diff --git a/convert-hf-to-gguf.py b/convert-hf-to-gguf.py index 5eee32016..5e065d39b 100755 --- a/convert-hf-to-gguf.py +++ b/convert-hf-to-gguf.py @@ -201,6 +201,7 @@ class Model(ABC): try: return cls._model_classes[arch] except KeyError: + print(f"{cls._model_classes}") raise NotImplementedError(f'Architecture {arch!r} not supported!') from None def _is_model_safetensors(self) -> bool: @@ -763,7 +764,142 @@ class BaichuanModel(Model): r = weights.shape[0] // 3 return weights[r * n_part:r * n_part + r, ...] +@Model.register("XverseForCausalLM") +class XverseModel(Model): + model_arch = gguf.MODEL_ARCH.XVERSE + def set_vocab(self): + assert (self.dir_model / "tokenizer.json").is_file() + dir_model = self.dir_model + hparams = self.hparams + tokens: list[bytearray] = [] + toktypes: list[int] = [] + + from transformers import AutoTokenizer + tokenizer = AutoTokenizer.from_pretrained(dir_model) + vocab_size = hparams.get("vocab_size", len(tokenizer.vocab)) + assert max(tokenizer.vocab.values()) < vocab_size + + reverse_vocab = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()} + added_vocab = tokenizer.get_added_vocab() + + for i in range(vocab_size): + if i not in reverse_vocab: + pad_token = f"[PAD{i}]".encode('utf-8') + tokens.append(bytearray(pad_token)) + toktypes.append(gguf.TokenType.USER_DEFINED) + elif reverse_vocab[i] in added_vocab: + tokens.append(reverse_vocab[i]) + if tokenizer.added_tokens_decoder[i].special: + toktypes.append(gguf.TokenType.CONTROL) + else: + toktypes.append(gguf.TokenType.USER_DEFINED) + else: + tokens.append(reverse_vocab[i]) + toktypes.append(gguf.TokenType.NORMAL) + + self.gguf_writer.add_tokenizer_model("xverse") + self.gguf_writer.add_token_list(tokens) + self.gguf_writer.add_token_types(toktypes) + + special_vocab = gguf.SpecialVocab(dir_model, load_merges=True) + special_vocab.add_to_gguf(self.gguf_writer) + + def set_gguf_parameters(self): + block_count = self.hparams["num_hidden_layers"] + head_count = self.hparams["num_attention_heads"] + head_count_kv = self.hparams.get("num_key_value_heads", head_count) + hf_repo = self.hparams.get("_name_or_path", "") + + ctx_length = 0 + if "max_sequence_length" in self.hparams: + ctx_length = self.hparams["max_sequence_length"] + elif "max_position_embeddings" in self.hparams: + ctx_length = self.hparams["max_position_embeddings"] + elif "model_max_length" in self.hparams: + ctx_length = self.hparams["model_max_length"] + else: + print("gguf: can not find ctx length parameter.") + sys.exit() + + 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) + self.gguf_writer.add_embedding_length(self.hparams["hidden_size"]) + self.gguf_writer.add_block_count(block_count) + self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"]) + self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"]) + self.gguf_writer.add_head_count(head_count) + self.gguf_writer.add_head_count_kv(head_count_kv) + self.gguf_writer.add_layer_norm_rms_eps(self.hparams["rms_norm_eps"]) + + if self.hparams.get("rope_scaling") is not None and "factor" in self.hparams["rope_scaling"]: + if self.hparams["rope_scaling"].get("type") == "linear": + self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.LINEAR) + self.gguf_writer.add_rope_scaling_factor(self.hparams["rope_scaling"]["factor"]) + + def write_tensors(self): + # Collect tensors from generator object + model_kv = dict(self.get_tensors()) + block_count = self.hparams["num_hidden_layers"] + head_count = self.hparams["num_attention_heads"] + tensor_map = gguf.get_tensor_name_map(self.model_arch, block_count) + head_count_kv = self.hparams.get("num_key_value_heads", head_count) + + for name, data_torch in model_kv.items(): + # we don't need these + if name.endswith(".rotary_emb.inv_freq"): + continue + + old_dtype = data_torch.dtype + + # convert any unsupported data types to float32 + if data_torch.dtype not in (torch.float16, torch.float32): + data_torch = data_torch.to(torch.float32) + + # HF models permute some of the tensors, so we need to undo that + if name.endswith(("q_proj.weight")): + data_torch = self._reverse_hf_permute(data_torch, head_count, head_count) + if name.endswith(("k_proj.weight")): + data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv) + + data = data_torch.squeeze().numpy() + + # map tensor names + new_name = tensor_map.get_name(name, try_suffixes=(".weight", ".bias")) + if new_name is None: + print(f"Can not map tensor {name!r}") + sys.exit() + + n_dims = len(data.shape) + data_dtype = data.dtype + + # if f32 desired, convert any float16 to float32 + if self.ftype == 0 and data_dtype == np.float16: + data = data.astype(np.float32) + + # TODO: Why cant we use these float16 as-is? There should be not reason to store float16 as float32 + if self.ftype == 1 and data_dtype == np.float16 and n_dims == 1: + data = data.astype(np.float32) + + # if f16 desired, convert any float32 2-dim weight tensors to float16 + if self.ftype == 1 and data_dtype == np.float32 and name.endswith(".weight") and n_dims == 2: + data = data.astype(np.float16) + + print(f"{name} -> {new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") + self.gguf_writer.add_tensor(new_name, data) + + def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor: + if n_kv_head is not None and n_head != n_kv_head: + n_head //= n_kv_head + + return ( + weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) + .swapaxes(1, 2) + .reshape(weights.shape) + ) + @Model.register("FalconForCausalLM", "RWForCausalLM") class FalconModel(Model): model_arch = gguf.MODEL_ARCH.FALCON diff --git a/gguf-py/gguf/constants.py b/gguf-py/gguf/constants.py index b23badb10..90bc58e3c 100644 --- a/gguf-py/gguf/constants.py +++ b/gguf-py/gguf/constants.py @@ -120,6 +120,7 @@ class MODEL_ARCH(IntEnum): GEMMA = auto() STARCODER2 = auto() MAMBA = auto() + XVERSE = auto() class MODEL_TENSOR(IntEnum): @@ -186,6 +187,7 @@ MODEL_ARCH_NAMES: dict[MODEL_ARCH, str] = { MODEL_ARCH.GEMMA: "gemma", MODEL_ARCH.STARCODER2: "starcoder2", MODEL_ARCH.MAMBA: "mamba", + MODEL_ARCH.XVERSE: "xverse", } TENSOR_NAMES: dict[MODEL_TENSOR, str] = { @@ -578,6 +580,22 @@ MODEL_TENSORS: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.SSM_D, MODEL_TENSOR.SSM_OUT, ], + MODEL_ARCH.XVERSE: [ + MODEL_TENSOR.TOKEN_EMBD, + MODEL_TENSOR.OUTPUT_NORM, + MODEL_TENSOR.OUTPUT, + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_NORM, + MODEL_TENSOR.ATTN_Q, + MODEL_TENSOR.ATTN_K, + MODEL_TENSOR.ATTN_V, + MODEL_TENSOR.ATTN_OUT, + MODEL_TENSOR.ATTN_ROT_EMBD, + MODEL_TENSOR.FFN_NORM, + MODEL_TENSOR.FFN_GATE, + MODEL_TENSOR.FFN_DOWN, + MODEL_TENSOR.FFN_UP, + ], # TODO } @@ -610,6 +628,10 @@ MODEL_TENSOR_SKIP: dict[MODEL_ARCH, list[MODEL_TENSOR]] = { MODEL_TENSOR.ROPE_FREQS, MODEL_TENSOR.ATTN_ROT_EMBD, ], + MODEL_ARCH.XVERSE: [ + MODEL_TENSOR.ROPE_FREQS, + MODEL_TENSOR.ATTN_ROT_EMBD, + ], } #