.fill_buffer` error for bfloat16 type

Yes and no. With an updated repro to account for MHLO changes: ``` from iree.compiler import compile_str CODE = """ module { func.func @main(%arg0: tensor, %arg1: tensor) -> tensor {...

[ROCm] Disable CUDA specific test for ROCm.

Sorry this dropped off my list of things to do. Apologies! I had missed that multiprocess support isn't hooked into ROCM/RCCL (which is what `nccl_unique_id_callback` does). I'd guess that's a...

Running pytest locally -- against AttributeError: num_generated_cases

I suspect the issue is we always run `pytest tests` not `pytest`. But it seems like a reasonable thing to want to do.

Different outputs every time fft2 is called

What version of CUDA do you have installed?

Different outputs every time fft2 is called

I'm unable to reproduce this on: a) a V100 cloud machine b) a T4 cloud machine with CUDA 11.7 c) a desktop 1080 GPU with CUDA 11.4 So there must...

Different outputs every time fft2 is called

I can reproduce this with CUDA 11.1 but not with CUDA 11.2 or newer. This at least suggests a workaround you can use: update to a newer CUDA release. I'm...

Different outputs every time fft2 is called

@sudhakarsingh27 Could you please determine whether this is a known CuFFT bug and if so whether there are workarounds we might use on the JAX side to get correct behavior...

Different outputs every time fft2 is called

I assume you mean "Plans with primes larger than 127 in FFT size decomposition or FFT size being a prime number bigger than 4093 do not perform calculations on second...