frozenleaves/peft
1
0
Fork 0
mirror of https://github.com/huggingface/peft.git synced 2025-10-20 23:43:47 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
2ea5377ee3069fa5aa49ed3930ac52bc706aed85
peft/tests/testing_common.py
Greenewald 293aea5df6 Support for Activated LoRA (#2609) 
This PR migrates Activated LoRA (aLoRA) support from a standalone Github (see above) to PEFT itself.

Note there is also an active PR for vLLM inference support for Activated LoRA: vllm-project/vllm#19710 . There are also collections of aLoRA models on huggingface (in the ibm-granite org), note that these preexisting models run off of the standalone github repo and will be updated to work with this new PEFT feature if merged.

Description of changes: Activated LoRA is a modification of the LoRA architecture to "activate" the adapter weights only on tokens coming after a specified invocation_string. This fact makes it so that KV values for the string coming before the activation matches KV values for the base model. This allows KV cache for the input to be interchangeable between the base model and adapter model, and allows for major speedups in inference pipelines (e.g. agentic pipelines) that want to use both base models and adapter models. See the paper for detailed exploration of use cases and further elaboration.

Other notes:

The crux of the changes are really in layer.py. Everything else is simply managing the alora_offsets quantity which defines where the weights start to be activated. This is determined by scanning input strings for the invocation_string defined in the aLoraConfig.
    
I believe that aLoRA really only makes sense for CausalLMs, hence I've only implemented this for that model type.

Merging doesn't make sense for aLoRA adapters since the weights are not universally applied to all tokens.
    
I used the LoRA code as a starting point, but did not implement various seemingly extra features in that code.

As of now, invocation_string should probably start and end with special tokens, to avoid tokenizer issues at the boundary. Open to suggestions on how to make this more general if needed.

---------

Co-authored-by: githubnemo <githubnemo@users.noreply.github.com>
2025-09-03 18:26:50 +02:00

1977 lines
85 KiB
Python
Raw Blame History

# Copyright 2023-present the HuggingFace Inc. team.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
import copy
import json
import os
import pickle
import platform
import re
import shutil
import tempfile
import warnings
from dataclasses import replace
import pytest
import torch
import yaml
from diffusers import StableDiffusionPipeline
from packaging import version
from safetensors.torch import save_file
from peft import (
AdaLoraConfig,
BOFTConfig,
BoneConfig,
CPTConfig,
FourierFTConfig,
HRAConfig,
IA3Config,
LNTuningConfig,
LoHaConfig,
LoKrConfig,
LoraConfig,
MissConfig,
OFTConfig,
PeftModel,
PeftType,
PrefixTuningConfig,
PromptEncoderConfig,
PromptLearningConfig,
PromptTuningConfig,
RandLoraConfig,
VBLoRAConfig,
VeraConfig,
get_peft_model,
get_peft_model_state_dict,
inject_adapter_in_model,
prepare_model_for_kbit_training,
)
from peft.tuners._buffer_dict import BufferDict
from peft.tuners.lora import LoraLayer
from peft.tuners.tuners_utils import BaseTunerLayer
from peft.utils import (
AuxiliaryTrainingWrapper,
ModulesToSaveWrapper,
TrainableTokensWrapper,
_get_submodules,
infer_device,
)
from .testing_utils import get_state_dict, hub_online_once
CONFIG_TESTING_KWARGS = (
# IA³
{
"target_modules": None,
"feedforward_modules": None,
},
# LoRA
{
"r": 8,
"lora_alpha": 32,
"target_modules": None,
"lora_dropout": 0.05,
"bias": "none",
},
# prefix tuning
{
"num_virtual_tokens": 10,
},
# prompt encoder
{
"num_virtual_tokens": 10,
"encoder_hidden_size": 32,
},
# prompt tuning
{
"num_virtual_tokens": 10,
},
# AdaLoRA
{
"target_modules": None,
"total_step": 1,
},
# BOFT
{
"target_modules": None,
},
# VeRA
{
"r": 8,
"target_modules": None,
"vera_dropout": 0.05,
"projection_prng_key": 0xFF,
"d_initial": 0.1,
"save_projection": True,
"bias": "none",
},
# FourierFT
{
"n_frequency": 10,
"target_modules": None,
},
# HRA
{
"target_modules": None,
},
# VBLoRA
{"target_modules": None, "vblora_dropout": 0.05, "vector_length": 1, "num_vectors": 2},
# OFT
{
"target_modules": None,
},
# Bone
{
"target_modules": None,
"r": 2,
},
# MiSS
{
"target_modules": None,
"r": 2,
},
# LoRA + trainable_tokens
{
"r": 8,
"lora_alpha": 32,
"target_modules": None,
"lora_dropout": 0.05,
"bias": "none",
"trainable_token_indices": [0, 1, 3],
},
# RandLoRA
{
"r": 32,
"randlora_alpha": 64,
"target_modules": None,
"randlora_dropout": 0.05,
"projection_prng_key": 0xFF,
"save_projection": True,
"bias": "none",
},
# CPT tuninig
{
"cpt_token_ids": [0, 1, 2, 3, 4, 5, 6, 7], # Example token IDs for testing
"cpt_mask": [1, 1, 1, 1, 1, 1, 1, 1],
"cpt_tokens_type_mask": [1, 2, 2, 2, 3, 3, 4, 4],
},
)
CLASSES_MAPPING = {
"ia3": (IA3Config, CONFIG_TESTING_KWARGS[0]),
"lora": (LoraConfig, CONFIG_TESTING_KWARGS[1]),
"prefix_tuning": (PrefixTuningConfig, CONFIG_TESTING_KWARGS[2]),
"prompt_encoder": (PromptEncoderConfig, CONFIG_TESTING_KWARGS[3]),
"prompt_tuning": (PromptTuningConfig, CONFIG_TESTING_KWARGS[4]),
"adalora": (AdaLoraConfig, CONFIG_TESTING_KWARGS[5]),
"boft": (BOFTConfig, CONFIG_TESTING_KWARGS[6]),
"vera": (VeraConfig, CONFIG_TESTING_KWARGS[7]),
"fourierft": (FourierFTConfig, CONFIG_TESTING_KWARGS[8]),
"hra": (HRAConfig, CONFIG_TESTING_KWARGS[9]),
"vblora": (VBLoRAConfig, CONFIG_TESTING_KWARGS[10]),
"oft": (OFTConfig, CONFIG_TESTING_KWARGS[11]),
"bone": (BoneConfig, CONFIG_TESTING_KWARGS[12]),
"miss": (MissConfig, CONFIG_TESTING_KWARGS[12]),
"lora+trainable_tokens": (LoraConfig, CONFIG_TESTING_KWARGS[13]),
"randlora": (RandLoraConfig, CONFIG_TESTING_KWARGS[14]),
}
DECODER_MODELS_EXTRA = {"cpt": (CPTConfig, CONFIG_TESTING_KWARGS[15])}
class PeftCommonTester:
r"""
A large testing suite for testing common functionality of the PEFT models.
Attributes:
torch_device (`torch.device`):
The device on which the tests will be run.
transformers_class (`transformers.PreTrainedModel`):
The transformers class that is being tested.
