Fix some bugs (#83)

* Fix tokenizer, dropout, bias for LoRA

* Update loader.py

* Fix LoRA downcasting

* Update _utils.py

* Saving to GGUF

* fix

* colab_quantize_to_gguf

* move save modules

* save module

* Update __init__.py

* Update save.py

* Temp downgrade due to TRL issue

* Fix up bugs

pyproject.toml

@@ -37,7 +37,7 @@ huggingface = [
     "datasets",
     "sentencepiece",
     "accelerate",
-    "trl==0.7.7",
+    "trl",
     "peft",
     "packaging",
     "ninja",

unsloth/kernels/fast_lora.py

@@ -22,7 +22,7 @@ def get_lora_parameters(proj):
     base_layer = (proj.base_layer if hasattr(proj, "base_layer") else proj)
     W = base_layer.weight
 
-    if proj.disable_adapters or proj.merged:
+    if not hasattr(proj, "disable_adapters") or proj.disable_adapters or proj.merged:
         return W, QUANT_STATE(W), None, None, None
     pass
 

unsloth/models/_utils.py

@@ -116,6 +116,11 @@ def patch_tokenizer(model, tokenizer):
 pass
 
 
+IGNORED_TOKENIZER_CHECKING = frozenset((
+    "CodeLlamaTokenizerFast",
+    "CodeLlamaTokenizer",
+))
+
 def check_tokenizer(
     model,
     tokenizer,
@@ -131,6 +136,11 @@ def check_tokenizer(
     # See https://huggingface.co/berkeley-nest/Starling-LM-7B-alpha/discussions/25
     # Seems like the Fast tokenizer in Rust breaks things!
 
+    # We ignore some of them!
+    if tokenizer.__repr__().split("(", 1)[0] in IGNORED_TOKENIZER_CHECKING:
+        return tokenizer
+    pass
+
     max_embedding_size = model.model.embed_tokens.weight.shape[0]
     added_tokens_fast = tokenizer.added_tokens_decoder
     added_tokens_fast = {index : str(value) for index, value in added_tokens_fast.items()}

unsloth/models/llama.py

@@ -714,6 +714,16 @@ class FastLlamaModel:
                 token = token,
             )
         pass
+
+        # Fix up config for transformers uploading PEFT
+        name = model.config._name_or_path
+        if name.startswith("unsloth/") and name.endswith("-bnb-4bit"):
+            name = name[:len(name) - len("-bnb-4bit")]
+            model.config.update({"_name_or_path" : name})
+        pass
+        # Log Unsloth version for future fastpaths for inference
+        model.config.update({"unsloth_version" : __version__})
+
         return model, tokenizer
     pass
 

unsloth/models/mistral.py

@@ -343,6 +343,16 @@ class FastMistralModel(FastLlamaModel):
                 token = token,
             )
         pass
+
+        # Fix up config for transformers uploading PEFT
+        name = model.config._name_or_path
+        if name.startswith("unsloth/") and name.endswith("-bnb-4bit"):
+            name = name[:len(name) - len("-bnb-4bit")]
+            model.config.update({"_name_or_path" : name})
+        pass
+        # Log Unsloth version for future fastpaths for inference
+        model.config.update({"unsloth_version" : __version__})
+        
         return model, tokenizer
     pass
 pass