Prepare for PEFT release of v0.14.0 (#2258)

- Bump versions
- Remove deprecated convert_pissa_to_lora argument
- Remove a pytest skip for older transformers versions
- Adjust some comments, docstrings

Author: BenjaminBossan

examples/pissa_finetuning/README.md

@@ -71,7 +71,7 @@ The main advantage of PiSSA is concentrated during the training phase. For a tra
 peft_model.save_pretrained(output_dir) 
 # Given the matrices $A_0$ and $B_0$, initialized by PiSSA and untrained, and the trained matrices $A$ and $B$, 
 # we can convert these to LoRA by setting $\Delta W = A \times B - A_0 \times B_0 = [A \mid A_0] \times [B \mid -B_0]^T = A'B'$.
-peft_model.save_pretrained(output_dir, convert_pissa_to_lora="pissa_init")
+peft_model.save_pretrained(output_dir, path_initial_model_for_weight_conversion="pissa_init")
 
 ```
 This conversion enables the loading of LoRA on top of a standard base model:

examples/pissa_finetuning/pissa_finetuning.py

@@ -136,7 +136,7 @@ class ScriptArguments(SFTConfig):
 if script_args.convert_pissa_to_lora:
     peft_model.save_pretrained(
         os.path.join(script_args.output_dir, "pissa_lora"),
-        convert_pissa_to_lora=os.path.join(script_args.residual_model_name_or_path, "pissa_init"),
+        path_initial_model_for_weight_conversion=os.path.join(script_args.residual_model_name_or_path, "pissa_init"),
     )
 else:
     peft_model.save_pretrained(

setup.py

@@ -15,7 +15,7 @@
 from setuptools import find_packages, setup
 
 
-VERSION = "0.13.3.dev0"
+VERSION = "0.14.0"
 
 extras = {}
 extras["quality"] = [

src/peft/__init__.py

@@ -17,7 +17,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "0.13.3.dev0"
+__version__ = "0.14.0"
 
 from .auto import (
     AutoPeftModel,

src/peft/peft_model.py

@@ -229,7 +229,6 @@ def save_pretrained(
         selected_adapters: Optional[list[str]] = None,
         save_embedding_layers: Union[str, bool] = "auto",
         is_main_process: bool = True,
-        convert_pissa_to_lora: Optional[str] = None,
         path_initial_model_for_weight_conversion: Optional[str] = None,
         **kwargs: Any,
     ) -> None:
@@ -253,8 +252,6 @@ def save_pretrained(
             is_main_process (`bool`, *optional*):
                 Whether the process calling this is the main process or not. Will default to `True`. Will not save the
                 checkpoint if not on the main process, which is important for multi device setups (e.g. DDP).
-            convert_pissa_to_lora (`str, *optional*`):
-                Deprecated. Use `path_initial_model_for_weight_conversion` instead.
             path_initial_model_for_weight_conversion (`str, *optional*`):
                 The path to the initialized adapter, which is obtained after initializing the model with PiSSA or OLoRA
                 and before performing any training. When `path_initial_model_for_weight_conversion` is not None, the
@@ -281,13 +278,6 @@ def save_pretrained(
                     f"You passed an invalid `selected_adapters` arguments, current supported adapter names are"
                     f" {list(self.peft_config.keys())} - got {selected_adapters}."
                 )
-        # TODO: remove deprecated parameter in PEFT v0.14.0
-        if convert_pissa_to_lora is not None:
-            warnings.warn(
-                "`convert_pissa_to_lora` is deprecated and will be removed in a future version. "
-                "Use `path_initial_model_for_weight_conversion` instead."
-            )
-            path_initial_model_for_weight_conversion = convert_pissa_to_lora
 
         def save_mutated_as_lora(peft_config, path_initial_model_for_weight_conversion, output_state_dict, kwargs):
             if peft_config.use_rslora and (peft_config.rank_pattern or peft_config.alpha_pattern):

tests/test_custom_models.py

@@ -912,7 +912,13 @@ def from_pretrained(cls, model_id, torch_dtype=None):
 
 