"""
torch_device = infer_device()
transformers_class = None
def prepare_inputs_for_common(self):
raise NotImplementedError
def check_modelcard(self, tmp_dirname, model):
# check the generated README.md
filename = os.path.join(tmp_dirname, "README.md")
assert os.path.exists(filename)
with open(filename, encoding="utf-8") as f:
readme = f.read()
metainfo = re.search(r"---\n(.*?)\n---", readme, re.DOTALL).group(1)
dct = yaml.safe_load(metainfo)
assert dct["library_name"] == "peft"
if hasattr(model, "config"):
assert dct["base_model"] == model.config.to_dict()["_name_or_path"]
else: # a custom model
assert "base_model" not in dct
# The Hub expects the lora tag to be set for PEFT LoRA models since they
# have explicit support for things like inference.
if model.active_peft_config.peft_type.value == "LORA":
assert "lora" in dct["tags"]
def check_config_json(self, tmp_dirname, model):
# check the generated config.json
filename = os.path.join(tmp_dirname, "adapter_config.json")
assert os.path.exists(filename)
with open(filename, encoding="utf-8") as f:
config = json.load(f)
if hasattr(model, "config"): # custom models don't have a config attribute
assert config["base_model_name_or_path"] == model.config.to_dict()["_name_or_path"]
def perturb_trainable_token_weights_if_used(self, model, config_kwargs, adapter_name="default", scale=1.0):
"""TrainableTokensLayer is initialized to be a no-op by default. Since there's currently no way to pass
`init_weights=False` to the trainable tokens layer when used in conjunction with LoRA, we have to do it like
this to make sure that it is *not* a no-op (essentially simulating "training" of the adapter).
"""
if "trainable_token_indices" not in config_kwargs:
return
token_wrapper = None
if hasattr(model, "get_input_embeddings"):
token_wrapper = model.get_input_embeddings()
else:
for module in model.modules():
if isinstance(module, TrainableTokensWrapper):
token_wrapper = module
break
# for a model with trainable_token_indices there should always be a trainable token wrapper somewhere.
# if not, then there's something broken.
assert token_wrapper is not None
token_wrapper.token_adapter.trainable_tokens_delta[adapter_name].data = (
torch.rand_like(token_wrapper.token_adapter.trainable_tokens_delta[adapter_name].data) * scale
)
def _test_model_attr(self, model_id, config_cls, config_kwargs):
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
assert hasattr(model, "save_pretrained")
assert hasattr(model, "from_pretrained")
assert hasattr(model, "push_to_hub")
def _test_adapter_name(self, model_id, config_cls, config_kwargs):
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config, adapter_name="test-adapter")
correctly_converted = False
for n, _ in model.named_parameters():
if "test-adapter" in n:
correctly_converted = True
break
assert correctly_converted
def _test_prepare_for_training(self, model_id, config_cls, config_kwargs):
if config_kwargs.get("trainable_token_indices", None) is not None:
# incompatible because trainable tokens is marking embeddings as trainable
self.skipTest("Trainable tokens is incompatible with this test.")
# some tests require specific tokenizers, make sure that they can be fetched as well
with hub_online_once(model_id + config_kwargs.get("tokenizer_name_or_path", "")):
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
dummy_input = self.prepare_inputs_for_testing()
dummy_output = model.get_input_embeddings()(dummy_input["input_ids"])
assert not dummy_output.requires_grad
# load with `prepare_model_for_kbit_training`
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
model = prepare_model_for_kbit_training(model)
for param in model.parameters():
assert not param.requires_grad
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
# For backward compatibility
if hasattr(model, "enable_input_require_grads"):
model.enable_input_require_grads()
else:
def make_inputs_require_grad(module, input, output):
output.requires_grad_(True)
model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
dummy_input = self.prepare_inputs_for_testing()
dummy_output = model.get_input_embeddings()(dummy_input["input_ids"])
assert dummy_output.requires_grad
def _test_load_model_low_cpu_mem_usage(self, model_id, config_cls, config_kwargs):
# Ensure that low_cpu_mem_usage=True works for from_pretrained and load_adapter and that the resulting model's
# parameters are on the correct device.
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
# note: not using the context manager here because it fails on Windows CI for some reason
tmp_dirname = tempfile.mkdtemp()
try:
model.save_pretrained(tmp_dirname)
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
model = PeftModel.from_pretrained(
model, tmp_dirname, torch_device=self.torch_device, low_cpu_mem_usage=True
)
assert {p.device.type for p in model.parameters()} == {self.torch_device}
model.load_adapter(tmp_dirname, adapter_name="other", low_cpu_mem_usage=True)
assert {p.device.type for p in model.parameters()} == {self.torch_device}
finally:
try:
shutil.rmtree(tmp_dirname)
except PermissionError:
# windows error
pass
# also test injecting directly
del model
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
inject_adapter_in_model(config, model, low_cpu_mem_usage=True) # check that there is no error
if not isinstance(config, LNTuningConfig):
# LN tuning does not add adapter layers that could be on meta device, it only changes the requires_grad.
# Therefore, there is no meta device for LN tuning.
assert "meta" in {p.device.type for p in model.parameters()}
def _test_save_pretrained(self, model_id, config_cls, config_kwargs, safe_serialization=True):
# ensure that the weights are randomly initialized
if issubclass(config_cls, LoraConfig):
config_kwargs = config_kwargs.copy()
config_kwargs["init_lora_weights"] = False
if issubclass(config_cls, IA3Config):
config_kwargs = config_kwargs.copy()
config_kwargs["init_ia3_weights"] = False
if hasattr(config_cls, "init_weights"):
config_kwargs = config_kwargs.copy()
config_kwargs["init_weights"] = False
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
with tempfile.TemporaryDirectory() as tmp_dirname:
if safe_serialization:
model.save_pretrained(tmp_dirname)
else:
model.save_pretrained(tmp_dirname, safe_serialization=False)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
with warnings.catch_warnings(record=True) as recs:
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname)
# ensure that there is no warning
assert not any("Found missing adapter keys" in str(rec.message) for rec in recs)
# check if the state dicts are equal
if issubclass(config_cls, PromptEncoderConfig):
# For prompt encoding, when loading the whole state_dict, there are differences, therefore, only load
# adapter-specific weights for comparison.
# TODO: is this expected?
state_dict = get_peft_model_state_dict(model, unwrap_compiled=True)
state_dict_from_pretrained = get_peft_model_state_dict(model_from_pretrained, unwrap_compiled=True)
else:
state_dict = get_state_dict(model, unwrap_compiled=True)
state_dict_from_pretrained = get_state_dict(model_from_pretrained, unwrap_compiled=True)
# check if tensors equal
for key in state_dict.keys():
assert torch.allclose(
state_dict[key].to(self.torch_device), state_dict_from_pretrained[key].to(self.torch_device)
)
target_adapter_filename = "adapter_model.safetensors" if safe_serialization else "adapter_model.bin"
# check if `adapter_model.safetensors` is present
assert os.path.exists(os.path.join(tmp_dirname, target_adapter_filename))
# check if `adapter_config.json` is present
assert os.path.exists(os.path.join(tmp_dirname, "adapter_config.json"))
# check if `model.safetensors` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "model.safetensors"))
# check if `config.json` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "config.json"))
self.check_modelcard(tmp_dirname, model)
self.check_config_json(tmp_dirname, model)
def _test_save_pretrained_selected_adapters(self, model_id, config_cls, config_kwargs, safe_serialization=True):
if issubclass(config_cls, AdaLoraConfig):
# AdaLora does not support adding more than 1 adapter
return pytest.skip(f"Test not applicable for {config_cls}")
# ensure that the weights are randomly initialized
if issubclass(config_cls, LoraConfig):
config_kwargs = config_kwargs.copy()
config_kwargs["init_lora_weights"] = False
elif issubclass(config_cls, IA3Config):
config_kwargs = config_kwargs.copy()
config_kwargs["init_ia3_weights"] = False
elif hasattr(config_cls, "init_weights"):
config_kwargs["init_weights"] = False
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
new_adapter_config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model.add_adapter("new_adapter", new_adapter_config)
with tempfile.TemporaryDirectory() as tmp_dirname:
if safe_serialization:
model.save_pretrained(tmp_dirname)
else:
model.save_pretrained(tmp_dirname, safe_serialization=False)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname)
new_adapter_dir = os.path.join(tmp_dirname, "new_adapter")
model_from_pretrained.load_adapter(new_adapter_dir, "new_adapter")
# check if the state dicts are equal
if issubclass(config_cls, PromptEncoderConfig):
