convert : support latest mistral-common (fix conversion with --mistral-format) (...

* fix convert_hf_to_gguf.py failing with --mistral-format using later mistral-common versions.

* use get_one_valid_tokenizer_file from mistral-common if available and fallback to old logic otherwise.

* use file name instead of file path for get_one_valid_tokenizer_file.

* fix --mistral-format tokenizer file failing for tokenizers in subdirectories.

* move get_one_valid_tokenizer_file import to avoid nested try-except.

diff --git a/gguf-py/gguf/vocab.py b/gguf-py/gguf/vocab.py
index 5c6817109ba571bdcb4ebc859f6c107307ddfa7a..028e5748e4d94fa07fd1da49ec1562ef36c19a14 100644 (file)
--- a/gguf-py/gguf/vocab.py
+++ b/gguf-py/gguf/vocab.py
@@ -31,6 +31,14 @@ except ImportError:
 else:
     _mistral_common_installed = True
 
+try:
+    from mistral_common.tokens.tokenizers.utils import ( # pyright: ignore[reportMissingImports]
+        get_one_valid_tokenizer_file,
+    )
+except ImportError:
+    # We still want the conversion to work with older mistral-common versions.
+    get_one_valid_tokenizer_file = None
+
 
 import gguf
 
@@ -673,24 +681,30 @@ class MistralVocab(Vocab):
 
         # Find the tokenizer files
         all_files = [f.as_posix() for f in base_path.glob("**/*") if f.is_file()]
-        valid_tokenizer_files = _filter_valid_tokenizer_files(all_files)
-
-        if len(valid_tokenizer_files) == 0:
-            raise ValueError(f"No tokenizer file found in the directory: {base_path}")
-        # If there are multiple tokenizer files, we use tekken.json if it exists, otherwise the versioned one.
-        if len(valid_tokenizer_files) > 1:
-            if "tekken.json" in valid_tokenizer_files:
-                tokenizer_file = "tekken.json"
-            else:
-                tokenizer_file = sorted(valid_tokenizer_files)[-1]
-            logger.warning(
-                f"Multiple tokenizer files found in {base_path}. Using {tokenizer_file}"
-            )
+
+        if get_one_valid_tokenizer_file is not None:
+            tokenizer_file_path = get_one_valid_tokenizer_file(all_files)
         else:
-            tokenizer_file = valid_tokenizer_files[0]
+            valid_tokenizer_files = _filter_valid_tokenizer_files(all_files)
+
+            if len(valid_tokenizer_files) == 0:
+                raise ValueError(f"No tokenizer file found in the directory: {base_path}")
+            # If there are multiple tokenizer files, we use tekken.json if it exists, otherwise the versioned one.
+            if len(valid_tokenizer_files) > 1:
+                if "tekken.json" in valid_tokenizer_files:
+                    tokenizer_file = "tekken.json"
+                else:
+                    tokenizer_file = sorted(valid_tokenizer_files)[-1]
+                logger.warning(
+                    f"Multiple tokenizer files found in {base_path}. Using {tokenizer_file}"
+                )
+            else:
+                tokenizer_file = valid_tokenizer_files[0]
+
+            tokenizer_file_path = base_path / tokenizer_file
 
         self.tokenizer = MistralTokenizer.from_file(
-            base_path / tokenizer_file
+            tokenizer_file_path
         ).instruct_tokenizer.tokenizer
         self.tokenizer_type = (
             MistralTokenizerType.tekken
@@ -698,7 +712,7 @@ class MistralVocab(Vocab):
             else MistralTokenizerType.spm
         )
         self.vocab_size = self.tokenizer.n_words
-        self.fname_tokenizer = base_path / tokenizer_file
+        self.fname_tokenizer = tokenizer_file_path
         self._name = (
             "mistral-" + self.tokenizer_type.value + "-" + self.tokenizer.version
         )