[HN Gopher] PyGraph: Robust Compiler Support for CUDA Graphs in ...
___________________________________________________________________
PyGraph: Robust Compiler Support for CUDA Graphs in PyTorch
Author : mfiguiere
Score : 79 points
Date : 2025-04-24 19:28 UTC (1 days ago)
(HTM) web link (arxiv.org)
(TXT) w3m dump (arxiv.org)
| infocollector wrote:
| The lack of a readily available, installable package (pip install
| pygraph - has no relation to this paper as far as i can tell)
| makes it difficult to fully assess the reproducibility and
| practical applicability of the work.
| easygenes wrote:
| There's a request code button here:
| https://www.catalyzex.com/paper/pygraph-robust-compiler-supp...
| bwfan123 wrote:
| why request code.. when all of pytorch2 is open, and this is
| built on top of it with some enhancements, why not put this
| out in the open
| tough wrote:
| I think that might be just a feature of the catalyzex
| platform for papers with no linked code yet that might
| internally add a +1 to code requested on their db and thats
| it
|
| some times papers come out a few weeks before code when its
| bleeding edge
| tho423i43234 wrote:
| Nice to see work by IISc show up on HN.
|
| Uday Bondhugula, the lead developer of Pluto framework for
| polyhedral comp. is also at IISc, whose group has spun out a
| startup,
|
| https://www.polymagelabs.com/
|
| Nice to see IISc support cool stuff like this (incl. their
| ArtPark initiative.)
| OutOfHere wrote:
| I don't see any source code.
| saagarjha wrote:
| This is neat, although it would be nice to see it merged into
| PyTorch instead of just a paper :) The key seems to be (beyond
| "obvious" optimizations like not running graphs that are measured
| to be slower) is that graphs "bake-in" parameters and if those
| change then the graph needs to be thrown away. The solution is
| indirecting more, so that what gets captured is a pointer that
| can remain constant, while the data behind it is changed. This
| also saves the need to copy in and out of a graph-captured buffer
| because you can just swap out the pointer instead. Of course
| there is overhead to this approach (I don't think the authors
| actually explore this much) in that you throw away information
| (divisibility, for example) that would allow for constructing
| better kernels, but often this is still worth it. (Or you could
| pass this through too.)
|
| Something worth exploring later would be getting better support
| for the rest of CUDA graphs into PyTorch, like conditional nodes.
| damnitbuilds wrote:
| Python can be used for many types of graphs. This package is for
| CUDA Graphs, so wouldn't "PyCudaGraph" be a better name?
___________________________________________________________________
(page generated 2025-04-25 23:02 UTC)