convert-*.py: parameter_class_attribute --> size_label

convert_hf_to_gguf.py

@@ -106,20 +106,20 @@ class Model:
             self.metadata.name = dir_model.name
 
         # Generate parameter weight class (useful for leader boards) if not yet determined
-        if self.metadata.parameter_class_attribute is None:
+        if self.metadata.size_label is None:
             expert_count = self.hparams.get("num_local_experts", 0)
             sum_weight_estimate = self.calculate_total_weight_count()
 
             # Calculate weight estimate per model
             per_model_weight_estimate: int = (sum_weight_estimate / expert_count) if (expert_count > 0) else sum_weight_estimate
 
-            self.metadata.parameter_class_attribute = gguf.parameter_class_attribute(expert_count, per_model_weight_estimate)
+            self.metadata.size_label = gguf.size_label(expert_count, per_model_weight_estimate)
 
         # Extracts and converts the encoding scheme from the given file type name. e.g. 'gguf.LlamaFileType.ALL_F32' --> 'F32'
         output_type = self.ftype.name.partition("_")[2]
 
         # Generate default filename based on model specification and available metadata
-        self.fname_default = gguf.naming_convention(self.metadata.name, self.metadata.basename, self.metadata.finetune, self.metadata.version, self.metadata.parameter_class_attribute, output_type)
+        self.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 fname_out is not None:

examples/convert_legacy_llama.py

@@ -802,8 +802,8 @@ class OutputFile:
             if metadata.quantized_by is not None:
                 self.gguf.add_quantized_by(metadata.quantized_by)
 
-            if metadata.parameter_class_attribute is not None:
+            if metadata.size_label is not None:
-                self.gguf.add_parameter_class_attribute(metadata.parameter_class_attribute)
+                self.gguf.add_size_label(metadata.size_label)
 
             if metadata.license is not None:
                 self.gguf.add_license(metadata.license)
@@ -1255,7 +1255,7 @@ def default_convention_outfile(file_type: GGMLFileType, expert_count:int | None,
     basename = metadata.basename if metadata.basename is not None else None
     finetune = metadata.finetune if metadata.finetune is not None else None
     version = metadata.version if metadata.version is not None else None
-    parameter_class_attribute = metadata.parameter_class_attribute if metadata.parameter_class_attribute is not None else gguf.parameter_class_attribute(expert_count, model_params_count)
+    size_label = metadata.size_label if metadata.size_label is not None else gguf.size_label(expert_count, model_params_count)
 
     output_type = {
         GGMLFileType.AllF32:    "F32",
@@ -1263,7 +1263,7 @@ def default_convention_outfile(file_type: GGMLFileType, expert_count:int | None,
         GGMLFileType.MostlyQ8_0: "Q8_0",
     }[file_type]
 
-    return gguf.naming_convention(name, basename, finetune, version, parameter_class_attribute, output_type)
+    return gguf.naming_convention(name, basename, finetune, version, size_label, output_type)
 
 
 def default_outfile(model_paths: list[Path], file_type: GGMLFileType, expert_count:int | None, model_params_count: int, metadata: gguf.Metadata) -> Path:
@@ -1427,7 +1427,7 @@ def main(args_in: list[str] | None = None) -> None:
     model   = convert_to_output_type(model, ftype)
     outfile = args.outfile or default_outfile(model_plus.paths, ftype, params.n_experts, model_params_count, metadata=metadata)
 
-    metadata.parameter_class_attribute = gguf.parameter_class_attribute(params.n_experts, model_params_count)
+    metadata.size_label = gguf.size_label(params.n_experts, model_params_count)
 
     params.ftype = ftype
     logger.info(f"Writing {outfile}, format {ftype}")

gguf-py/gguf/constants.py

@@ -37,7 +37,7 @@ class Keys:
         DESCRIPTION                = "general.description"
         QUANTIZED_BY               = "general.quantized_by"
 
-        PARAMETER_CLASS_ATTRIBUTE  = "general.parameter_class_attribute"
+        SIZE_LABEL  = "general.size_label"
 