# For prompt encoding, when loading the whole state_dict, there are differences, therefore, only load
# adapter-specific weights for comparison.
# TODO: is this expected?
state_dict = get_peft_model_state_dict(model, unwrap_compiled=True)
state_dict_from_pretrained = get_peft_model_state_dict(model_from_pretrained, unwrap_compiled=True)
else:
state_dict = get_state_dict(model, unwrap_compiled=True)
state_dict_from_pretrained = get_state_dict(model_from_pretrained, unwrap_compiled=True)
# check if same keys
assert state_dict.keys() == state_dict_from_pretrained.keys()
# check if tensors equal
for key in state_dict.keys():
assert torch.allclose(
state_dict[key].to(self.torch_device), state_dict_from_pretrained[key].to(self.torch_device)
)
target_adapter_filename = "adapter_model.safetensors" if safe_serialization else "adapter_model.bin"
# check if `adapter_model.safetensors` is present
assert os.path.exists(os.path.join(tmp_dirname, target_adapter_filename))
assert os.path.exists(os.path.join(new_adapter_dir, target_adapter_filename))
# check if `adapter_config.json` is present
assert os.path.exists(os.path.join(tmp_dirname, "adapter_config.json"))
assert os.path.exists(os.path.join(new_adapter_dir, "adapter_config.json"))
# check if `model.safetensors` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "model.safetensors"))
assert not os.path.exists(os.path.join(new_adapter_dir, "model.safetensors"))
# check if `config.json` is not present
assert not os.path.exists(os.path.join(tmp_dirname, "config.json"))
assert not os.path.exists(os.path.join(new_adapter_dir, "config.json"))
self.check_modelcard(tmp_dirname, model)
self.check_config_json(tmp_dirname, model)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname, selected_adapters=["default"])
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname)
assert "default" in model_from_pretrained.peft_config.keys()
assert "new_adapter" not in model_from_pretrained.peft_config.keys()
def _test_from_pretrained_config_construction(self, model_id, config_cls, config_kwargs):
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(base_model_name_or_path=model_id, **config_kwargs)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(
model_from_pretrained, tmp_dirname, is_trainable=False, config=config
)
assert model_from_pretrained.peft_config["default"].inference_mode
assert model_from_pretrained.peft_config["default"] is config
def _test_load_multiple_adapters(self, model_id, config_cls, config_kwargs):
# just ensure that this works and raises no error
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
del model
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
model = PeftModel.from_pretrained(model, tmp_dirname, torch_device=self.torch_device)
load_result1 = model.load_adapter(tmp_dirname, adapter_name="other")
load_result2 = model.load_adapter(tmp_dirname, adapter_name="yet-another")
# VBLoRA uses a shared "vblora_vector_bank" across all layers, causing it to appear
# in the missing keys list, which leads to failed test cases. So
# skipping the missing keys check for VBLoRA.
if config.peft_type != "VBLORA":
assert load_result1.missing_keys == []
assert load_result2.missing_keys == []
def _test_merge_layers_fp16(self, model_id, config_cls, config_kwargs):
if (
config_cls not in (LoraConfig, IA3Config, AdaLoraConfig, LoHaConfig, LoKrConfig, VBLoRAConfig)
or config_kwargs.get("alora_invocation_tokens") is not None
):
# Merge layers only supported for LoRA and IA³, and not for Activated LoRA (aLoRA)
if config_kwargs.get("alora_invocation_tokens") is None:
return pytest.skip(f"Test not applicable for {config_cls}")
else:
return pytest.skip("Test not applicable for Activated LoRA")
if ("gpt2" in model_id.lower()) and (config_cls != LoraConfig):
self.skipTest("Merging GPT2 adapters not supported for IA³ (yet)")
if (self.torch_device in ["cpu"]) and (version.parse(torch.__version__) <= version.parse("2.1")):
self.skipTest("PyTorch 2.1 not supported for Half of addmm_impl_cpu_ ")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id, torch_dtype=torch.float16)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(device=self.torch_device, dtype=torch.float16)
model.eval()
# This should simply work
_ = model.merge_and_unload()
def _test_merge_layers_nan(self, model_id, config_cls, config_kwargs):
if (
config_cls
not in (
LoraConfig,
IA3Config,
AdaLoraConfig,
LoHaConfig,
LoKrConfig,
VeraConfig,
FourierFTConfig,
)
or config_kwargs.get("alora_invocation_tokens") is not None
):
# Merge layers only supported for LoRA and IA³, and not for Activated LoRA (aLoRA)
return
if ("gpt2" in model_id.lower()) and (config_cls != LoraConfig):
self.skipTest("Merging GPT2 adapters not supported for IA³ (yet)")
if "gemma" in model_id.lower():
# TODO: could be related to tied weights
self.skipTest("Merging currently fails with gemma")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
self.perturb_trainable_token_weights_if_used(model, config_kwargs)
dummy_input = self.prepare_inputs_for_testing()
model.eval()
# This should work
logits_unmerged = model(**dummy_input)[0]
model = model.merge_and_unload()
logits_merged = model(**dummy_input)[0]
assert torch.allclose(logits_unmerged, logits_merged, atol=1e-3, rtol=1e-3)
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
for name, module in model.named_parameters():
if (
"lora_A" in name
or "ia3" in name
or "lora_E" in name
or "lora_B" in name
or "vera_lambda" in name
or "fourierft_spectrum" in name
):
module.data[0] = torch.nan
with pytest.raises(
ValueError, match="NaNs detected in the merged weights. The adapter default seems to be broken"
):
model = model.merge_and_unload(safe_merge=True)
for name, module in model.named_parameters():
if (
"lora_A" in name
or "ia3" in name
or "lora_E" in name
or "lora_B" in name
or "vera_lambda" in name
or "fourierft_spectrum" in name
):
module.data[0] = torch.inf
with pytest.raises(
ValueError, match="NaNs detected in the merged weights. The adapter default seems to be broken"
):
model = model.merge_and_unload(safe_merge=True)
def _test_merge_layers(self, model_id, config_cls, config_kwargs):
if issubclass(config_cls, PromptLearningConfig):
return pytest.skip(f"Test not applicable for {config_cls}")
if issubclass(config_cls, (OFTConfig, BOFTConfig)):
return pytest.skip(f"Test not applicable for {config_cls}")
if config_kwargs.get("alora_invocation_tokens") is not None:
return pytest.skip("Merging not applicable to aLoRA")
if ("gpt2" in model_id.lower()) and (config_cls != LoraConfig):
self.skipTest("Merging GPT2 adapters not supported for IA³ (yet)")
if "gemma" in model_id.lower():
# TODO: could be related to tied weights
self.skipTest("Merging currently fails with gemma")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
self.perturb_trainable_token_weights_if_used(model, config_kwargs)
dummy_input = self.prepare_inputs_for_testing()
model.eval()
logits = model(**dummy_input)[0]
model.merge_adapter()
logits_merged = model(**dummy_input)[0]
model.unmerge_adapter()
logits_unmerged = model(**dummy_input)[0]
model = model.merge_and_unload()
# check that PEFT layers are completely removed
assert not any(isinstance(module, BaseTunerLayer) for module in model.modules())
logits_merged_unloaded = model(**dummy_input)[0]
conv_ids = ["Conv2d", "Conv3d", "Conv2d2"]
atol, rtol = 1e-4, 1e-4
if self.torch_device in ["mlu"]:
atol, rtol = 1e-3, 1e-3 # MLU
if config.peft_type == "ADALORA":
# AdaLoRA is a bit flaky on CI, but this cannot be reproduced locally
atol, rtol = 1e-2, 1e-2
if (config.peft_type in {"IA3", "LORA"}) and (model_id in conv_ids):
# for some reason, the Conv introduces a larger error
atol, rtol = 0.3, 0.01
if model_id == "trl-internal-testing/tiny-Llama4ForCausalLM":
# also getting larger errors here, not exactly sure why
atol, rtol = 0.3, 0.01
assert torch.allclose(logits, logits_merged, atol=atol, rtol=rtol)
assert torch.allclose(logits, logits_unmerged, atol=atol, rtol=rtol)
assert torch.allclose(logits, logits_merged_unloaded, atol=atol, rtol=rtol)
