Dango233

Update cutout method

1

Update the cutout method to increase the variety of the overview cuts. Thanks @dicknascarsixtynine for flagging the issue

Stucked in REALY mode when connect to router, but can DIRECT to the PC behind it

2

I have to admit this is a weird situation: I can DIRECT connect to my PC running Zerotier behind my OpenWRT router using the PC Zerotier address, but I can...

Torch JIT breaks when memory_efficient_attention

13

# 🐛 Bug torch.jit.trace breaks with the following error: `RuntimeError: unsupported output type: int, from operator: xformers::efficient_attention_forward_generic` The output of the ops contains an int that can't be traced by...

* Use Cutlass ops when possible to +15% speed

1

Use Cutlass ops when possible to +15% speed (for free) Sampling speed 1.80it/s->2.08it/s on RTX5000. Won't trigger if MemoryEfficientAttentionCutlassOp is not available.

Fix mid block checker to fix custom model loading

4

The from_pretrained_2d runs into `ValueError: unknown mid_block_type: UNetMidBlock2DCrossAttn` when ` "mid_block_type": "UNetMidBlock2DCrossAttn"` is specified in the config file of pertained unet. This happens with some custom models (model converters could...

* When load network saved using previous version of oneflow, as the cpg attribute not exist, the `self.cpg` reference will fail. * PR address this problem by adding an additional...