         # Licensing details
         LICENSE                    = "general.license"

gguf-py/gguf/gguf_writer.py

@@ -480,8 +480,8 @@ class GGUFWriter:
     def add_quantized_by(self, quantized: str) -> None:
         self.add_string(Keys.General.QUANTIZED_BY, quantized)
 
-    def add_parameter_class_attribute(self, parameter_class_attribute: str) -> None:
+    def add_size_label(self, size_label: str) -> None:
-        self.add_string(Keys.General.PARAMETER_CLASS_ATTRIBUTE, parameter_class_attribute)
+        self.add_string(Keys.General.SIZE_LABEL, size_label)
 
     def add_license(self, license: str) -> None:
         self.add_string(Keys.General.LICENSE, license)

gguf-py/gguf/metadata.py

@@ -23,7 +23,7 @@ class Metadata:
     basename: Optional[str] = None
     description: Optional[str] = None
     quantized_by: Optional[str] = None
-    parameter_class_attribute: Optional[str] = None
+    size_label: Optional[str] = None
     url: Optional[str] = None
     doi: Optional[str] = None
     uuid: Optional[str] = None
@@ -69,7 +69,7 @@ class Metadata:
         metadata.description               = metadata_override.get(Keys.General.DESCRIPTION,               metadata.description)
         metadata.quantized_by              = metadata_override.get(Keys.General.QUANTIZED_BY,              metadata.quantized_by)
 
-        metadata.parameter_class_attribute = metadata_override.get(Keys.General.PARAMETER_CLASS_ATTRIBUTE, metadata.parameter_class_attribute)
+        metadata.size_label = metadata_override.get(Keys.General.SIZE_LABEL, metadata.size_label)
         metadata.license_name              = metadata_override.get(Keys.General.LICENSE_NAME,              metadata.license_name)
         metadata.license_link              = metadata_override.get(Keys.General.LICENSE_LINK,              metadata.license_link)
 
@@ -164,7 +164,7 @@ class Metadata:
         # Regular expression to extract model name components
         # Heuristic to match against cases such as 'Mixtral-8x7B-Instruct-v0.1' or 'Codestral-22B-v0.1'
         regex_match = re.compile(r'^(?P<basename>[A-Za-z0-9\s]*(?:(?:-(?:(?:[A-Za-z\s][A-Za-z0-9\s]*)|(?:[0-9\s]*)))*))'
-                                 r'(?:-(?P<parameter_class_attribute>(?:\d+x)?\d+[A-Za-z]+)(?:-(?P<finetune>[A-Za-z0-9\s-]+))?)?'
+                                 r'(?:-(?P<size_label>(?:\d+x)?\d+[A-Za-z]+)(?:-(?P<finetune>[A-Za-z0-9\s-]+))?)?'
                                  r'(?:-(?P<version>v\d+(?:\.\d+)*))?$').match(model_full_name_component)
 
         if not regex_match:
@@ -174,9 +174,9 @@ class Metadata:
         basename = components.get("basename")
         finetune = components.get("finetune")
         version = components.get("version")
-        parameter_class_attribute = components.get("parameter_class_attribute")
+        size_label = components.get("size_label")
 
-        return model_full_name_component, org_component, basename, finetune, version, parameter_class_attribute
+        return model_full_name_component, org_component, basename, finetune, version, size_label
 
     @staticmethod
     def apply_metadata_heuristic(metadata: Metadata, model_card: Optional[dict] = None, hf_params: Optional[dict] = None, model_path: Optional[Path] = None) -> Metadata:
@@ -218,7 +218,7 @@ class Metadata:
                     metadata.base_models = []
 