# For this test to work, weights should not be initialized to identity transform (e.g.
# init_lora_weights should be False).
transformers_model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
logits_transformers = transformers_model(**dummy_input)[0]
assert not torch.allclose(logits_merged, logits_transformers, atol=1e-10, rtol=1e-10)
# test that the logits are identical after a save-load-roundtrip
if hasattr(model, "save_pretrained"):
# model is a transformers model
tmp_dirname = tempfile.mkdtemp()
# note: not using the context manager here because it fails on Windows CI for some reason
try:
model.save_pretrained(tmp_dirname)
model_from_pretrained = self.transformers_class.from_pretrained(tmp_dirname).to(self.torch_device)
finally:
try:
shutil.rmtree(tmp_dirname)
except PermissionError:
# windows error
pass
else:
# model is not a transformers model
model_from_pretrained = pickle.loads(pickle.dumps(model))
logits_merged_from_pretrained = model_from_pretrained(**dummy_input)[0]
assert torch.allclose(logits_merged, logits_merged_from_pretrained, atol=atol, rtol=rtol)
def _test_merge_layers_multi(self, model_id, config_cls, config_kwargs):
supported_peft_types = [
PeftType.LORA,
PeftType.LOHA,
PeftType.LOKR,
PeftType.IA3,
PeftType.OFT,
PeftType.BOFT,
PeftType.HRA,
PeftType.BONE,
PeftType.MISS,
]
if ("gpt2" in model_id.lower()) and (config_cls == IA3Config):
self.skipTest("Merging GPT2 adapters not supported for IA³ (yet)")
if config_kwargs.get("trainable_token_indices", None) is not None:
self.skipTest(
"Merging two adapters with trainable tokens is tested elsewhere since adapters with "
"the same token indices cannot be merged."
)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
if config.peft_type not in supported_peft_types or config_kwargs.get("alora_invocation_tokens") is not None:
return
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
dummy_input = self.prepare_inputs_for_testing()
model.eval()
with torch.inference_mode():
logits_adapter_1 = model(**dummy_input)[0]
model.add_adapter("adapter-2", config)
model.set_adapter("adapter-2")
model.eval()
# sanity check: each adapter layer with a 'default' adapter should also have 'adapter-2'
containers = (torch.nn.ModuleDict, torch.nn.ParameterDict, BufferDict)
num_default = len([m for m in model.modules() if isinstance(m, containers) and "default" in m])
num_adapter2 = len([m for m in model.modules() if isinstance(m, containers) and "adapter-2" in m])
assert num_default > 0
assert num_default == num_adapter2
with torch.inference_mode():
logits_adapter_2 = model(**dummy_input)[0]
assert not torch.allclose(logits_adapter_1, logits_adapter_2, atol=1e-3, rtol=1e-3)
model.set_adapter("default")
with torch.inference_mode():
logits_adapter_1_after_set = model(**dummy_input)[0]
assert torch.allclose(logits_adapter_1_after_set, logits_adapter_1, atol=1e-3, rtol=1e-3)
model_copy = copy.deepcopy(model)
model_copy_2 = copy.deepcopy(model)
model_merged_all = model.merge_and_unload(adapter_names=["adapter-2", "default"])
with torch.inference_mode():
logits_merged_all = model_merged_all(**dummy_input)[0]
assert not torch.allclose(logits_merged_all, logits_adapter_2, atol=1e-3, rtol=1e-3)
assert not torch.allclose(logits_merged_all, logits_adapter_1, atol=1e-3, rtol=1e-3)
model_merged_adapter_2 = model_copy.merge_and_unload(adapter_names=["adapter-2"])
with torch.inference_mode():
logits_merged_adapter_2 = model_merged_adapter_2(**dummy_input)[0]
assert torch.allclose(logits_merged_adapter_2, logits_adapter_2, atol=1e-3, rtol=1e-3)
model_merged_adapter_default = model_copy_2.merge_and_unload(adapter_names=["default"])
with torch.inference_mode():
logits_merged_adapter_default = model_merged_adapter_default(**dummy_input)[0]
assert torch.allclose(logits_merged_adapter_default, logits_adapter_1, atol=1e-3, rtol=1e-3)
def _test_merge_layers_is_idempotent(self, model_id, config_cls, config_kwargs):
if config_kwargs.get("alora_invocation_tokens") is not None:
# Merging not supported for Activated LoRA (aLoRA)
return pytest.skip("Test not applicable for Activated LoRA (aLoRA)")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
model.eval()
torch.manual_seed(0)
model.merge_adapter()
logits_0 = model(**self.prepare_inputs_for_testing())[0]
# merging again should not change anything
# also check warning:
with pytest.warns(UserWarning, match="All adapters are already merged, nothing to do"):
model.merge_adapter()
logits_1 = model(**self.prepare_inputs_for_testing())[0]
assert torch.allclose(logits_0, logits_1, atol=1e-6, rtol=1e-6)
def _test_safe_merge(self, model_id, config_cls, config_kwargs):
if config_kwargs.get("alora_invocation_tokens") is not None:
# Merging not supported for Activated LoRA (aLoRA)
return pytest.skip("Test not applicable for Activated LoRA (aLoRA)")
torch.manual_seed(0)
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = model.to(self.torch_device).eval()
inputs = self.prepare_inputs_for_testing()
logits_base = model(**inputs)[0]
model = get_peft_model(model, config).eval()
logits_peft = model(**inputs)[0]
atol, rtol = 1e-6, 1e-6 # default
# Initializing with LN tuning cannot be configured to change the outputs (unlike init_lora_weights=False)
if not issubclass(config_cls, LNTuningConfig):
# sanity check that the logits are different
assert not torch.allclose(logits_base, logits_peft, atol=atol, rtol=rtol)
model_unloaded = model.merge_and_unload(safe_merge=True)
logits_unloaded = model_unloaded(**inputs)[0]
if self.torch_device in ["mlu"]:
atol, rtol = 1e-3, 1e-3 # MLU
conv_ids = ["Conv2d", "Conv3d", "Conv2d2"]
if issubclass(config_cls, (IA3Config, LoraConfig)) and model_id in conv_ids: # more instability with Conv
atol, rtol = 1e-3, 1e-3
# check that the logits are the same after unloading
assert torch.allclose(logits_peft, logits_unloaded, atol=atol, rtol=rtol)
# Ensure that serializing with safetensors works, there was an error when weights were not contiguous
with tempfile.TemporaryDirectory() as tmp_dirname:
# serializing with torch.save works
torch.save(model_unloaded.state_dict(), os.path.join(tmp_dirname, "model.bin"))
# serializing with safetensors works
save_file(model_unloaded.state_dict(), os.path.join(tmp_dirname, "model.safetensors"))
def _test_mixed_adapter_batches(self, model_id, config_cls, config_kwargs):
# Test for mixing different adapters in a single batch by passing the adapter_names argument
if config_cls not in (LoraConfig,):
return pytest.skip(f"Mixed adapter batches not supported for {config_cls}")
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
torch.manual_seed(0)
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config, adapter_name="adapter0").eval()
model.add_adapter("adapter1", config)
model = model.to(self.torch_device).eval()
self.perturb_trainable_token_weights_if_used(model, config_kwargs, adapter_name="adapter0")
self.perturb_trainable_token_weights_if_used(model, config_kwargs, adapter_name="adapter1")
dummy_input = self.prepare_inputs_for_testing()
# ensure that we have at least 3 samples for this test
dummy_input = {k: torch.cat([v for _ in range(3)]) for k, v in dummy_input.items()}
with torch.inference_mode():
with model.disable_adapter():
output_base = model(**dummy_input)[0]
logits_base = model.generate(**dummy_input, return_dict_in_generate=True, output_scores=True).scores[0]
model.set_adapter("adapter0")
with torch.inference_mode():
output_adapter0 = model(**dummy_input)[0]
logits_adapter0 = model.generate(**dummy_input, return_dict_in_generate=True, output_scores=True).scores[0]
model.set_adapter("adapter1")
with torch.inference_mode():
output_adapter1 = model(**dummy_input)[0]
logits_adapter1 = model.generate(**dummy_input, return_dict_in_generate=True, output_scores=True).scores[0]
atol, rtol = 1e-4, 1e-4
