Revert xfail for MoE LoRA tests now that peft#3602 fixed the autocast dtype bug (#6919)

Author: albertvillanova

tests/test_dpo_trainer.py

@@ -733,14 +733,6 @@ def test_train_dense_with_peft_config_lora(self):
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
     @require_peft
-    @pytest.mark.xfail(
-        is_peft_available() and Version(peft.__version__).is_devrelease,
-        reason=(
-            "peft's LoRA parametrization for MoE expert parameters returns bf16 under autocast, which "
-            "`register_parametrization` rejects (see #6914)."
-        ),
-        strict=True,
-    )
     def test_train_moe_with_peft_config(self):
         model_id = "trl-internal-testing/tiny-GptOssForCausalLM"
         model = AutoModelForCausalLM.from_pretrained(model_id, dtype="float32")
@@ -809,14 +801,6 @@ def test_train_peft_model(self):
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
     @require_peft
-    @pytest.mark.xfail(
-        is_peft_available() and Version(peft.__version__).is_devrelease,
-        reason=(
-            "peft's LoRA parametrization for MoE expert parameters returns bf16 under autocast, which "
-            "`register_parametrization` rejects (see #6914)."
-        ),
-        strict=True,
-    )
     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

@@ -930,14 +930,6 @@ def test_train_peft_model(self):
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
     @require_peft
-    @pytest.mark.xfail(
-        is_peft_available() and Version(peft.__version__).is_devrelease,
-        reason=(
-            "peft's LoRA parametrization for MoE expert parameters returns bf16 under autocast, which "
-            "`register_parametrization` rejects (see #6914)."
-        ),
-        strict=True,
-    )
     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

@@ -813,14 +813,6 @@ def test_train_peft_model(self):
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
     @require_peft
-    @pytest.mark.xfail(
-        is_peft_available() and Version(peft.__version__).is_devrelease,
-        reason=(
-            "peft's LoRA parametrization for MoE expert parameters returns bf16 under autocast, which "
-            "`register_parametrization` rejects (see #6914)."
-        ),
-        strict=True,
-    )
     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

@@ -18,7 +18,6 @@
 import pytest
 import torch
 from datasets import DatasetDict, IterableDatasetDict, load_dataset
-from packaging.version import Version
 from transformers import AutoModelForSequenceClassification, AutoTokenizer
 from transformers.utils import is_peft_available
 
@@ -29,7 +28,6 @@
 
 
 if is_peft_available():
-    import peft
     from peft import LoraConfig, get_peft_model
 
 
@@ -344,14 +342,6 @@ def test_train_dense_with_peft_config(self):
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
     @require_peft
-    @pytest.mark.xfail(
-        is_peft_available() and Version(peft.__version__).is_devrelease,
-        reason=(
-            "peft's LoRA parametrization for MoE expert parameters returns bf16 under autocast, which "
-            "`register_parametrization` rejects (see #6914)."
-        ),
-        strict=True,
-    )
     def test_train_moe_with_peft_config(self):
         model_id = "trl-internal-testing/tiny-Qwen3MoeForCausalLM"
         model = AutoModelForSequenceClassification.from_pretrained(model_id, dtype="float32")

tests/test_rloo_trainer.py

@@ -559,14 +559,6 @@ def test_train_peft_model(self):
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
     @require_peft
-    @pytest.mark.xfail(
-        is_peft_available() and Version(peft.__version__).is_devrelease,
-        reason=(
-            "peft's LoRA parametrization for MoE expert parameters returns bf16 under autocast, which "
-            "`register_parametrization` rejects (see #6914)."
-        ),
-        strict=True,
-    )
     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

@@ -753,14 +753,6 @@ def test_train_with_peft_config_prompt_tuning(self, peft_type):
                 assert not torch.equal(param, new_param), f"Parameter {n} has not changed."
 
     @require_peft
-    @pytest.mark.xfail(
-        is_peft_available() and Version(peft.__version__).is_devrelease,
-        reason=(
-            "peft's LoRA parametrization for MoE expert parameters returns bf16 under autocast, which "
-            "`register_parametrization` rejects (see #6914)."
-        ),
-        strict=True,
-    )
     def test_train_moe_with_peft_config(self):
         model_id = "trl-internal-testing/tiny-GptOssForCausalLM"
         model = AutoModelForCausalLM.from_pretrained(model_id, dtype="float32")