Dependencies

pyproject.toml

@@ -42,8 +42,8 @@ huggingface = [
     "psutil",
     "wheel>=0.42.0",
     "numpy",
-    "accelerate>=0.26.1",
-    "trl>=0.7.9,!=0.9.0,!=0.9.1,!=0.9.2,!=0.9.3",
+    "accelerate>=0.34.1",
+    "trl>=0.7.9,!=0.9.0,!=0.9.1,!=0.9.2,!=0.9.3,<=0.11.1",
     "peft>=0.7.1,!=0.11.0",
     "protobuf<4.0.0",
     "huggingface_hub",
@@ -224,8 +224,8 @@ colab-new = [
     "hf_transfer",
 ]
 colab-no-deps = [
-    "accelerate>=0.26.1",
-    "trl>=0.7.9,!=0.9.0,!=0.9.1,!=0.9.2,!=0.9.3",
+    "accelerate>=0.34.1",
+    "trl>=0.7.9,!=0.9.0,!=0.9.1,!=0.9.2,!=0.9.3,<=0.11.1",
     "peft>=0.7.1",
     "xformers<0.0.27",
     "bitsandbytes>=0.43.3",

unsloth/tokenizer_utils.py

@@ -1135,7 +1135,39 @@ def check_nvidia():
 import trl.trainer.sft_trainer
 from trl.trainer.sft_trainer import *
 from transformers.trainer import *
-from trl.trainer.sft_trainer import neftune_post_forward_hook
+try:
+    from trl.trainer.sft_trainer import neftune_post_forward_hook
+except:
+    def neftune_post_forward_hook(module, input, output):
+        """
+        Implements the NEFTune forward pass for the model using forward hooks. Note this works only for
+        torch.nn.Embedding layers. This method is slightly adapted from the original source code
+        that can be found here: https://github.com/neelsjain/NEFTune
+
+        Simply add it to your model as follows:
+        ```python
+        model = ...
+        model.embed_tokens.neftune_noise_alpha = 0.1
+        model.embed_tokens.register_forward_hook(neftune_post_forward_hook)
+        ```
+
+        Args:
+            module (`torch.nn.Module`):
+                The embedding module where the hook is attached. Note that you need to set
+                `module.neftune_noise_alpha` to the desired noise alpha value.
+            input (`torch.Tensor`):
+                The input tensor to the model.
+            output (`torch.Tensor`):
+                The output tensor of the model (i.e. the embeddings).
+        """
+        if module.training:
+            dims = torch.tensor(output.size(1) * output.size(2))
+            mag_norm = module.neftune_noise_alpha / torch.sqrt(dims)
+            output = output + torch.zeros_like(output).uniform_(-mag_norm, mag_norm)
+        return output
+    pass
+pass
+
 
 def patch_sft_trainer_tokenizer():
     """