Ma Mingfei

> @mingfeima the profiler will affect the e2e runs. In general, did you repro the regression on your side? I can't reproduce the regression on my side. `hf_BigBird` is slightly...

@xuzhao9 Hi any update for this issue? One more thing to clarify is that performance improvement from https://github.com/pytorch/pytorch/pull/84404 can not be properly reflected by the listed modes in this issue...

@NicolasHug First of all, yes our priority is inference. And the most requested model from our customers are `MaskedRCNN` and its variants. So from this point of view, the key...

@NicolasHug @vfdev-5 Oh sorry for the late response, super busy recently, just got time to take a look at this last weekend ... I opened https://github.com/pytorch/pytorch/pull/87053 to address `mode=bilinear (3,...

> Hopefully there will be support for uint8 type input and an accelerated version of it for `interpolate()` as mentioned in [pytorch/pytorch#86361 (comment)](https://github.com/pytorch/pytorch/pull/86361#issuecomment-1269822386) and [pytorch/pytorch#5580](https://github.com/pytorch/pytorch/issues/5580) . sum up the status...

Ok, I see. Then it is better to optimize scatter_reduce in torch. Just checked the code, scatter_add and scatter_reduce share the same kernel in torch so they have the same...

> > I'm a complete newbie to this, so my question is to learn not suggest something. Can you explain what you're intending to change here? It sounds like you...

current benchmark profiling result uses the default setting. Some scripts, for example [to_hetero_mag](https://github.com/pyg-team/pytorch_geometric/blob/master/examples/hetero/to_hetero_mag.py#L29) would explicitly set the `num_workers`, if not the pytorch default setting will be 4. `DataLoader` time in...

Updates on `scatter_add` optimizations, PR submitted at https://github.com/pytorch/pytorch/pull/82703 ### Initiative Depending type of the `edge_index`, message passing will choose different paths: a) `scatter_add` for dense tensor; b) `spmm` for `SparseTensor`....

> Just to understand: Does this mean that we first sort `index` and then do a segment reduction? In that case it might be good to preserve the information that...