Felipe97/llama-cpp-compiled

1from __future__ import annotations2 3import re4 5from typing import Iterable, TYPE_CHECKING6 7if TYPE_CHECKING:8    from torch import Tensor9 10from .base import ModelBase, TextModel, gguf11 12 13@ModelBase.register("XverseForCausalLM")14@ModelBase.example("xverse/XVERSE-7B")15class XverseModel(TextModel):16    model_arch = gguf.MODEL_ARCH.XVERSE17 18    def set_vocab(self):19        assert (self.dir_model / "tokenizer.json").is_file()20        dir_model = self.dir_model21        hparams = self.hparams22 23        tokens: list[bytes] = []24        toktypes: list[int] = []25 26        from transformers import AutoTokenizer27        tokenizer = AutoTokenizer.from_pretrained(dir_model)28        vocab_size = hparams.get("vocab_size", len(tokenizer.vocab))  # ty: ignore[unresolved-attribute]29        # Since we are checking the maximum index, we need to ensure it's strictly less than vocab_size,30        # because vocab_size is the count of items, and indexes start at 0.31        max_vocab_index = max(tokenizer.get_vocab().values())  # ty: ignore[unresolved-attribute]32        if max_vocab_index >= vocab_size:33            raise ValueError("Vocabulary size exceeds expected maximum size.")34 35        reverse_vocab: dict[int, str] = {id_: encoded_tok for encoded_tok, id_ in tokenizer.vocab.items()}  # ty: ignore[unresolved-attribute]36        added_vocab = tokenizer.get_added_vocab()  # ty: ignore[unresolved-attribute]37 38        for token_id in range(vocab_size):39            token_text = reverse_vocab[token_id].encode('utf-8')40            # replace "\x00" to string with length > 041            if token_text == b"\x00":42                toktype = gguf.TokenType.BYTE  # special43                token_text = f"<{token_text}>".encode('utf-8')44            elif re.fullmatch(br"<0x[0-9A-Fa-f]{2}>", token_text):45                toktype = gguf.TokenType.BYTE  # special46            elif reverse_vocab[token_id] in added_vocab:47                if tokenizer.added_tokens_decoder[token_id].special:  # ty: ignore[unresolved-attribute]48                    toktype = gguf.TokenType.CONTROL49                else:50                    toktype = gguf.TokenType.USER_DEFINED51            else:52                toktype = gguf.TokenType.NORMAL53 54            tokens.append(token_text)55            toktypes.append(toktype)56 57        self.gguf_writer.add_tokenizer_model("llama")58        self.gguf_writer.add_tokenizer_pre("default")59        self.gguf_writer.add_token_list(tokens)60        self.gguf_writer.add_token_types(toktypes)61 62        special_vocab = gguf.SpecialVocab(dir_model, n_vocab=len(tokens))63        special_vocab.add_to_gguf(self.gguf_writer)64 65    def set_gguf_parameters(self):66        super().set_gguf_parameters()67 68        self.gguf_writer.add_tensor_data_layout("Meta AI original pth")69        self.gguf_writer.add_rope_dimension_count(self.hparams["hidden_size"] // self.hparams["num_attention_heads"])70 71    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:72        head_count = self.hparams["num_attention_heads"]73        head_count_kv = self.hparams.get("num_key_value_heads", head_count)74 75        # HF models permute some of the tensors, so we need to undo that76        if name.endswith("q_proj.weight"):77            data_torch = self._reverse_hf_permute(data_torch, head_count, head_count)78        if name.endswith("k_proj.weight"):79            data_torch = self._reverse_hf_permute(data_torch, head_count, head_count_kv)80 81        yield from super().modify_tensors(data_torch, name, bid)82 83    def _reverse_hf_permute(self, weights: Tensor, n_head: int, n_kv_head: int | None = None) -> Tensor:84        if n_kv_head is not None and n_head != n_kv_head:85            n_head //= n_kv_head86 87        return (88            weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])89            .swapaxes(1, 2)90            .reshape(weights.shape)91        )92