# sanity check that there are enough outputs and that they are different
assert len(output_base) == len(output_adapter0) == len(output_adapter1) >= 3
assert len(logits_base) == len(logits_adapter0) == len(logits_adapter1) >= 3
assert not torch.allclose(output_base, output_adapter0, atol=atol, rtol=rtol)
assert not torch.allclose(output_base, output_adapter1, atol=atol, rtol=rtol)
assert not torch.allclose(output_adapter0, output_adapter1, atol=atol, rtol=rtol)
assert not torch.allclose(logits_base, logits_adapter0, atol=atol, rtol=rtol)
assert not torch.allclose(logits_base, logits_adapter1, atol=atol, rtol=rtol)
assert not torch.allclose(logits_adapter0, logits_adapter1, atol=atol, rtol=rtol)
# alternate between base model, adapter0, and adapter1
adapters = ["__base__", "adapter0", "adapter1"]
dummy_input["adapter_names"] = [adapters[i % 3] for i in (range(len(dummy_input["input_ids"])))]
with torch.inference_mode():
output_mixed = model(**dummy_input)[0]
logits_mixed = model.generate(**dummy_input, return_dict_in_generate=True, output_scores=True).scores[0]
assert torch.allclose(output_base[::3], output_mixed[::3], atol=atol, rtol=rtol)
assert torch.allclose(output_adapter0[1::3], output_mixed[1::3], atol=atol, rtol=rtol)
assert torch.allclose(output_adapter1[2::3], output_mixed[2::3], atol=atol, rtol=rtol)
assert torch.allclose(logits_base[::3], logits_mixed[::3], atol=atol, rtol=rtol)
assert torch.allclose(logits_adapter0[1::3], logits_mixed[1::3], atol=atol, rtol=rtol)
assert torch.allclose(logits_adapter1[2::3], logits_mixed[2::3], atol=atol, rtol=rtol)
def _test_generate_with_mixed_adapter_batches_and_beam_search(self, model_id, config_cls, config_kwargs):
# Test generating with beam search and with mixing different adapters in a single batch by passing the
# adapter_names argument. See #2283.
if config_cls not in (LoraConfig,):
return pytest.skip(f"Mixed adapter batches not supported for {config_cls}")
if config_kwargs.get("alora_invocation_tokens") is not None:
return pytest.skip("Beam search not yet supported for aLoRA") # beam search not yet fully supported
if config_kwargs.get("trainable_token_indices", None) is not None:
# for some configurations this test will fail since the adapter values don't differ.
# this is probably a problem with the test setup and not with the implementation.
return pytest.skip("Trainable token indices is not supported here (yet).")
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
torch.manual_seed(0)
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config, adapter_name="adapter0").eval()
model.add_adapter("adapter1", config)
# In contrast to forward, for generate, it can sometimes happen that we get the same results as the base model
# even with LoRA applied because the impact of LoRA is not big enough. Therefore, use this "trick" to make LoRA
# stronger.
for name, param in model.named_parameters():
if model.base_model.prefix in name:
param.data.mul_(10.0)
model = model.to(self.torch_device).eval()
dummy_input = self.prepare_inputs_for_testing()
# ensure that we have at least 3 samples for this test
dummy_input = {k: torch.cat([v for _ in range(3)]) for k, v in dummy_input.items()}
gen_kwargs = {**dummy_input, "max_length": 20, "num_beams": 10, "early_stopping": True}
with torch.inference_mode():
with model.disable_adapter():
gen_base = model.generate(**gen_kwargs)
model.set_adapter("adapter0")
with torch.inference_mode():
gen_adapter0 = model.generate(**gen_kwargs)
model.set_adapter("adapter1")
with torch.inference_mode():
gen_adapter1 = model.generate(**gen_kwargs)
def remove_padding(seq, pad_value):
lst = list(seq)
while lst and (lst[-1] == pad_value):
lst.pop()
return lst
def gens_are_same(gen0, gen1):
# Special function to compare generations. We cannot use torch.allclose it will raise an error when sequence
# lengths differ. Morevoer, we need to remove the padding from the sequences. This is because, even though
# normally identical sequences should have the same length, when we do mixed adapter batches, each sample
# will be padded to the longest sequence in that mixed batch, which can be different from the longest
# sequence without mixed adapter batches.
pad_value = model.config.eos_token_id
for sample0, sample1 in zip(gen0, gen1):
sample0 = remove_padding(sample0, pad_value)
sample1 = remove_padding(sample1, pad_value)
if (len(sample0) != len(sample1)) or (sample0 != sample1):
# at least one sample differs, the generations are not identical
return False
return True
# sanity check that there are enough outputs and that they are different
assert len(gen_base) == len(gen_adapter0) == len(gen_adapter1)
assert len(gen_adapter1) >= 3
assert not gens_are_same(gen_base, gen_adapter0)
assert not gens_are_same(gen_base, gen_adapter1)
assert not gens_are_same(gen_adapter0, gen_adapter1)
# alternate between base model, adapter0, and adapter1
adapters = ["__base__", "adapter0", "adapter1"]
gen_kwargs["adapter_names"] = [adapters[i % 3] for i in (range(len(dummy_input["input_ids"])))]
with torch.inference_mode():
gen_mixed = model.generate(**gen_kwargs)
assert gens_are_same(gen_base[::3], gen_mixed[::3])
assert gens_are_same(gen_adapter0[1::3], gen_mixed[1::3])
assert gens_are_same(gen_adapter1[2::3], gen_mixed[2::3])
def _test_generate(self, model_id, config_cls, config_kwargs):
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `generate` works
_ = model.generate(**inputs)
def _test_generate_pos_args(self, model_id, config_cls, config_kwargs, raises_err: bool):
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
if raises_err:
with pytest.raises(TypeError):
# check if `generate` raises an error if positional arguments are passed
_ = model.generate(inputs["input_ids"])
else:
# check if `generate` works if positional arguments are passed
_ = model.generate(inputs["input_ids"])
def _test_generate_half_prec(self, model_id, config_cls, config_kwargs):
if config_cls not in (IA3Config, LoraConfig, PrefixTuningConfig):
return pytest.skip(f"Test not applicable for {config_cls}")
if self.torch_device == "mps": # BFloat16 is not supported on MPS
return pytest.skip("BFloat16 is not supported on MPS")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id, torch_dtype=torch.bfloat16)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
input_ids = torch.LongTensor([[1, 1, 1], [2, 1, 2]]).to(self.torch_device)
attention_mask = torch.LongTensor([[1, 1, 1], [1, 0, 1]]).to(self.torch_device)
# check if `generate` works
_ = model.generate(input_ids=input_ids, attention_mask=attention_mask)
def _test_prefix_tuning_half_prec_conversion(self, model_id, config_cls, config_kwargs):
if config_cls not in (PrefixTuningConfig,):
return pytest.skip(f"Test not applicable for {config_cls}")
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config)
model = model.half()
assert model.base_model_torch_dtype == torch.float16
def _test_training(self, model_id, config_cls, config_kwargs):
if issubclass(config_cls, PromptLearningConfig):
return pytest.skip(f"Test not applicable for {config_cls}")
if (config_cls == AdaLoraConfig) and ("roberta" in model_id.lower()):
# TODO: no gradients on the "dense" layer, other layers work, not sure why
self.skipTest("AdaLora with RoBERTa does not work correctly")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
loss = output.sum()
loss.backward()
parameter_prefix = model.prefix
for n, param in model.named_parameters():
if (parameter_prefix in n) or ("modules_to_save" in n) or ("token_adapter.trainable_tokens" in n):
assert param.grad is not None
else:
assert param.grad is None
def _test_inference_safetensors(self, model_id, config_cls, config_kwargs):
if (config_cls == PrefixTuningConfig) and ("deberta" in model_id.lower()):
# TODO: raises an error:
# TypeError: DebertaModel.forward() got an unexpected keyword argument 'past_key_values'
self.skipTest("DeBERTa with PrefixTuning does not work correctly")
