Fix GGUF (#731)

* Update mapper.py

* Update Model Conversion Command in `save.py` to `convert_hf_to_gguf.py` (#730)

* Updated convert_hf_to_gguf.py call to align with changes in llama.cpp repository

* Update save.py

---------

Co-authored-by: Daniel Han <danielhanchen@gmail.com>

* Typo Fix (#690)

---------

Co-authored-by: M. Ali Bayram <malibayram91@gmail.com>
Co-authored-by: johnpaulbin <johnpaulbin@gmail.com>

unsloth/models/loader.py

@@ -108,7 +108,7 @@ class FastLanguageModel(FastLlamaModel):
         # Cannot be both!
         if is_model and is_peft:
             raise RuntimeError(
-                "Unsloth: You repo has a LoRA adapter and a base model.\n"\
+                "Unsloth: Your repo has a LoRA adapter and a base model.\n"\
                 "You have 2 files `config.json` and `adapter_config.json`.\n"\
                 "We must only allow one config file.\n"\
                 "Please separate the LoRA and base models to 2 repos."

unsloth/models/mapper.py

@@ -199,6 +199,17 @@ __INT_TO_FLOAT_MAPPER = \
         "unsloth/gemma-2-27b",
         "google/gemma-2-27b",
     ),
+    "unsloth/gemma-2-9b-it-bnb-4bit" : (
+        "unsloth/gemma-2-9b-it",
+        "google/gemma-2-9b-it",
+    ),
+    "unsloth/gemma-2-27b-it-bnb-4bit" : (
+        "unsloth/gemma-2-27b-it",
+        "google/gemma-2-27b-it",
+    ),
+    "unsloth/Phi-3-mini-4k-instruct-v0-bnb-4bit" : ( # Old Phi pre July
+        "unsloth/Phi-3-mini-4k-instruct-v0",
+    ),
 }
 
 INT_TO_FLOAT_MAPPER = {}

unsloth/save.py

@@ -800,8 +800,8 @@ def install_llama_cpp_old(version = -10):
     # Check if successful
     if not os.path.exists("llama.cpp/quantize") and not os.path.exists("llama.cpp/llama-quantize"):
         raise RuntimeError(
-            "Unsloth: llama.cpp GGUF seems to be too buggy to install.\n"\
-            "File a report to llama.cpp's main repo since this is not an Unsloth issue."
+            "Unsloth: The file 'llama.cpp/llama-quantize' or `llama.cpp/quantize` does not exist.\n"\
+            "But we expect this file to exist! Maybe the llama.cpp developers changed the name?"
         )
     pass
 pass
@@ -945,9 +945,28 @@ def save_to_gguf(
         quantize_location = "llama.cpp/quantize"
     elif os.path.exists("llama.cpp/llama-quantize"):
         quantize_location = "llama.cpp/llama-quantize"
+    else:
+        raise RuntimeError(
+            "Unsloth: The file 'llama.cpp/llama-quantize' or 'llama.cpp/quantize' does not exist.\n"\
+            "But we expect this file to exist! Maybe the llama.cpp developers changed the name?"
+        )
     pass
 
-    if error != 0 or quantize_location is None:
+    # See https://github.com/unslothai/unsloth/pull/730
+    # Filenames changed again!
+    convert_location = None
+    if os.path.exists("llama.cpp/convert-hf-to-gguf.py"):
+        convert_location = "llama.cpp/convert-hf-to-gguf.py"
+    elif os.path.exists("llama.cpp/convert_hf_to_gguf.py"):
+        convert_location = "llama.cpp/convert_hf_to_gguf.py"
+    else:
+        raise RuntimeError(
+            "Unsloth: The file 'llama.cpp/convert-hf-to-gguf.py' or 'llama.cpp/convert_hf_to_gguf.py' does not exist.\n"\
+            "But we expect this file to exist! Maybe the llama.cpp developers changed the name?"
+        )
+    pass
+
+    if error != 0 or quantize_location is None or convert_location is None:
         print(f"Unsloth: llama.cpp error code = {error}.")
         install_llama_cpp_old(-10)
     pass
@@ -1035,7 +1054,7 @@ def save_to_gguf(
             f"--outfile {final_location} --vocab-type {vocab_type} "\
             f"--outtype {first_conversion} --concurrency {n_cpus} --pad-vocab"
     else:
-        command = f"python llama.cpp/convert-hf-to-gguf.py {model_directory} "\
+        command = f"python {convert_location} {model_directory} "\
             f"--outfile {final_location} "\
             f"--outtype {first_conversion}"
     pass

unsloth/tokenizer_utils.py

@@ -376,7 +376,7 @@ def fix_sentencepiece_gguf(saved_location):
     """
         Fixes sentencepiece tokenizers which did not extend the vocabulary with
         user defined tokens.
-        Inspiration from https://github.com/ggerganov/llama.cpp/blob/master/convert-hf-to-gguf.py
+        Inspiration from https://github.com/ggerganov/llama.cpp/blob/master/convert_hf_to_gguf.py
     """
     from copy import deepcopy
     from transformers.utils import sentencepiece_model_pb2