fix(quantize_registry): compare activation_attrs against other, not self, in QuantizeConfig.__eq__

[Anai-Guo]

Problem

Default8BitQuantizeConfig.__eq__ and DefaultNBitQuantizeConfig.__eq__ compare activation_attrs against self instead of other:

  def __eq__(self, other):
    if not isinstance(other, Default8BitQuantizeConfig):
      return False

    return (self.weight_attrs == other.weight_attrs and
            self.activation_attrs == self.activation_attrs and   # <-- always True
            self.weight_quantizer == other.weight_quantizer and
            self.activation_quantizer == other.activation_quantizer and
            self.quantize_output == other.quantize_output)

self.activation_attrs == self.activation_attrs is always True, so the term contributes nothing to the and chain: activation_attrs is effectively excluded from equality, and two configs that differ only in which activations they quantize compare equal (and __ne__, which delegates to __eq__, reports them as not-unequal).

Every one of the four sibling terms in the same expression compares self.X == other.X, and every other __eq__ in the package follows that idiom — e.g. LastValueQuantizer.__eq__ in quantizers.py:

    return (self.num_bits == other.num_bits and
            self.per_axis == other.per_axis and
            self.symmetric == other.symmetric and
            self.narrow_range == other.narrow_range)

The same line is duplicated verbatim in the default_n_bit copy, so both are fixed here.

• default_8bit/default_8bit_quantize_registry.py:369
• experimental/default_n_bit/default_n_bit_quantize_registry.py:415

Fix

self.activation_attrs == self.activation_attrs → self.activation_attrs == other.activation_attrs, in both files. Two lines, no other change.

Verification

I don't have a TF runtime on this box, so I verified behaviourally by extracting __eq__ verbatim from the file with ast.get_source_segment, exec-ing it against a minimal stand-in that only holds the five compared attributes, and running the current line against the patched one:

case	expected	master	patched
differs only in activation_attrs	False	True ❌	False ✅
identical	True	True	True
differs in weight_attrs (control)	False	False	False
empty vs non-empty activation_attrs	False	True ❌	False ✅

The control row is there to show the check isn't vacuous: weight_attrs, which is compared correctly, already discriminates on master, and the patch leaves it alone. Only the two rows that isolate activation_attrs change.

A concrete pair that is wrongly equal on master:

a = Default8BitQuantizeConfig(['kernel'], ['activation'], False)
b = Default8BitQuantizeConfig(['kernel'], ['recurrent_activation'], False)
a == b   # True on master, False with this patch

🤖 Generated with Claude Code

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.