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
logits = output[0]
loss = output.sum()
loss.backward()
# set to eval mode, since things like dropout can affect the output otherwise
model.eval()
logits = model(**inputs)[0][0]
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname, safe_serialization=True)
assert "adapter_model.safetensors" in os.listdir(tmp_dirname)
assert "adapter_model.bin" not in os.listdir(tmp_dirname)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname).to(
self.torch_device
)
logits_from_pretrained = model_from_pretrained(**inputs)[0][0]
assert torch.allclose(logits, logits_from_pretrained, atol=1e-4, rtol=1e-4)
def _test_training_layer_indexing(self, model_id, config_cls, config_kwargs):
if config_cls not in (LoraConfig,):
return pytest.skip(f"Test not applicable for {config_cls}")
config = config_cls(
base_model_name_or_path=model_id,
layers_to_transform=[0],
**config_kwargs,
)
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
logits = output[0]
loss = output.sum()
loss.backward()
has_trainable_tokens = config_kwargs.get("trainable_token_indices", None) is not None
nb_trainable = 0
for n, param in model.named_parameters():
if model.prefix in n or (has_trainable_tokens and "trainable_tokens" in n):
assert param.grad is not None
nb_trainable += 1
else:
assert param.grad is None
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
model_from_pretrained = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname).to(
self.torch_device
)
logits_from_pretrained = model_from_pretrained(**inputs)[0][0]
assert torch.allclose(logits, logits_from_pretrained, atol=1e-4, rtol=1e-4)
# check the nb of trainable params again but without layers_to_transform
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
nb_trainable_all = 0
for n, param in model.named_parameters():
if model.prefix in n or (has_trainable_tokens and "trainable_tokens" in n):
nb_trainable_all += 1
mod_list = next((m for m in model.modules() if isinstance(m, torch.nn.ModuleList)), None)
if mod_list and len(mod_list) == 1:
# there is only a single layer
assert nb_trainable == nb_trainable_all
else:
# more than 1 layer, i.e. setting layers_to_transform=[0] should target fewer layers
assert nb_trainable < nb_trainable_all
def _test_training_gradient_checkpointing(self, model_id, config_cls, config_kwargs):
if config_cls == PrefixTuningConfig:
return pytest.skip(f"Test not applicable for {config_cls}")
if (config_cls == AdaLoraConfig) and ("roberta" in model_id.lower()):
# TODO: no gradients on the "dense" layer, other layers work, not sure why
self.skipTest("AdaLora with RoBERTa does not work correctly")
if (config_cls == OFTConfig) and ("deberta" in model_id.lower()):
# TODO: no gradients on the "dense" layer, other layers work, not sure why
self.skipTest("OFT with Deberta does not work correctly")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
if not getattr(model, "supports_gradient_checkpointing", False):
return pytest.skip(f"Model {model_id} does not support gradient checkpointing")
model.gradient_checkpointing_enable()
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
loss = output.sum()
loss.backward()
for n, param in model.named_parameters():
if "prompt_encoder." in n: # prompt tuning methods
if not issubclass(config_cls, CPTConfig):
assert param.grad is not None
elif (
"delta_embedding" in n
): # delta_embedding is the embedding that should be updated with grads in CPT
assert param.grad is not None
elif hasattr(model, "prefix") and (model.prefix in n): # non-prompt tuning methods
assert param.grad is not None
elif "trainable_tokens_" in n: # trainable tokens layer
assert param.grad is not None
else:
assert param.grad is None
def _test_peft_model_device_map(self, model_id, config_cls, config_kwargs):
if config_cls not in (LoraConfig, VBLoRAConfig):
return pytest.skip(f"Test not applicable for {config_cls}")
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
with tempfile.TemporaryDirectory() as tmp_dirname:
model.save_pretrained(tmp_dirname)
model_from_pretrained = self.transformers_class.from_pretrained(model_id)
_ = PeftModel.from_pretrained(model_from_pretrained, tmp_dirname, device_map={"": "cpu"}).to(
self.torch_device
)
def _test_training_prompt_learning_tasks(self, model_id, config_cls, config_kwargs):
if not issubclass(config_cls, PromptLearningConfig):
return pytest.skip(f"Test not applicable for {config_cls}")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
inputs = self.prepare_inputs_for_testing()
# check if `training` works
output = model(**inputs)[0]
loss = output.sum()
loss.backward()
if issubclass(config_cls, CPTConfig):
parameters = []
for name, param in model.prompt_encoder.named_parameters():
if name != "default.embedding.weight":
parameters.append(param)
else:
parameters = model.prompt_encoder.parameters()
# check that prompt encoder has grads
for param in parameters:
assert param.grad is not None
def _test_delete_adapter(self, model_id, config_cls, config_kwargs):
supported_peft_types = [
PeftType.LORA,
PeftType.LOHA,
PeftType.LOKR,
PeftType.IA3,
PeftType.OFT,
PeftType.BOFT,
PeftType.VERA,
PeftType.FOURIERFT,
PeftType.HRA,
PeftType.VBLORA,
PeftType.BONE,
PeftType.MISS,
]
# IA3 does not support deleting adapters yet, but it just needs to be added
# AdaLora does not support multiple adapters
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
if config.peft_type not in supported_peft_types:
return pytest.skip(f"Test not applicable for {config.peft_type}")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
adapter_to_delete = "delete_me"
model = get_peft_model(model, config)
model.add_adapter(adapter_to_delete, config)
model.set_adapter(adapter_to_delete)
model = model.to(self.torch_device)
model.delete_adapter(adapter_to_delete)
assert adapter_to_delete not in model.peft_config
assert model.active_adapters == ["default"]
key_list = [key for key, _ in model.named_modules()]
for key in key_list:
_, target, _ = _get_submodules(model, key)
attributes_to_check = getattr(target, "adapter_layer_names", []) + getattr(
target, "other_param_names", []
)
for attr in attributes_to_check:
assert adapter_to_delete not in getattr(target, attr)
# check auxiliary modules
for module in model.modules():
if isinstance(module, AuxiliaryTrainingWrapper):
assert adapter_to_delete not in module._adapters
assert module.active_adapters == ["default"]
if isinstance(module, ModulesToSaveWrapper):
assert adapter_to_delete not in module.modules_to_save
elif isinstance(module, TrainableTokensWrapper):
assert adapter_to_delete not in module.token_adapter.trainable_tokens_delta
assert adapter_to_delete not in module.token_adapter.trainable_tokens_original
# check that we can also delete the last remaining adapter
model.delete_adapter("default")
assert "default" not in model.peft_config
assert model.active_adapters == []
for module in model.modules():
if isinstance(module, AuxiliaryTrainingWrapper):
assert "default" not in module._adapters
assert module.active_adapters == []
if isinstance(module, ModulesToSaveWrapper):
assert "default" not in module.modules_to_save
elif isinstance(module, TrainableTokensWrapper):
assert "default" not in module.token_adapter.trainable_tokens_delta
assert "default" not in module.token_adapter.trainable_tokens_original
input = self.prepare_inputs_for_testing()
# note: we cannot call model(**input) because PeftModel always expects there to be at least one adapter
model.base_model(**input) # should not raise an error
def _test_delete_inactive_adapter(self, model_id, config_cls, config_kwargs):
# same as test_delete_adapter, but this time an inactive adapter is deleted
supported_peft_types = [
PeftType.LORA,
PeftType.LOHA,
PeftType.LOKR,
PeftType.IA3,
PeftType.OFT,
PeftType.BOFT,
PeftType.FOURIERFT,
PeftType.HRA,
PeftType.VBLORA,
PeftType.BONE,
PeftType.MISS,
