Many bug fixes (#1162)

* Fix TRL

* Update mistral.py

* Patch processing_class

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* Update tokenizer_utils.py

* Installation guide (#1165)

* chore: update chat_templates.py (#1166)

orginal -> original

* Disable Flex Attention

* Update tokenizer_utils.py

* Update _utils.py

---------

Co-authored-by: timothelaborie <97834767+timothelaborie@users.noreply.github.com>
Co-authored-by: Ikko Eltociear Ashimine <eltociear@gmail.com>

Author: 3 people

README.md

@@ -181,6 +181,18 @@ x = x.format(cuda.replace(".", ""), "-ampere" if is_ampere else "")
 print(f'pip install --upgrade pip && pip install "unsloth[{x}] @ git+https://github.com/unslothai/unsloth.git"')
 ```
 
+### Windows Installation
+
+To run Unsloth directly on Windows:
+- Install Triton from this Windows fork and follow the instructions: https://github.com/woct0rdho/triton-windows
+- In the SFTTrainer, set `dataset_num_proc=1` to avoid a crashing issue:
+```python
+trainer = SFTTrainer(
+    dataset_num_proc=1,
+    ...
+)
+```
+
 For **advanced installation instructions** or if you see weird errors during installations:
 
 1. Install `torch` and `triton`. Go to https://pytorch.org to install it. For example `pip install torch torchvision torchaudio triton`

pyproject.toml

@@ -44,7 +44,7 @@ huggingface = [
     "wheel>=0.42.0",
     "numpy",
     "accelerate>=0.34.1",
-    "trl>=0.7.9,!=0.9.0,!=0.9.1,!=0.9.2,!=0.9.3,<=0.11.1",
+    "trl>=0.7.9,!=0.9.0,!=0.9.1,!=0.9.2,!=0.9.3",
     "peft>=0.7.1,!=0.11.0",
     "protobuf<4.0.0",
     "huggingface_hub",
@@ -227,7 +227,7 @@ colab-new = [
 ]
 colab-no-deps = [
     "accelerate>=0.34.1",
-    "trl>=0.7.9,!=0.9.0,!=0.9.1,!=0.9.2,!=0.9.3,<=0.11.1",
+    "trl>=0.7.9,!=0.9.0,!=0.9.1,!=0.9.2,!=0.9.3",
     "peft>=0.7.1",
     "xformers<0.0.27",
     "bitsandbytes>=0.43.3",

unsloth/chat_templates.py

@@ -678,7 +678,7 @@
 {{- end }}
 {{- if .Tools }}
 
-You are a helpful assistant with tool calling capabilities. When you receive a tool call response, use the output to format an answer to the orginal use question.
+You are a helpful assistant with tool calling capabilities. When you receive a tool call response, use the output to format an answer to the original use question.
 {{- end }}
 {{- end }}<|eot_id|>
 {{- range $i, $_ := .Messages }}

unsloth/kernels/flex_attention.py

@@ -31,7 +31,7 @@
         create_block_mask as _create_block_mask,
     )
     _flex_attention = torch.compile(_flex_attention, dynamic = True, options = torch_compile_options)
-    HAS_FLEX_ATTENTION = True
+    HAS_FLEX_ATTENTION = False
 except:
     HAS_FLEX_ATTENTION = False
 pass

unsloth/models/_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2024.10.3"
+__version__ = "2024.10.4"
 
 __all__ = [
     "prepare_model_for_kbit_training",
@@ -1194,8 +1194,8 @@ def patch_gradient_accumulation_fix(Trainer):
         logger.warning_once(
             "Unsloth: We fixed a gradient accumulation bug, "\
             "but it seems like you don't have the latest transformers version!\n"\
-            "Please update transformers via:\n"\
-            '`pip uninstall transformers -y && pip install --upgrade --no-cache-dir "git+https://github.com/huggingface/transformers.git"`'
+            "Please update transformers, TRL and unsloth via:\n"\
+            '`pip install --upgrade --no-cache-dir unsloth git+https://github.com/huggingface/transformers.git git+https://github.com/huggingface/trl.git`'
         )
     pass
 pass

unsloth/models/mistral.py

@@ -254,8 +254,9 @@ def MistralForCausalLM_fast_forward(
 
         shift_labels = torch.hstack((labels[..., 1:], self.extra_ignored_labels[:labels.shape[0]]))
         loss = fast_cross_entropy_loss(
-            logits = shift_logits,
-            labels = shift_labels,
