Can filter_fn include conv

[Freed-Wu]

Currently, default filter_fn is nn.Linear, don't include nn.Conv. Can it support conv?

[jerryzh168]

we were actually thinking of removing filter_fn in the future and have people express everything through config (FqnToConfig) itself, and users can generate the config with code:

fqn_to_config = {}
for n, m in model.named_modules():
      if isinstance(m, torch.nn.Conv2d):
          fqn_to_config[n] = conv_config

torchao_config = FqnToConfig(fqn_to_config)

any concerns for this?

also we have limited support for conv for now I think, only on B200 and above

[Freed-Wu]

any concerns for this?

Any document about FqnToConfig ?

[jcaip]

Yes please see Support for quantizing any parameter by its fully qualified name (FQN) via FqnToConfig in the latest release notes: https://github.com/pytorch/ao/releases/tag/v0.15.0

Please note that this is just support for applying the config, depending on the quantization config you are using, conv support may or may not be implement

[Freed-Wu]

Can the document be added to website not release tag?