]
# IA3 does not support deleting adapters yet, but it just needs to be added
# AdaLora does not support multiple adapters
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
if config.peft_type not in supported_peft_types:
return pytest.skip(f"Test not applicable for {config.peft_type}")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
adapter_to_delete = "delete_me"
model = get_peft_model(model, config)
model.add_adapter(adapter_to_delete, config)
# "delete_me" is added but not activated
model = model.to(self.torch_device)
model.delete_adapter(adapter_to_delete)
assert adapter_to_delete not in model.peft_config
assert model.active_adapters == ["default"]
key_list = [key for key, _ in model.named_modules()]
for key in key_list:
_, target, _ = _get_submodules(model, key)
attributes_to_check = getattr(target, "adapter_layer_names", []) + getattr(
target, "other_param_names", []
)
for attr in attributes_to_check:
assert adapter_to_delete not in getattr(target, attr)
# check auxiliary modules
for module in model.modules():
if isinstance(module, AuxiliaryTrainingWrapper):
assert adapter_to_delete not in module._adapters
assert module.active_adapters == ["default"]
if isinstance(module, ModulesToSaveWrapper):
assert adapter_to_delete not in module.modules_to_save
elif isinstance(module, TrainableTokensWrapper):
assert adapter_to_delete not in module.token_adapter.trainable_tokens_delta
assert adapter_to_delete not in module.token_adapter.trainable_tokens_original
# check that we can also delete the last remaining adapter
model.delete_adapter("default")
assert "default" not in model.peft_config
assert model.active_adapters == []
for module in model.modules():
if isinstance(module, AuxiliaryTrainingWrapper):
assert "default" not in module._adapters
assert module.active_adapters == []
if isinstance(module, ModulesToSaveWrapper):
assert "default" not in module.modules_to_save
elif isinstance(module, TrainableTokensWrapper):
assert "default" not in module.token_adapter.trainable_tokens_delta
assert "default" not in module.token_adapter.trainable_tokens_original
input = self.prepare_inputs_for_testing()
# note: we cannot call model(**input) because PeftModel always expects there to be at least one adapter
model.base_model(**input) # should not raise an error
def _test_delete_unknown_adapter_raises(self, model_id, config_cls, config_kwargs):
# Check that we get a nice error message when trying to delete an adapter that does not exist.
config = config_cls(base_model_name_or_path=model_id, **config_kwargs)
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
adapter_to_delete = "delete_me"
model = get_peft_model(model, config)
msg = "Adapter unknown-adapter does not exist"
with pytest.raises(ValueError, match=msg):
model.delete_adapter("unknown-adapter")
def _test_unload_adapter(self, model_id, config_cls, config_kwargs):
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
num_params_base = len(model.state_dict())
dummy_input = self.prepare_inputs_for_testing()
with torch.inference_mode():
logits_transformers = model(**dummy_input)[0]
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config)
model = model.to(self.torch_device)
if isinstance(config, PromptLearningConfig):
# prompt learning does not support unloading
with pytest.raises(AttributeError):
model = model.unload()
else:
self.perturb_trainable_token_weights_if_used(model, config_kwargs)
with torch.inference_mode():
logits_with_adapter = model(**dummy_input)[0]
model.eval()
model = model.unload()
num_params_unloaded = len(model.state_dict())
with torch.inference_mode():
logits_unload = model(**dummy_input)[0]
# check that PEFT layers are completely removed
assert not any(isinstance(module, BaseTunerLayer) for module in model.modules())
assert not torch.allclose(logits_with_adapter, logits_unload, atol=1e-10, rtol=1e-10)
assert torch.allclose(logits_transformers, logits_unload, atol=1e-4, rtol=1e-4)
assert num_params_base == num_params_unloaded
def _test_weighted_combination_of_adapters_lora(self, model, config, adapter_list, weight_list):
model.add_adapter(adapter_list[1], config)
model.add_adapter(adapter_list[2], replace(config, r=20))
model = model.to(self.torch_device)
# test re-weighting single adapter
model.add_weighted_adapter([adapter_list[0]], [weight_list[0]], "single_adapter_reweighting")
# test svd re-weighting with multiple adapters
model.add_weighted_adapter(adapter_list[1:], weight_list[1:], "multi_adapter_svd_reweighting")
# test ties_svd re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_ties_svd_reweighting",
combination_type="ties_svd",
density=0.5,
)
# test dare_linear_svd re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_dare_linear_svd_reweighting",
combination_type="dare_linear_svd",
density=0.5,
)
# test dare_ties_svd re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_dare_ties_svd_reweighting",
combination_type="dare_ties_svd",
density=0.5,
)
# test magnitude_prune_svd re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_magnitude_prune_svd_reweighting",
combination_type="magnitude_prune_svd",
density=0.5,
)
# test cat re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[1:], weight_list[1:], "multi_adapter_cat_reweighting", combination_type="cat"
)
# test linear re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[:2], weight_list[:2], "multi_adapter_linear_reweighting", combination_type="linear"
)
# test ties re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[:2], weight_list[:2], "multi_adapter_ties_reweighting", combination_type="ties", density=0.5
)
# test dare_linear re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[:2],
weight_list[:2],
"multi_adapter_dare_linear_reweighting",
combination_type="dare_linear",
density=0.5,
)
# test dare_ties re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[:2],
weight_list[:2],
"multi_adapter_dare_ties_reweighting",
combination_type="dare_ties",
density=0.5,
)
# test magnitude_prune re-weighting with multiple adapters
model.add_weighted_adapter(
adapter_list[:2],
weight_list[:2],
"multi_adapter_magnitude_prune_reweighting",
combination_type="magnitude_prune",
density=0.5,
)
# test linear re-weighting with multiple adapters with only first adapter having non zero weight
model.add_weighted_adapter(
adapter_list[:2],
[weight_list[0], 0],
"multi_adapter_linear_reweighting_single_enabled",
combination_type="linear",
)
with pytest.raises(ValueError):
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_linear_reweighting_uneven_r",
combination_type="linear",
)
with pytest.raises(ValueError):
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_ties_reweighting_uneven_r",
combination_type="ties",
density=0.5,
)
with pytest.raises(ValueError):
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_dare_linear_reweighting_uneven_r",
combination_type="dare_linear",
density=0.5,
)
with pytest.raises(ValueError):
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_dare_ties_reweighting_uneven_r",
combination_type="dare_ties",
density=0.5,
)
with pytest.raises(ValueError):
model.add_weighted_adapter(
adapter_list[1:],
weight_list[1:],
"multi_adapter_magnitude_prune_reweighting_uneven_r",
combination_type="magnitude_prune",
density=0.5,
)
new_adapters = [
"single_adapter_reweighting",
"multi_adapter_svd_reweighting",
"multi_adapter_ties_svd_reweighting",
"multi_adapter_dare_linear_svd_reweighting",
"multi_adapter_dare_ties_svd_reweighting",
"multi_adapter_magnitude_prune_svd_reweighting",
"multi_adapter_cat_reweighting",
"multi_adapter_linear_reweighting",
"multi_adapter_linear_reweighting_single_enabled",
"multi_adapter_ties_reweighting",
"multi_adapter_dare_linear_reweighting",
"multi_adapter_dare_ties_reweighting",
"multi_adapter_magnitude_prune_reweighting",
]
for new_adapter in new_adapters:
assert new_adapter in model.peft_config
