Require peft>=0.17.0 and drop the target_parameters guards (#7640)

Author: albertvillanova

.github/workflows/tests.yml

@@ -248,6 +248,7 @@ jobs:
             ".[dev]" \
             "accelerate==1.10.1" \
             "datasets==4.7.0" \
+            "peft==0.17.0" \
             "torch==2.6.0" \
             "transformers==4.56.2"
 

pyproject.toml

@@ -71,7 +71,7 @@ openreward = [
     "openreward>=0.1.109",
 ]
 peft = [
-    "peft>=0.13.0"
+    "peft>=0.17.0"
 ]
 quality = [
     "pre-commit",
@@ -108,7 +108,7 @@ dev = [
     # openreward
     "openreward>=0.1.109",
     # peft
-    "peft>=0.13.0",
+    "peft>=0.17.0",
     # quality
     "pre-commit",
     "hf-doc-builder",

tests/test_dpo_trainer.py

@@ -32,7 +32,6 @@
     require_kernels,
     require_liger_kernel,
     require_peft,
-    require_peft_target_parameters,
     require_torch_accelerator,
     require_vision,
 )
@@ -791,7 +790,7 @@ def test_train_dense_with_peft_config_lora(self):
             elif "base_layer" not in n:  # We expect the peft params to be different (except for the base layer)
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
-    @require_peft_target_parameters
+    @require_peft
     def test_train_moe_with_peft_config(self):
         model_id = "trl-internal-testing/tiny-GptOssForCausalLM"
         model = AutoModelForCausalLM.from_pretrained(model_id, dtype="float32")
@@ -859,7 +858,7 @@ def test_train_peft_model(self):
             elif "base_layer" not in n and "ref" not in n:  # and the peft params to be different (except base and ref)
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
-    @require_peft_target_parameters
+    @require_peft
     def test_train_moe_peft_model(self):
         # Regression test for https://github.com/huggingface/trl/issues/5222. Before PEFT 0.20.0, only one adapter per
         # model was supported when the LoRA config uses `target_parameters` (see peft#3340, fixed in peft#3350), so no

tests/test_grpo_trainer.py

@@ -43,7 +43,6 @@
     require_bitsandbytes,
     require_liger_kernel,
     require_peft,
-    require_peft_target_parameters,
     require_response_parsing,
     require_torch_accelerator,
     require_vision,
@@ -1125,7 +1124,7 @@ def test_ref_model_like_model_instance(self):
         assert trainer.ref_model.dtype == torch.bfloat16
         assert trainer.ref_model.config._attn_implementation == "eager"
 
-    @require_peft_target_parameters
+    @require_peft
     def test_train_moe_peft_model(self):
         # Regression test for https://github.com/huggingface/trl/issues/5222. Before PEFT 0.20.0, only one adapter per
         # model was supported when the LoRA config uses `target_parameters` (see peft#3340, fixed in peft#3350), so no

tests/test_kto_trainer.py

@@ -33,7 +33,6 @@
     require_bitsandbytes,
     require_liger_kernel,
     require_peft,
-    require_peft_target_parameters,
     require_torch_accelerator,
     require_vision,
 )
@@ -881,7 +880,7 @@ def test_train_peft_model(self):
             elif "base_layer" not in n and "ref" not in n:  # and the peft params to be different (except base and ref)
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
-    @require_peft_target_parameters
+    @require_peft
     def test_train_moe_peft_model(self):
         # Regression test for https://github.com/huggingface/trl/issues/5222. Before PEFT 0.20.0, only one adapter per
         # model was supported when the LoRA config uses `target_parameters` (see peft#3340, fixed in peft#3350), so no

tests/test_reward_trainer.py

@@ -28,7 +28,6 @@
     TrlTestCase,
     require_bitsandbytes,
     require_peft,
-    require_peft_target_parameters,
     require_torch_accelerator,
 )
 
