covert-hf-to-gguf.py: refactor to use get_model_name()

Author: mofosyne

convert-hf-to-gguf.py

@@ -84,6 +84,7 @@ class SentencePieceTokenTypes(IntEnum):
 class Model:
     _model_classes: dict[str, type[Model]] = {}
 
+    model_name: str
     dir_model: Path
     ftype: int
     is_big_endian: bool
@@ -132,20 +133,28 @@ def __init__(self, dir_model: Path, ftype: gguf.LlamaFileType, fname_out: Path,
                 logger.info(f"choosing --outtype bf16 from first tensor type ({first_tensor.dtype})")
                 self.ftype = gguf.LlamaFileType.MOSTLY_BF16
 
-        # Generate default filename based on model specification and available metadata
-        def get_model_name(metadata, dir_model):
+        # Set model name based on latest metadata either provided or calculated from environment
+        def get_model_name(metadata, hyperparameters, dir_model, model_arch):
             if metadata is not None and metadata.name is not None:
+                # Explicit Metadata Was Provided By User
                 return metadata.name
+            elif hyperparameters is not None and "model_type" in hyperparameters:
+                # Hugging Face Hyperparameter Model Type Already Provided
+                return hyperparameters["model_type"]
             elif dir_model is not None and dir_model.name is not None:
+                # Use directory folder name
                 return dir_model.name
-            return None
+            else:
+                return gguf.MODEL_ARCH_NAMES[model_arch]
+        self.model_name = get_model_name(self.metadata, self.hparams, self.dir_model, self.model_arch)
+
+        # Generate default filename based on model specification and available metadata
         def extract_encoding_scheme(ftype):
             # Extracts and converts the encoding scheme from the given file type name.
             # e.g. 'gguf.LlamaFileType.ALL_F32' --> 'F32'
             return ftype.name.partition("_")[2].upper()
-        model_name = get_model_name(metadata, dir_model)
         expert_count = self.hparams["num_local_experts"] if "num_local_experts" in self.hparams else None
-        self.fname_default = f"{gguf.naming_convention(model_name, self.metadata.version, expert_count, self.parameter_count(), extract_encoding_scheme(self.ftype))}"
+        self.fname_default = f"{gguf.naming_convention(self.model_name, self.metadata.version, expert_count, self.parameter_count(), extract_encoding_scheme(self.ftype))}"
 
         # Filename Output
         if fname_out is not None:
@@ -244,15 +253,7 @@ def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", "
         return new_name
 
     def set_gguf_meta_model(self):
-
-        # Metadata About The Model And Its Provenence
-        name = "LLaMA"
-        if self.metadata is not None and self.metadata.name is not None:
-            name = metadata.name
-        elif self.dir_model is not None:
-            name = self.dir_model.name
-
-        self.gguf_writer.add_name(name)
+        self.gguf_writer.add_name(self.model_name)
 
         if self.metadata is not None:
             if self.metadata.author is not None:
@@ -770,7 +771,6 @@ class GPTNeoXModel(Model):
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
 
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
@@ -826,7 +826,6 @@ class BloomModel(Model):
     model_arch = gguf.MODEL_ARCH.BLOOM
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_name("Bloom")
         n_embed = self.hparams.get("hidden_size", self.hparams.get("n_embed"))
         n_head = self.hparams.get("n_head", self.hparams.get("num_attention_heads"))
         self.gguf_writer.add_context_length(self.hparams.get("seq_length", n_embed))
@@ -903,7 +902,6 @@ def set_vocab(self):
 
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layers"]
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
         self.gguf_writer.add_embedding_length(self.hparams["d_model"])
         self.gguf_writer.add_block_count(block_count)
@@ -955,7 +953,6 @@ def set_gguf_parameters(self):
             raise ValueError("gguf: can not find ctx length parameter.")
 
         self.gguf_writer.add_file_type(self.ftype)
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_source_hf_repo(hf_repo)
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
@@ -992,7 +989,6 @@ def set_gguf_parameters(self):
         else:
             raise ValueError("gguf: can not find ctx length parameter.")
 
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_source_hf_repo(hf_repo)
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
@@ -1115,7 +1111,6 @@ def set_gguf_parameters(self):
         else:
             raise ValueError("gguf: can not find ctx length parameter.")
 
