Fix DPO stats (#906)

* Update pyproject.toml

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* Update _utils.py

* Update _utils.py

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* fix_tokenizer

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* Update save.py

* Update save.py

* Update save.py

* Update save.py

* Update save.py

* Update loader.py

* Update pyproject.toml

* Update _utils.py

* Update gemma2.py

* Update gemma2.py

* Update _utils.py

* gemma 2 mask

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update _utils.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update _utils.py

* Update llama.py

* Update llama.py

* Update llama.py

* Update _utils.py

* Update _utils.py

* Update _utils.py

* Update _utils.py

* Update _utils.py

* Update _utils.py

* Torch 2.4 Xformers 0.0.27post2

* Update llama.py

* Update llama.py

* Update llama.py

* Update llama.py

* Gemma 2 fixes

* Update gemma2.py

* Update llama.py

* Update llama.py

* Update save.py

* Update save.py

* Update llama.py

* Update cross_entropy_loss.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

* Update dpo.py

unsloth/kernels/cross_entropy_loss.py

@@ -303,6 +303,7 @@ class Fast_CrossEntropyLoss(torch.autograd.Function):
 pass
 
 
+@torch._disable_dynamo
 def fast_cross_entropy_loss(logits, labels, logit_softcapping = 0):
     """
     Arguments:

unsloth/models/dpo.py

@@ -12,6 +12,10 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+__all__ = [
+    "PatchDPOTrainer",
+]
+
 try:
     from transformers.utils.notebook import (
         IntervalStrategy,
@@ -22,6 +26,12 @@ try:
 except:
     HAS_NOTEBOOK = False
 pass
+import torch
+from ._utils import torch_compile_options
+import inspect
+import torch.nn as nn
+from typing import Any, Callable, Dict, List, Literal, Optional, Tuple, Union
+
 
 DPOTrainer_metrics = [
     "rewards/chosen",
@@ -37,11 +47,11 @@ set_DPOTrainer_metrics = frozenset(DPOTrainer_metrics)
 
 
 def NotebookProgressCallback_on_train_begin(self, args, state, control, **kwargs):
-    self.first_column = "Epoch" if args.evaluation_strategy == IntervalStrategy.EPOCH else "Step"
+    self.first_column = "Epoch" if args.eval_strategy == IntervalStrategy.EPOCH else "Step"
     self.training_loss = 0
     self.last_log = 0
     column_names = [self.first_column] + ["Training Loss"]
-    if args.evaluation_strategy != IntervalStrategy.NO:
+    if args.eval_strategy != IntervalStrategy.NO:
         column_names.append("Validation Loss")
     column_names += [x.replace("/", " / ") for x in DPOTrainer_metrics]
     self.training_tracker = NotebookTrainingTracker(state.max_steps, column_names)
@@ -50,7 +60,7 @@ pass
 
 def NotebookProgressCallback_on_log(self, args, state, control, logs=None, **kwargs):
     # Only for when there is no evaluation
-    if args.evaluation_strategy == IntervalStrategy.NO and "loss" in logs:
+    if args.eval_strategy == IntervalStrategy.NO and "loss" in logs:
         values = {"Training Loss": logs["loss"]}
         for metric in DPOTrainer_metrics:
             values[metric.replace("/", " / ")] = logs[metric]

unsloth/models/llama.py

@@ -961,6 +961,7 @@ def CausalLM_fast_forward(fast_forward_inference):
 pass
 
 
+@torch._disable_dynamo
 def PeftModelForCausalLM_fast_forward(
     self,
     input_ids=None,