Vadim Kantorov

Maybe also worth allowing passing these predefined scale factors directly `1/2`, `1/4`, `1/8` instead of `size_hint`

@SeanNaren @titu1994 An option of using SDPA is a good thing also because a Triton-based version of FAv2 with custom `attn_bias` support (FlexAttention) is being added into PyTorch core: https://github.com/pytorch/pytorch/pull/130250#issuecomment-2216318222,...

@SeanNaren @titu1994 haha, and now that FAv3 is out, probably PyTorch would integrate it as well in some near term - for maximum brr on H100 :) so having Nemo's...

stale bump

@pzelasko I think the first step should be running SDPA with the most basic `math` algo selected: ```python with torch.backends.cuda.sdpa_kernel(backends=[SDPBackend.MATH]): # needs to be checked if this breaks torch.compile(fullgraph=True) or...

Okay, I figured out how to get nice names for TexLive ISO files. I adapted the `auto` mode from https://github.com/clalancette/pycdlib/blob/master/tools/pycdlib-extract-files. Maybe it would be nice if such `auto` mode was...

Yeah, I've got the same question. E.g. if I have `.tar.xz` files nested inside `.iso`. Does libarchive have some builtin mechanism for automatic decompression of this nested structure? @kientzle Or...

The self-contained GSoC repo by @diyang seems to be in https://github.com/hiDiYANG/gPb-GSoC/

New `--channels discrete` can be used to force uncoupled encoding, but it would still be nice to have a mode for super-storage-space-efficient silent frame encoding (using VAD or by passing...

Related on DTX: - https://github.com/xiph/opus-tools/issues/89