Felipe97/llama-cpp-compiled

1from __future__ import annotations2 3from typing import Iterable, TYPE_CHECKING4 5import torch6 7if TYPE_CHECKING:8    from torch import Tensor9 10from .base import ModelBase, TextModel, gguf, logger11 12 13@ModelBase.register("SmallThinkerForCausalLM")14@ModelBase.example("PowerInfer/SmallThinker-4BA0.6B-Instruct")15class SmallThinkerModel(TextModel):16    model_arch = gguf.MODEL_ARCH.SMALLTHINKER17 18    def set_gguf_parameters(self):19        super().set_gguf_parameters()20        if (n_experts := self.hparams.get("moe_num_primary_experts")) is not None:21            self.gguf_writer.add_expert_count(n_experts)22        if (n_experts_used := self.hparams.get("moe_num_active_primary_experts")) is not None:23            self.gguf_writer.add_expert_used_count(n_experts_used)24        if (moe_intermediate_size := self.hparams.get("moe_ffn_hidden_size")) is not None:25            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)26            self.gguf_writer.add_feed_forward_length(moe_intermediate_size)27            logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")28        if (self.hparams.get('moe_primary_router_apply_softmax')):29            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)30        else:31            self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SIGMOID)32 33        sliding_window_layout = self.hparams.get("sliding_window_layout")34        if sliding_window_layout:35            for i in sliding_window_layout:36                if i != 0:37                    sliding_window = self.hparams.get("sliding_window_size")38                    if sliding_window:39                        self.gguf_writer.add_sliding_window(sliding_window)40                    break41 42    _experts: list[dict[str, Tensor]] | None = None43 44    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:45        # process the experts separately46        if name.find("experts") != -1:47            n_experts = self.hparams.get("moe_num_primary_experts") or self.find_hparam(["num_local_experts", "num_experts"])48            assert bid is not None49 50            if self._experts is None:51                self._experts = [{} for _ in range(self.block_count)]52 53            self._experts[bid][name] = data_torch54 55            if len(self._experts[bid]) >= n_experts * 3:56                # merge the experts into a single 3d tensor57                for w_name in ["down", "gate", "up"]:58                    datas: list[Tensor] = []59 60                    for xid in range(n_experts):61                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{w_name}.weight"62                        datas.append(self._experts[bid][ename])63                        del self._experts[bid][ename]64 65                    data_torch = torch.stack(datas, dim=0)66 67                    merged_name = f"model.layers.{bid}.block_sparse_moe.experts.{w_name}.weight"68 69                    yield from super().modify_tensors(data_torch, merged_name, bid)70                return71            else:72                return73 74        yield from super().modify_tensors(data_torch, name, bid)75 76    def prepare_tensors(self):77        super().prepare_tensors()78 79        if self._experts is not None:80            # flatten `list[dict[str, Tensor]]` into `list[str]`81            experts = [k for d in self._experts for k in d.keys()]82            if len(experts) > 0:83                raise ValueError(f"Unprocessed experts: {experts}")84