Vadim Kantorov
Vadim Kantorov
More 2-bit packing related to quantization: https://github.com/pytorch/pytorch/pull/43077
4-bit tensor dtype: https://github.com/pytorch/pytorch/pull/44574
torch.sum and torch.count_nonzero for bool types should be equivalent. is it the case? Do they allocate any intermediate tensors and upcast? In the ideal case, the should also not upcast...
@vak maybe file a separate issue about this
> For whatever reason the default image maxWidth in the playground is 500px but the default image maxWidth when importing from markdown in the playground is 800px. Thank you! This...
I agree that that there are workaround (and thanks for bringing attention to `copyto`!), but many other regular functions already have `out=` argument, so having it supported also for often-allocating...
The new setting `Preserve newlines in Markdown` in v0.25.0 playground  solves the extra-newline appearing (might be nice to have it enabled by default), but the `javascript` lang-specifier still appears...
For me this was happening on the latest Chrome on Win11, flow: - I open playground - delete all default content - click on M↓ to go to markdown source...
Maybe this issue is duplicate of: - https://github.com/facebook/lexical/issues/4355 The advantage of this dup is that it has some playground screenshots :)
+ maybe a separate search/filter bar would be helpful for keybindings window