@@ -428,7 +427,7 @@ def test_train_peft_dora_and_quantization(self):
                 assert param.dtype == torch.bfloat16, f"Parameter {n} is not in bfloat16."
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
-    @require_peft_target_parameters
+    @require_peft
     def test_train_moe_with_peft_config(self):
         model_id = "trl-internal-testing/tiny-GptOssForCausalLM"
         model = AutoModelForSequenceClassification.from_pretrained(model_id, dtype="float32")

tests/test_rloo_trainer.py

@@ -37,7 +37,6 @@
     require_bitsandbytes,
     require_liger_kernel,
     require_peft,
-    require_peft_target_parameters,
     require_torch_accelerator,
     require_vision,
     require_vllm,
@@ -784,7 +783,7 @@ def test_train_peft_model(self):
             elif "base_layer" not in n and "ref" not in n:  # and the peft params to be different (except base and ref)
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
-    @require_peft_target_parameters
+    @require_peft
     def test_train_moe_peft_model(self):
         # Regression test for https://github.com/huggingface/trl/issues/5222. Before PEFT 0.20.0, only one adapter per
         # model was supported when the LoRA config uses `target_parameters` (see peft#3340, fixed in peft#3350), so no

tests/test_sft_trainer.py

@@ -41,7 +41,6 @@
     require_kernels,
     require_liger_kernel,
     require_peft,
-    require_peft_target_parameters,
     require_torch_accelerator,
     require_vision,
 )
@@ -704,7 +703,7 @@ def test_train_with_peft_config_prompt_tuning(self, peft_type):
             else:  # We expect the peft params to be different
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
-    @require_peft_target_parameters
+    @require_peft
     def test_train_moe_with_peft_config(self):
         model_id = "trl-internal-testing/tiny-GptOssForCausalLM"
         model = AutoModelForCausalLM.from_pretrained(model_id, dtype="float32")

tests/testing_utils.py

@@ -20,7 +20,6 @@
 import psutil
 import pytest
 import torch
-from packaging.version import Version
 from transformers import is_bitsandbytes_available, is_comet_available, is_wandb_available
 from transformers.testing_utils import backend_device_count, torch_device
 from transformers.utils import (
@@ -47,10 +46,6 @@
 )
 
 
-if is_peft_available():
-    import peft
-
-
 require_bitsandbytes = pytest.mark.skipif(not is_bitsandbytes_available(), reason="test requires bitsandbytes")
 require_comet = pytest.mark.skipif(not is_comet_available(), reason="test requires comet_ml")
 require_harbor = pytest.mark.skipif(not is_harbor_available(), reason="test requires harbor")
@@ -60,11 +55,6 @@
 require_mergekit = pytest.mark.skipif(not is_mergekit_available(), reason="test requires mergekit")
 require_openreward = pytest.mark.skipif(not is_openreward_available(), reason="test requires openreward")
 require_peft = pytest.mark.skipif(not is_peft_available(), reason="test requires peft")
-# `LoraConfig.target_parameters` was added in peft 0.17.0; on older versions the field doesn't exist at all.
-require_peft_target_parameters = pytest.mark.skipif(
-    not is_peft_available() or Version(peft.__version__) < Version("0.17.0"),
-    reason="test requires peft>=0.17.0 for `LoraConfig.target_parameters`",
-)
 require_requests = pytest.mark.skipif(not is_requests_available(), reason="test requires requests")
 # Response parsing needs jmespath only on transformers < 5.13, which ships the legacy `response_schema` parser; the
 # new-style `response_template` parser doesn't use it. See `_SUPPORTS_RESPONSE_TEMPLATE`.

trl/trainer/dpo_trainer.py

@@ -669,12 +669,11 @@ def __init__(
             # 0.20.0, only one adapter per model was supported when the LoRA config uses `target_parameters` (see
             # peft#3340, fixed in peft#3350), so in that case we skip the "ref" adapter and compute the reference log
             # probs with adapters disabled, i.e. with the base model. The fix only allows adapters targeting the same
-            # parameters, which holds here since the "ref" adapter reuses the "default" config. `target_parameters`
-            # itself was only added in PEFT 0.17.0, so the version check is bounded on both sides.
+            # parameters, which holds here since the "ref" adapter reuses the "default" config.
             default_config = model.peft_config["default"]
             if (
                 isinstance(default_config, LoraConfig)
-                and Version("0.17.0") <= Version(peft.__version__) < Version("0.20.0")
+                and Version(peft.__version__) < Version("0.20.0")
                 and default_config.target_parameters
             ):
                 logger.warning(