                 for model_id in metadata_base_models:
-                    model_full_name_component, org_component, basename, finetune, version, parameter_class_attribute = Metadata.get_model_id_components(model_id)
+                    model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id)
                     base_model = {}
                     if model_full_name_component is not None:
                         base_model["name"] = Metadata.id_to_title(model_full_name_component)
@@ -297,7 +297,7 @@ class Metadata:
                 # Use _name_or_path only if its actually a model name and not some computer path
                 # e.g. 'meta-llama/Llama-2-7b-hf'
                 model_id = hf_name_or_path
-                model_full_name_component, org_component, basename, finetune, version, parameter_class_attribute = Metadata.get_model_id_components(model_id)
+                model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id)
                 if metadata.name is None and model_full_name_component is not None:
                     metadata.name = Metadata.id_to_title(model_full_name_component)
                 if metadata.organization is None and org_component is not None:
@@ -308,14 +308,14 @@ class Metadata:
                     metadata.finetune = finetune
                 if metadata.version is None and version is not None:
                     metadata.version = version
-                if metadata.parameter_class_attribute is None and parameter_class_attribute is not None:
+                if metadata.size_label is None and size_label is not None:
-                    metadata.parameter_class_attribute = parameter_class_attribute
+                    metadata.size_label = size_label
 
         # Directory Folder Name Fallback Heuristics
         ############################################
         if model_path is not None:
             model_id = model_path.name
-            model_full_name_component, org_component, basename, finetune, version, parameter_class_attribute = Metadata.get_model_id_components(model_id)
+            model_full_name_component, org_component, basename, finetune, version, size_label = Metadata.get_model_id_components(model_id)
             if metadata.name is None and model_full_name_component is not None:
                 metadata.name = Metadata.id_to_title(model_full_name_component)
             if metadata.organization is None and org_component is not None:
@@ -326,8 +326,8 @@ class Metadata:
                 metadata.finetune = finetune
             if metadata.version is None and version is not None:
                 metadata.version = version
-            if metadata.parameter_class_attribute is None and parameter_class_attribute is not None:
+            if metadata.size_label is None and size_label is not None:
-                metadata.parameter_class_attribute = parameter_class_attribute
+                metadata.size_label = size_label
 
         return metadata
 
@@ -352,8 +352,8 @@ class Metadata:
         if self.quantized_by is not None:
             gguf_writer.add_quantized_by(self.quantized_by)
 
-        if self.parameter_class_attribute is not None:
+        if self.size_label is not None:
-            gguf_writer.add_parameter_class_attribute(self.parameter_class_attribute)
+            gguf_writer.add_size_label(self.size_label)
 
         if self.license is not None:
             gguf_writer.add_license(self.license)

gguf-py/gguf/utility.py

@@ -34,7 +34,7 @@ def model_weight_count_rounded_notation(model_params_count: int) -> str:
     return f"{round(scaled_model_params)}{scale_suffix}"
 
 
-def parameter_class_attribute(expert_count_int:int | None, model_params_count: int) -> str:
+def size_label(expert_count_int:int | None, model_params_count: int) -> str:
     per_model_rounded_weight_estimate = model_weight_count_rounded_notation(model_params_count)
 
     if expert_count_int is not None and expert_count_int > 0:
@@ -45,7 +45,7 @@ def parameter_class_attribute(expert_count_int:int | None, model_params_count: i
     return size_class
 
 
-def naming_convention(model_name: str | None, base_name: str | None, finetune_string:str | None, version_string:str | None, parameter_class_attribute: str | None, output_type: str | None) -> str:
+def naming_convention(model_name: str | None, base_name: str | None, finetune_string:str | None, version_string:str | None, size_label: str | None, output_type: str | None) -> str:
     # Reference: https://github.com/ggerganov/ggml/blob/master/docs/gguf.md#gguf-naming-convention
 
     if base_name is not None:
@@ -55,7 +55,7 @@ def naming_convention(model_name: str | None, base_name: str | None, finetune_st
     else:
         name = "ggml-model"
 
