Feat (examples): add support for Stable Diffusion XL #909
Conversation
| # Workaround to expose `in_features` attribute from the Hook Wrapper | ||
| for m in pipe.unet.modules(): | ||
| if isinstance(m, KwargsForwardHook) and hasattr(m.module, 'in_features'): | ||
| m.in_features = m.module.in_features |
There was a problem hiding this comment.
@costigt-dev This (and similarly, line 129-131) is why I need the wrapper to expose the attributes of the wrapped module.
In this case, the wrapping is only temporary so I need to keep some sort of reference of wrapper vs wrapped
| parser.add_argument( | ||
| '--weight-quant-granularity', | ||
| type=str, | ||
| default='per_group', | ||
| default='per_channel', |
There was a problem hiding this comment.
could you explain a bit of context why this changes from per group to per channel?
There was a problem hiding this comment.
We have more robust support for per_channel quantization, including export representation.
We can export per_group quantized network but we're still defining the best way to represent that quantization style.
| if args.export_target == 'torchscript': | ||
| if args.weight_quant_granularity == 'per_group': | ||
| export_manager = BlockQuantProxyLevelManager | ||
| else: | ||
| export_manager = TorchQCDQManager | ||
| export_manager.change_weight_export(export_weight_q_node=True) | ||
| export_manager.change_weight_export(export_weight_q_node=args.export_weight_q_node) |
There was a problem hiding this comment.
does the value from args always resolve to True/False type?
There was a problem hiding this comment.
Because there are mostly input and argument changes in this PR my only real question would be - have you run it or tested it to double check there isn't any clash with arguments? I'd like to give it a run myself (with supervision) since it's a hot topic demo but I don't want to hold up the PR, so I'm hitting 'approve'.
Great to see some improvements to comments in tricky code and notes on PR to add context!
I asked a couple of "put me in context here" questions, but I don't see any code that looks like it could be a bug. I suggest a bit of background info in PR descriptions would be good for next PRs.
Missing:
Possibly Learned Round but it could be very slowTorchscript export is deprecated since it is not used and difficult to make compatible with SD-XL.
More investigations are needed to fix it in case it is required.