key_list = [key for key, _ in model.named_modules()]
for key in key_list:
_, target, _ = _get_submodules(model, key)
if isinstance(target, LoraLayer):
for adapter_name in new_adapters:
if "single" in adapter_name:
new_delta_weight = target.get_delta_weight(adapter_name)
weighted_original_delta_weights = target.get_delta_weight(adapter_list[0]) * weight_list[0]
assert torch.allclose(new_delta_weight, weighted_original_delta_weights, atol=1e-4, rtol=1e-4)
elif "svd" in adapter_name:
assert target.r[adapter_name] == 20
elif "linear" in adapter_name:
assert target.r[adapter_name] == 8
elif "cat" in adapter_name:
assert target.r[adapter_name] == 28
dummy_input = self.prepare_inputs_for_testing()
model.eval()
for adapter_name in new_adapters:
# ensuring new adapters pass the forward loop
model.set_adapter(adapter_name)
assert model.active_adapter == adapter_name
assert model.active_adapters == [adapter_name]
model(**dummy_input)[0]
def _test_weighted_combination_of_adapters_ia3(self, model, config, adapter_list, weight_list):
model.add_adapter(adapter_list[1], config)
model.add_adapter(adapter_list[2], config)
model = model.to(self.torch_device)
# test re-weighting single adapter
model.add_weighted_adapter([adapter_list[0]], [weight_list[0]], "single_adapter_reweighting")
# test re-weighting with multiple adapters
model.add_weighted_adapter(adapter_list[1:], weight_list[1:], "multi_adapter_reweighting")
new_adapters = [
"single_adapter_reweighting",
"multi_adapter_reweighting",
]
for new_adapter in new_adapters:
assert new_adapter in model.peft_config
dummy_input = self.prepare_inputs_for_testing()
model.eval()
for adapter_name in new_adapters:
# ensuring new adapters pass the forward loop
model.set_adapter(adapter_name)
assert model.active_adapter == adapter_name
assert model.active_adapters == [adapter_name]
model(**dummy_input)[0]
def _test_weighted_combination_of_adapters(self, model_id, config_cls, config_kwargs):
if issubclass(config_cls, AdaLoraConfig):
# AdaLora does not support adding more than 1 adapter
return pytest.skip(f"Test not applicable for {config_cls}")
if model_id.endswith("qwen2"):
# Qwen2 fails with weighted adapter combinations using SVD
return pytest.skip(f"Test does not work with model {model_id}")
if "gemma" in model_id.lower():
return pytest.skip("Combining Gemma adapters with SVD is currently failing")
adapter_list = ["adapter1", "adapter_2", "adapter_3"]
weight_list = [0.5, 1.5, 1.5]
# Initialize the config
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
if not isinstance(config, (LoraConfig, IA3Config)):
# This test is only applicable for Lora and IA3 configs
return pytest.skip(f"Test not applicable for {config}")
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config, adapter_list[0])
if isinstance(config, LoraConfig):
self._test_weighted_combination_of_adapters_lora(model, config, adapter_list, weight_list)
elif isinstance(config, IA3Config):
self._test_weighted_combination_of_adapters_ia3(model, config, adapter_list, weight_list)
else:
pytest.skip(f"Test not applicable for {config}")
def _test_disable_adapter(self, model_id, config_cls, config_kwargs):
task_type = config_kwargs.get("task_type")
if (task_type == "SEQ_2_SEQ_LM") and (config_cls in (PromptTuningConfig, PromptEncoderConfig)):
self.skipTest("Seq2Seq + prompt tuning/prompt encoder does not work with disabling adapters")
def get_output(model):
# helper function that works with different model types
torch.manual_seed(0)
if hasattr(model, "generate"):
# let's check the scores, not the output ids, since the latter can easily be identical even if the
# weights are slightly changed
output = model.generate(**input, return_dict_in_generate=True, output_scores=True).scores[0]
# take element 0, as output is a tuple
else:
output = model(**input)
if hasattr(output, "images"): # for SD
import numpy as np
img = output.images[0]
return torch.from_numpy(np.array(img))
return output
# initialize model
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id).to(self.torch_device)
# output from BASE MODEL
input = self.prepare_inputs_for_testing()
output_before = get_output(model)
# output from PEFT MODEL
if hasattr(self, "instantiate_sd_peft"):
# SD models are instantiated differently
peft_model = self.instantiate_sd_peft(model_id, config_cls, config_kwargs)
else:
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
peft_model = get_peft_model(model, config)
# trainable_token_indices doesn't have support for `init_weights` so we have to do this manually
self.perturb_trainable_token_weights_if_used(model, config_kwargs)
output_peft = get_output(peft_model)
# first check trivial case is not true that peft does not affect the output; for this to work, init_weight
# must be False (if the config supports it)
if isinstance(peft_model, StableDiffusionPipeline):
# for SD, check that most pixels have different values
assert (output_before != output_peft).float().mean() > 0.8
else:
assert not torch.allclose(output_before, output_peft)
# output with DISABLED ADAPTER
if isinstance(peft_model, StableDiffusionPipeline):
with peft_model.unet.disable_adapter():
with peft_model.text_encoder.disable_adapter():
output_peft_disabled = get_output(peft_model)
# for SD, very rarely, a pixel can differ
assert (output_before != output_peft_disabled).float().mean() < 1e-4
else:
atol, rtol = 1e-6, 1e-6
if (platform.system() == "Windows") and (model_id == "trl-internal-testing/tiny-Llama4ForCausalLM"):
# for some reason, Windows CI fails with stricter tolerance
atol, rtol = 1e-5, 1e-5
with peft_model.disable_adapter():
output_peft_disabled = get_output(peft_model)
assert torch.allclose(output_before, output_peft_disabled, atol=atol, rtol=rtol)
# after leaving the disable_adapter context, the output should be the same as with enabled adapter again
# see #1501
output_peft_after_disabled = get_output(peft_model)
assert torch.allclose(output_peft, output_peft_after_disabled, atol=atol, rtol=rtol)
# TODO: add tests to check if disabling adapters works after calling merge_adapter
def _test_adding_multiple_adapters_with_bias_raises(self, model_id, config_cls, config_kwargs):
# When trying to add multiple adapters with bias in Lora, AdaLora or BOFTConfig, an error should be
# raised. Also, the peft model should not be left in a half-initialized state.
if not issubclass(config_cls, (LoraConfig, AdaLoraConfig, BOFTConfig)):
return pytest.skip(f"Test not applicable for {config_cls}")
with hub_online_once(model_id):
config_kwargs = config_kwargs.copy()
config_kwargs["bias"] = "all"
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = self.transformers_class.from_pretrained(model_id)
model = get_peft_model(model, config, "adapter0")
if config_cls == LoraConfig or config_cls == AdaLoraConfig:
with pytest.raises(ValueError):
model.add_adapter("adapter1", replace(config, r=20))
if config_cls == BOFTConfig:
with pytest.raises(ValueError):
model.add_adapter("adapter1", replace(config, boft_block_num=1, boft_block_size=0))
# (superficial) test that the model is not left in a half-initialized state when adding an adapter fails
assert "adapter1" not in model.peft_config
assert "adapter1" not in model.base_model.peft_config
def _test_passing_input_embeds_works(self, test_name, model_id, config_cls, config_kwargs):
# https://github.com/huggingface/peft/issues/727
with hub_online_once(model_id):
model = self.transformers_class.from_pretrained(model_id)
config = config_cls(
base_model_name_or_path=model_id,
**config_kwargs,
)
model = get_peft_model(model, config, adapter_name="test-adapter").to(self.torch_device)
dummy_input = self.prepare_inputs_for_testing()
inputs_embeds = model.get_input_embeddings()(dummy_input["input_ids"])
# just check that no error is raised
model.forward(inputs_embeds=inputs_embeds)
Reference in New Issue View Git Blame Copy Permalink