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_source_hf_repo(hf_repo)
         self.gguf_writer.add_tensor_data_layout("Meta AI original pth")
         self.gguf_writer.add_context_length(ctx_length)
@@ -1175,7 +1170,6 @@ def set_gguf_parameters(self):
         if n_head_kv is None:
             n_head_kv = self.hparams.get("n_head_kv", 1)  # old name
 
-        self.gguf_writer.add_name("Falcon")
         self.gguf_writer.add_context_length(2048)  # not in config.json
         self.gguf_writer.add_tensor_data_layout("jploski")  # qkv tensor transform
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
@@ -1220,7 +1214,6 @@ class StarCoderModel(Model):
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
-        self.gguf_writer.add_name("StarCoder")
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
@@ -1256,7 +1249,6 @@ def set_gguf_parameters(self):
 
         block_count = self.hparams["n_layer"]
 
-        self.gguf_writer.add_name("Refact")
         # refact uses Alibi. So this is from config.json which might be used by training.
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
@@ -1311,7 +1303,6 @@ def set_gguf_parameters(self):
         hparams = self.hparams
         block_count = hparams["num_hidden_layers"]
 
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
@@ -1501,7 +1492,6 @@ def __init__(self, *args, **kwargs):
 
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
-        self.gguf_writer.add_name("Grok")
 
     _experts: list[dict[str, Tensor]] | None = None
 
@@ -1550,7 +1540,6 @@ class DbrxModel(Model):
     def set_gguf_parameters(self):
         ffn_config = self.hparams["ffn_config"]
         attn_config = self.hparams["attn_config"]
-        self.gguf_writer.add_name(self.hparams["model_type"])
         self.gguf_writer.add_block_count(self.hparams["n_layers"])
 
         self.gguf_writer.add_context_length(self.hparams["max_seq_len"])
@@ -1620,7 +1609,6 @@ class MiniCPMModel(Model):
 
     def set_gguf_parameters(self):
         block_count = self.hparams["num_hidden_layers"]
-        self.gguf_writer.add_name("MiniCPM")
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
@@ -1690,7 +1678,6 @@ def set_vocab(self):
         self._set_vocab_qwen()
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_name("Qwen")
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
@@ -1775,7 +1762,6 @@ class GPT2Model(Model):
     model_arch = gguf.MODEL_ARCH.GPT2
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_block_count(self.hparams["n_layer"])
         self.gguf_writer.add_context_length(self.hparams["n_ctx"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
@@ -1818,7 +1804,6 @@ def set_gguf_parameters(self):
         n_embd = self.find_hparam(["hidden_size", "n_embd"])
         n_head = self.find_hparam(["num_attention_heads", "n_head"])
 
-        self.gguf_writer.add_name("Phi2")
         self.gguf_writer.add_context_length(self.find_hparam(["n_positions", "max_position_embeddings"]))
 
         self.gguf_writer.add_embedding_length(n_embd)
@@ -1940,7 +1925,6 @@ def set_gguf_parameters(self):
         orig_max_pos_embds = self.find_hparam(["original_max_position_embeddings"])
         rope_dims = n_embd // n_head
 
-        self.gguf_writer.add_name("Phi3")
         self.gguf_writer.add_context_length(max_pos_embds)
         self.gguf_writer.add_rope_scaling_orig_ctx_len(orig_max_pos_embds)
         self.gguf_writer.add_embedding_length(n_embd)
@@ -1997,7 +1981,6 @@ def set_gguf_parameters(self):
         hparams = self.hparams
         block_count = hparams["num_hidden_layers"]
 
-        self.gguf_writer.add_name("PLaMo")
         self.gguf_writer.add_context_length(4096)  # not in config.json
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_feed_forward_length(hparams["intermediate_size"])
@@ -2042,7 +2025,6 @@ class CodeShellModel(Model):
     def set_gguf_parameters(self):
         block_count = self.hparams["n_layer"]
 
-        self.gguf_writer.add_name("CodeShell")
         self.gguf_writer.add_context_length(self.hparams["n_positions"])
         self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
         self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
@@ -2176,7 +2158,6 @@ def _hf_permute_qk(self, weights, n_head: int, n_head_kv: int):
                 .reshape(weights.shape))
 
     def set_gguf_parameters(self):
-        self.gguf_writer.add_name("InternLM2")
         self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_block_count(self.hparams["num_hidden_layers"])
         self.gguf_writer.add_embedding_length(self.hparams["hidden_size"])
@@ -2342,7 +2323,6 @@ def set_gguf_parameters(self):
         hparams = self.hparams
         block_count = hparams["num_hidden_layers"]
 
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_context_length(hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hparams["hidden_size"])
         self.gguf_writer.add_block_count(block_count)
@@ -2442,7 +2422,6 @@ def set_gguf_parameters(self):
         # Fail early for models which don't have a block expansion factor of 2
         assert d_inner == 2 * d_model
 
-        self.gguf_writer.add_name(self.dir_model.name)
         self.gguf_writer.add_context_length(2**20) # arbitrary value; for those who use the default
         self.gguf_writer.add_embedding_length(d_model)
         self.gguf_writer.add_feed_forward_length(0) # unused, but seemingly required when loading