 class PeftCustomModelTester(unittest.TestCase, PeftCommonTester):
-    """TODO"""
+    """
+    Implements the tests for custom models.
+
+    Most tests should just call the parent class, e.g. test_save_pretrained calls self._test_save_pretrained. Override
+    this if custom models don't work with the parent test method.
+
+    """
 
     transformers_class = MockTransformerWrapper
 

tests/test_initialization.py

@@ -583,44 +583,6 @@ def test_pissa_alpha_pattern_and_rslora_raises(self, tmp_path):
                 tmp_path / "pissa-model", path_initial_model_for_weight_conversion=tmp_path / "init-model"
             )
 
-    # TODO: remove test for deprecated arg in PEFT v0.14.0
-    def test_lora_pissa_conversion_same_output_after_loading_with_deprecated_arg(self, data, tmp_path):
-        model = self.get_model()
-        config = LoraConfig(init_lora_weights="pissa", target_modules=["linear"], r=8)
-        peft_model = get_peft_model(deepcopy(model), config)
-        peft_model.peft_config["default"].init_lora_weights = True
-        peft_model.save_pretrained(tmp_path / "init-model")
-        peft_model.peft_config["default"].init_lora_weights = "pissa"
-
-        tol = 1e-06
-        peft_model.base_model.linear.lora_B["default"].weight.data *= 2.0
-        output_pissa = peft_model(data)[0]
-
-        peft_model.save_pretrained(tmp_path / "pissa-model-converted", convert_pissa_to_lora=tmp_path / "init-model")
-        model_converted = PeftModel.from_pretrained(deepcopy(model), tmp_path / "pissa-model-converted")
-        output_converted = model_converted(data)[0]
-
-        assert torch.allclose(output_pissa, output_converted, atol=tol, rtol=tol)
-        assert model_converted.peft_config["default"].r == 16
-        assert model_converted.base_model.model.linear.lora_A["default"].weight.shape[0] == 16
-        assert torch.allclose(
-            model.linear.weight, model_converted.base_model.model.linear.base_layer.weight, atol=tol, rtol=tol
-        )
-
-    # TODO: remove test for deprecated warning in PEFT v0.14.0
-    def test_lora_pissa_conversion_deprecated_warning(self, data, tmp_path):
-        model = self.get_model()
-        config = LoraConfig(init_lora_weights="pissa", target_modules=["linear"], r=8)
-        peft_model = get_peft_model(deepcopy(model), config)
-        peft_model.peft_config["default"].init_lora_weights = True
-        peft_model.save_pretrained(tmp_path / "init-model")
-        warning_message = "`convert_pissa_to_lora` is deprecated and will be removed in a future version. Use `path_initial_model_for_weight_conversion` instead."
-        # Test the warning
-        with pytest.warns(UserWarning, match=warning_message):
-            peft_model.save_pretrained(
-                tmp_path / "pissa-model-converted", convert_pissa_to_lora=tmp_path / "init-model"
-            )
-
     def test_olora_conversion_same_output_after_loading(self, data, tmp_path):
         model = self.get_model()
         output_base = model(data)[0]

tests/test_xlora.py

@@ -15,10 +15,8 @@
 import os
 
 import huggingface_hub
-import packaging
 import pytest
 import torch
-import transformers
 from safetensors.torch import load_file
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
@@ -27,9 +25,6 @@
 from peft.utils import infer_device
 
 
-uses_transformers_4_45 = packaging.version.parse(transformers.__version__) >= packaging.version.parse("4.45.0")
-
-
 class TestXlora:
     torch_device = infer_device()
 
@@ -133,8 +128,7 @@ def test_functional(self, tokenizer, model):
         )
         assert torch.isfinite(outputs[: inputs.shape[1] :]).all()
 
-    # TODO: remove the skip when 4.45 is released!
-    @pytest.mark.skipif(not uses_transformers_4_45, reason="Requires transformers >= 4.45")
+    # TODO: fix the xfailing test
     @pytest.mark.xfail
     def test_scalings_logging_methods(self, tokenizer, model):
         model.enable_scalings_logging()