-    parameters = f"-{parameter_class_attribute}" if parameter_class_attribute is not None else ""
+    parameters = f"-{size_label}" if size_label is not None else ""
 
     finetune = f"-{finetune_string.strip().title().replace(' ', '-')}" if finetune_string is not None else ""
 

gguf-py/tests/test_metadata.py

@@ -42,17 +42,17 @@ class TestMetadataMethod(unittest.TestCase):
             'base_model': ["EmbeddedLLM/Mistral-7B-Merge-14-v0", "janai-hq/trinity-v1"]
         }
         got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), model_card, None, None)
-        expect = gguf.Metadata(name=None, author=None, version=None, organization=None, finetune=None, basename=None, description=None, quantized_by=None, parameter_class_attribute=None, url=None, doi=None, uuid=None, repo_url=None, license=None, license_name=None, license_link=None, base_models=[{'name': 'Mistral 7B Merge 14 v0', 'organization': 'EmbeddedLLM', 'repo_url': 'https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0'}, {'name': 'Trinity v1', 'organization': 'Janai Hq', 'repo_url': 'https://huggingface.co/janai-hq/trinity-v1'}], tags=['Llama-3', 'instruct', 'finetune', 'chatml', 'DPO', 'RLHF', 'gpt4', 'synthetic data', 'distillation', 'function calling', 'json mode', 'axolotl'], languages=['en'], datasets=['teknium/OpenHermes-2.5'])
+        expect = gguf.Metadata(name=None, author=None, version=None, organization=None, finetune=None, basename=None, description=None, quantized_by=None, size_label=None, url=None, doi=None, uuid=None, repo_url=None, license=None, license_name=None, license_link=None, base_models=[{'name': 'Mistral 7B Merge 14 v0', 'organization': 'EmbeddedLLM', 'repo_url': 'https://huggingface.co/EmbeddedLLM/Mistral-7B-Merge-14-v0'}, {'name': 'Trinity v1', 'organization': 'Janai Hq', 'repo_url': 'https://huggingface.co/janai-hq/trinity-v1'}], tags=['Llama-3', 'instruct', 'finetune', 'chatml', 'DPO', 'RLHF', 'gpt4', 'synthetic data', 'distillation', 'function calling', 'json mode', 'axolotl'], languages=['en'], datasets=['teknium/OpenHermes-2.5'])
         self.assertEqual(got, expect)
 
     def test_apply_metadata_heuristic_from_hf_parameters(self):
         hf_params = {"_name_or_path": "./hermes-2-pro-llama-3-8b-DPO"}
         got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), None, hf_params, None)
-        expect = gguf.Metadata(name='Hermes 2 Pro Llama 3 8b DPO', author=None, version=None, organization=None, finetune='DPO', basename='hermes-2-pro-llama-3', description=None, quantized_by=None, parameter_class_attribute='8b', url=None, doi=None, uuid=None, repo_url=None, license=None, license_name=None, license_link=None, base_models=None, tags=None, languages=None, datasets=None)
+        expect = gguf.Metadata(name='Hermes 2 Pro Llama 3 8b DPO', author=None, version=None, organization=None, finetune='DPO', basename='hermes-2-pro-llama-3', description=None, quantized_by=None, size_label='8b', url=None, doi=None, uuid=None, repo_url=None, license=None, license_name=None, license_link=None, base_models=None, tags=None, languages=None, datasets=None)
         self.assertEqual(got, expect)
 
     def test_apply_metadata_heuristic_from_model_dir(self):
         model_dir_path = Path("./hermes-2-pro-llama-3-8b-DPO")
         got = gguf.Metadata.apply_metadata_heuristic(gguf.Metadata(), None, None, model_dir_path)
-        expect = gguf.Metadata(name='Hermes 2 Pro Llama 3 8b DPO', author=None, version=None, organization=None, finetune='DPO', basename='hermes-2-pro-llama-3', description=None, quantized_by=None, parameter_class_attribute='8b', url=None, doi=None, uuid=None, repo_url=None, license=None, license_name=None, license_link=None, base_models=None, tags=None, languages=None, datasets=None)
+        expect = gguf.Metadata(name='Hermes 2 Pro Llama 3 8b DPO', author=None, version=None, organization=None, finetune='DPO', basename='hermes-2-pro-llama-3', description=None, quantized_by=None, size_label='8b', url=None, doi=None, uuid=None, repo_url=None, license=None, license_name=None, license_link=None, base_models=None, tags=None, languages=None, datasets=None)
         self.assertEqual(got, expect)