LORA/MODEL: Use max rank of pattern for add_weighted_adapter

[Beinsezii]

Otherwise it's possible that the new merged adapter won't allocate enough space when a patterned rank is larger than config.r.

For example, the cat method uses sum(adapters_ranks) for the empty tensor alloc, so if cat(loras) for a given layer is ever > the sum of config.r the assignment will overflow.

Local test fails but it's all entirely the TestEvaInitialization module which appears completely unrelated.

[BenjaminBossan]

Thanks for the PR, indeed the rank_pattern. Would you please be so kind to also add a unit test for this? A good starting point could be this test:

peft/tests/test_custom_models.py

Lines 1970 to 1981 in 2d90569

	def test_multiple_adapters_mixed_modules_to_save_merging_adapters(self):
	# See issue 1574
	# This test is similar to test_multiple_adapters_mixed_modules_to_save, but it also checks that merging adapter
	# weights works when one adapter has a modules_to_save and the other hasn't
	config0 = LoraConfig(target_modules=["lin0"], modules_to_save=["lin1"])
	config1 = LoraConfig(target_modules=["lin0"])
	model = MLP()
	model = get_peft_model(model, config0).to(self.torch_device)
	model.add_adapter("other", config1)
	# check that this does not raise
	model.add_weighted_adapter(["default", "other"], weights=[1.0, 1.0], adapter_name="merged")

I think it can be copied and slightly adjusted by adding a config with rank_pattern.

I wonder if we need even more checks for rank_pattern, especially when not using the cat method. But this can be dealt with in another PR.

[bghira]

thanks Ben, i'd requested this pull request be opened since it seems trivial enough but I'm not sure we'll have time to add unit tests here, especially considering none of us are particularly familiar with the unit test ecosystem for PEFT.

[BenjaminBossan]

Thanks for providing more context @bghira. I wrote the code required to test this change:

    @parameterized.expand([IA3Config, LoHaConfig, LoKrConfig, LoraConfig, HRAConfig, BoneConfig])
    def test_add_weighted_adapter_cat_with_rank_pattern(self, config_cls):
        # Fixes a bug described in #2512, which resulted from the rank_pattern not being taken into account
        config0 = LoraConfig(target_modules=["lin0", "lin1"], r=8)
        config1 = LoraConfig(target_modules=["lin0", "lin1"], r=8, rank_pattern={"lin0": 16})
        model = MLP()
        model = get_peft_model(model, config0).to(self.torch_device)
        model.add_adapter("other", config1)
        model.add_weighted_adapter(["default", "other"], weights=[1.0, 1.0], adapter_name="merged", combination_type="cat")

I can best be added in this line. @Beinsezii would you be so kind to add it to your PR?

[Beinsezii]

Verified it fails on main and passes on beinsezii/weighted_adapter_max_rank_pattern