convert-*.py: Add naming_convention_vocab_only()

convert_hf_to_gguf.py

@@ -337,7 +337,7 @@ class Model:
 
                 self.gguf_writer.add_tensor(new_name, data, raw_dtype=data_qtype)
 
-    def prepare_metadata(self):
+    def prepare_metadata(self, vocab_only: bool):
 
         # Fallback to model directory name if metadata name is still missing
         if self.metadata.name is None:
@@ -350,20 +350,34 @@ class Model:
             if (total_params > 0):
                 self.metadata.size_label = gguf.size_label(total_params, shared_params, expert_params, expert_count)
 
-        # Extract the encoding scheme from the file type name. e.g. 'gguf.LlamaFileType.MOSTLY_Q8_0' --> 'Q8_0'
-        output_type = self.ftype.name.partition("_")[2]
-
-        # Generate default filename based on model specification and available metadata
-        fname_default = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, self.metadata.size_label, output_type)
-
         # Filename Output
-        if self.fname_out is not None:
+        if self.fname_out is not None and not self.fname_out.is_dir():
-            if not self.fname_out.is_dir():
+            # Output path is a custom defined templated filename
-                # custom defined filename and path was provided
+            # Note: `not is_dir()` is used because `.is_file()` will not detect
-                # allow templating the file name with the output ftype, useful with the "auto" ftype
+            #       file template strings as it doesn't actually exist as a file
+
+            # Extract the encoding scheme from the file type name. e.g. 'gguf.LlamaFileType.MOSTLY_Q8_0' --> 'Q8_0'
+            output_type: str = self.ftype.name.partition("_")[2]
+
+            # Process templated file name with the output ftype, useful with the "auto" ftype
             self.fname_out = self.fname_out.parent / gguf.fill_templated_filename(self.fname_out.name, output_type)
         else:
-                # the target file is a directory
+
+            # Generate default filename based on model specification and available metadata
+            if vocab_only:
+                # Vocab based default filename
+                fname_default: str = gguf.naming_convention_vocab_only(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version)
+            else:
+
+                # Extract the encoding scheme from the file type name. e.g. 'gguf.LlamaFileType.MOSTLY_Q8_0' --> 'Q8_0'
+                output_type: str = self.ftype.name.partition("_")[2]
+
+                # Standard default filename
+                fname_default: str = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, self.metadata.size_label, output_type)
+
+            # Check if preferred output directory path was provided
+            if self.fname_out is not None and self.fname_out.is_dir():
+                # output path is a directory
                 self.fname_out = self.fname_out / f"{fname_default}.gguf"
             else:
                 # output in the same directory as the model by default
@@ -384,7 +398,7 @@ class Model:
 
     def write(self):
         self.prepare_tensors()
-        self.prepare_metadata()
+        self.prepare_metadata(vocab_only=False)
         self.gguf_writer.write_header_to_file(path=self.fname_out)
         self.gguf_writer.write_kv_data_to_file()
         self.gguf_writer.write_tensors_to_file(progress=True)
@@ -394,7 +408,7 @@ class Model:
         if len(self.gguf_writer.tensors) != 1:
             raise ValueError('Splitting the vocabulary is not supported')
 
-        self.prepare_metadata()
+        self.prepare_metadata(vocab_only=True)
         self.gguf_writer.write_header_to_file(path=self.fname_out)
         self.gguf_writer.write_kv_data_to_file()
         self.gguf_writer.close()

gguf-py/gguf/utility.py

@@ -1,10 +1,10 @@
 from __future__ import annotations
 
 
-def fill_templated_filename(filename: str, output_type: str):
+def fill_templated_filename(filename: str, output_type: str | None) -> str:
     # Given a file name fill in any type templates e.g. 'some-model-name.{ftype}.gguf'
-    ftype_uppercase: str = output_type.upper()
+    ftype_lowercase: str = output_type.lower() if output_type is not None else ""
-    ftype_lowercase: str = output_type.lower()
+    ftype_uppercase: str = output_type.upper() if output_type is not None else ""
     return filename.format(ftype_lowercase,
                            outtype=ftype_lowercase, ftype=ftype_lowercase,
                            OUTTYPE=ftype_uppercase, FTYPE=ftype_uppercase)
@@ -63,3 +63,20 @@ def naming_convention(model_name: str | None, base_name: str | None, finetune_st
     precision = f"-{output_type.strip().replace(' ', '-').upper()}" if output_type is not None else ""
 
     return f"{name}{parameters}{finetune}{version}{precision}"
+
+
+def naming_convention_vocab_only(model_name: str | None, base_name: str | None, finetune_string: str | None, version_string: str | None) -> str:
+    # Reference: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#gguf-naming-convention
+
+    if base_name is not None:
+        name = base_name.strip().title().replace(' ', '-').replace('/', '-')
+    elif model_name is not None:
+        name = model_name.strip().title().replace(' ', '-').replace('/', '-')
+    else:
+        name = "ggml-model"
+
+    finetune = f"-{finetune_string.strip().title().replace(' ', '-')}" if finetune_string is not None else ""
+
+    version = f"-{version_string.strip().replace(' ', '-')}" if version_string is not None else ""
+
+    return f"{name}{finetune}{version}-vocab"