Brunobkr/llama.cpp_AlgMor24_github

1from __future__ import annotations2 3from typing import Iterable, TYPE_CHECKING4 5if TYPE_CHECKING:6    from torch import Tensor7 8from .base import ModelBase, TextModel, gguf9 10 11@ModelBase.register("GPTRefactForCausalLM")12class RefactModel(TextModel):13    model_arch = gguf.MODEL_ARCH.REFACT14 15    def set_vocab(self):16        super().set_vocab()17 18        # TODO: how to determine special FIM tokens automatically?19        special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False,20                                          special_token_types = ['prefix', 'suffix', 'middle', 'eot'])21        special_vocab._set_special_token("prefix", 1)22        special_vocab._set_special_token("suffix", 3)23        special_vocab._set_special_token("middle", 2)24        special_vocab.chat_template = None  # do not add it twice25        special_vocab.add_to_gguf(self.gguf_writer)26 27    def set_gguf_parameters(self):28        hidden_dim = self.hparams["n_embd"]29        inner_dim = 4 * hidden_dim30        hidden_dim = int(2 * inner_dim / 3)31        multiple_of = 25632        ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)33 34        # refact uses Alibi. So this is from config.json which might be used by training.35        self.gguf_writer.add_context_length(self.hparams["n_positions"])36        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])37 38        self.gguf_writer.add_feed_forward_length(ff_dim)39        self.gguf_writer.add_block_count(self.block_count)40        self.gguf_writer.add_head_count(self.hparams["n_head"])41        self.gguf_writer.add_head_count_kv(1)42        self.gguf_writer.add_layer_norm_rms_eps(self.hparams["layer_norm_epsilon"])43        self.gguf_writer.add_file_type(self.ftype)44 45    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:46        hidden_dim = self.hparams["n_embd"]47        inner_dim = 4 * hidden_dim48        hidden_dim = int(2 * inner_dim / 3)49        multiple_of = 25650        ff_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)51        n_head = self.hparams["n_head"]52        n_head_kv = 153        head_dim = self.hparams["n_embd"] // n_head54 55        if bid is not None:56            if name == f"transformer.h.{bid}.attn.kv.weight":57                yield from super().modify_tensors(data_torch[:n_head_kv * head_dim], self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K, bid), bid)58                yield from super().modify_tensors(data_torch[n_head_kv * head_dim:], self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V, bid), bid)59                return60            if name == f"transformer.h.{bid}.attn.q.weight":61                yield from super().modify_tensors(data_torch, self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_Q, bid), bid)62                return63            if name == f"transformer.h.{bid}.mlp.gate_up_proj.weight":64                yield from super().modify_tensors(data_torch[:ff_dim], self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE, bid), bid)65                yield from super().modify_tensors(data_torch[ff_dim:], self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP, bid), bid)66                return67 68        yield from super().modify_tensors(data_torch, name, bid)69