+            logits  = shift_logits,
+            labels  = shift_labels,
+            n_items = kwargs.get("num_items_in_batch", None) or kwargs.get("n_items", None),
         )
     pass
 

unsloth/tokenizer_utils.py

@@ -831,6 +831,7 @@ def check_nvidia():
 PRE_CHECK = check_nvidia()
 
 
+import inspect
 from inspect import getsource
 import trl.trainer.sft_trainer
 from trl.trainer.sft_trainer import *
@@ -869,6 +870,35 @@ def neftune_post_forward_hook(module, input, output):
 pass
 
 
+def patch_trl_tokenizer_processing_class(trainer_name):
+    # New TRL removes tokenizer!
+    # We return it back!
+    exec(f"from trl import {trainer_name}", globals())
+    if str(eval(f"{trainer_name}").__name__).startswith("Unsloth"): return None
+    parameters = eval(f"inspect.signature({trainer_name}).parameters")
+    if "tokenizer" in parameters: return None
+
+    args = {
+        key : \
+            value.default \
+            if type(value.default) is not str else \
+            f"'{value.default}'" \
+        for key, value in parameters.items()
+    }
+    args["tokenizer"] = None
+    new_args = args.copy()
+    del new_args["tokenizer"]
+    del new_args["processing_class"]
+    new_args = ",\n".join(f"{' '*12}{key} = {key}" for key in new_args) + \
+        f",\n{' '*12}processing_class = tokenizer if tokenizer else processing_class"
+    args = ",\n".join(f"{' '*8}{key} = {value}" for key, value in args.items())
+    args = f"def __init__(\n" + f"{' '*8}self,\n" + args + "):"
+    args += f"\n{' '*8}\n{' '*8}super().__init__(\n{new_args}\n{' '*8})"
+    new_class = f"""class Unsloth{trainer_name}({trainer_name}):\n{' '*4}{args}\n"""
+    return new_class
+pass
+
+
 def patch_sft_trainer_tokenizer():
     """
         Patches the trainer with changes
@@ -884,7 +914,8 @@ def patch_sft_trainer_tokenizer():
 
         check_text = \
         "\n"\
-        "test_text = dataset[0][dataset_text_field] if (formatting_func is None or not use_formatting_func) else formatting_func(dataset[0])[0]\n"\
+        "if 'tokenizer' not in locals(): tokenizer = processing_class\n"\
+        "test_text = dataset[0][dataset_text_field] if (formatting_func is not None and dataset_text_field is None) else formatting_func(dataset[0])[0]\n"\
         "chat_template = getattr(tokenizer, 'chat_template', None)\n"\
         "chat_template = '' if chat_template is None else chat_template\n"\
         "has_bos_token_already = (test_text.startswith(tokenizer.bos_token) or tokenizer.bos_token in chat_template) "\
@@ -941,7 +972,8 @@ def patch_sft_trainer_tokenizer():
         "        from transformers import __version__ as transformers_version\n"\
         "        from packaging.version import Version\n"\
         "        if Version(transformers_version) <= Version('4.45.2'):\n"\
-        "            print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers and Unsloth!')\n"\
+        "            print('**** Unsloth: Please use our fixed gradient_accumulation_steps by updating transformers, TRL and Unsloth!\\n'\\\n"\
+        "                  '`pip install --upgrade --no-cache-dir unsloth git+https://github.com/huggingface/transformers.git git+https://github.com/huggingface/trl.git`')\n"\
         "except:\n"\
         "    pass\n"\
         "\n\n"
@@ -981,4 +1013,13 @@ def patch_sft_trainer_tokenizer():
     pass
 pass
 
+# Fix TRL trainers with removed tokenizer args (got replaced with processing_class)
+for trainer_name in ("SFTTrainer", "DPOTrainer", "KTOTrainer"):
+    trainer_text = patch_trl_tokenizer_processing_class(trainer_name)
+    if trainer_text is None: continue
+    exec(trainer_text, globals())
+    exec(f"trl.trainer.{trainer_name} = Unsloth{trainer_name}", globals())
+pass
+
+# FInally patch TRL tokenizer things
 patch_sft_trainer_tokenizer()