Syed Tousif Ahmed comments

Results 26 comments of


                                            Syed Tousif Ahmed

`ap_hs` doesn't work with `hls::stream`

Hi @dhananjays, I don't have a workaround for getting the exact signals that `ap_hs` used to infer. We ended up changing our interfaces to `axis`. So that would mean you...

Implements user buffer registration using MemPool

Thanks @kwen2501 ! Addressed all the review. Will merge after CI is green and posting a local run of the test in a multicast supported machine.

Implements user buffer registration using MemPool

Ran test in 8xH100 system with nvswitch (supports multicast): `python test/distributed/test_c10d_nccl.py -k test_nccl_user_buffer_registration -v` ``` INFO:numba.cuda.cudadrv.driver:init Detected CUDA files, patching ldflags Emitting ninja build file /root/.cache/torch_extensions/py312_cu126/nccl_allocator/build.ninja... Building extension module nccl_allocator......

Implements user buffer registration using MemPool

@pytorchbot merge

Implements user buffer registration using MemPool

@pytorchbot merge

Implements user buffer registration using MemPool

@pytorchbot merge

Fixes MX formats build for blackwell

> This meeans we end up shipping 3 more .so s right? I wonder if ther isn't a better way to do this by including the files in the extension...

Fixes MX formats build for blackwell

Sounds good! Thanks.

Fixes MX formats build for blackwell

Closing, has been merged in https://github.com/pytorch/ao/pull/2278.

Add initial support for torch.utils.checkpoint

@IvanYashchuk You might wanna checkout selective activation checkpointing available in PyTorch nightlies: https://pytorch.org/docs/main/checkpoint.html#torch.utils.checkpoint.create_selective_checkpoint_contexts to specify which activations to save for backward.