Llama 3.3

README.md

@@ -469,6 +469,18 @@ Two Tesla T4s on Kaggle
 ![](https://i.ibb.co/sJ7RhGG/image-41.png)
 <br>
 
+### Citing
+
+You can cite the Unsloth repo as follows:
+```bibtex
+@software{unsloth,
+  author = {Daniel Han, Michael Han and Unsloth team},
+  title = {Unsloth},
+  url = {https://github.com/unslothai/unsloth},
+  year = {2023}
+}
+```
+
 ### Thank You to
 - [HuyNguyen-hust](https://github.com/HuyNguyen-hust) for making [RoPE Embeddings 28% faster](https://github.com/unslothai/unsloth/pull/238)
 - [RandomInternetPreson](https://github.com/RandomInternetPreson) for confirming WSL support

unsloth/models/__init__.py

@@ -12,6 +12,8 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+
+from .granite import FastGraniteModel
 from .loader  import FastLanguageModel, FastVisionModel
 from .llama   import FastLlamaModel
 from .mistral import FastMistralModel

unsloth/models/_utils.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-__version__ = "2024.12.2"
+__version__ = "2024.12.3"
 
 __all__ = [
     "prepare_model_for_kbit_training",
@@ -188,7 +188,7 @@ def patch_mistral_nemo_config(config):
 
 from transformers import __version__ as transformers_version
 from transformers import PretrainedConfig
-model_architectures = ["llama", "mistral", "gemma", "gemma2", "qwen2",]
+model_architectures = ["llama", "mistral", "gemma", "gemma2", "qwen2", "granite"]
 
 for model_name in model_architectures:
     config_filepath = f"transformers.models.{model_name}.configuration_{model_name}"

unsloth/models/gemma2.py

@@ -193,7 +193,7 @@ def Gemma2DecoderLayer_fast_forward(
             output_attentions=output_attentions,
             use_cache=use_cache,
             padding_mask=padding_mask,
-            _flag_for_generation=True,
+            _flag_for_generation=self._flag_for_generation,
         )
         hidden_states = fast_rms_layernorm_inference_gemma(self.post_attention_layernorm, hidden_states, out_weight)
         hidden_states += residual