-
### š Describe the bug
I have a model from @murphyk that's OOM'ing unless I explicitly disable the inductor pattern matcher. cc @ezyang @soumith @wconstab @ngimel @bdhirsh @cpuhrsch - cuda graphs hā¦
-
I'm in the situation where a single energy/force evaluation is only slightly too memory-expensive to fit on a single node. Based on the examples in the benchmark folder, I gather that the MEMORY_CUT aā¦
-
We're trying to tracing a module with dynamic shape, and ran into some problem:
repro:
```
import torch
import torchdynamo
class Module(torch.nn.Module):
def __init__(self):
sā¦
-
### š Describe the bug
We cut a [release branch](https://github.com/pytorch/pytorch/tree/release/2.1) for the 2.1.0 release.
Our plan from this point from this point is roughly:
* Phase 1 (unā¦
-
The script to run:
https://gist.github.com/AmosLewis/465795bfa1a4004fb96c742666515027
THE FINAL TOSA FILE:
[distilgpt2_tosa_20230123_elide.mlir](https://storage.googleapis.com/shark_tank/chi-nod/ā¦
-
See pytorch/pytorch#93757 for list of missing ops.
I've seen this op recently show up in `pytorch_CycleGAN_and_pix2pix` and `pytorch_stargan` when using the latest pytorch. Args are:
```
[WARNINā¦
-
### š Describe the bug
[test_compile](https://github.com/pyg-team/pytorch_geometric/blob/master/test/nn/test_compile.py) passes for dynamic and static shapes on simple gather scatter ops. Using my ā¦
-
This is supported by Eager PyTorch, see minimal repro:
```
import torchvision
import torch
import torchdynamo
dtype = torch.float16
a = torch.randn((1,), device="cuda", dtype=torch.floatā¦
-
(this is a brain dump for discussion and referencing code as we start to plumb things; design will change as we learn, and we should iterate/file more specific issues/etc as we go)
The goal is to hā¦
-
#trac 10632
This patch adds the following new features for crystals:
* CrystalOfTableaux now also accepts half-integer shapes for spin tableaux for type `B_n` and `D_n`
* Spin crystals have a newā¦