Brunobkr/llama.cpp_AlgMor24_github

1from __future__ import annotations2 3import sys4 5from typing import Iterable, TYPE_CHECKING6 7import torch8 9if TYPE_CHECKING:10    from torch import Tensor11 12from .base import ModelBase, TextModel, gguf, logger13 14 15@ModelBase.register("GrokForCausalLM", "Grok1ForCausalLM")16class GrokModel(TextModel):17    model_arch = gguf.MODEL_ARCH.GROK18 19    def set_vocab(self):20        if (self.dir_model / 'tokenizer.model').is_file():21            self._set_vocab_sentencepiece()22            return23 24        if not (self.dir_model / 'tokenizer.json').is_file() or not (self.dir_model / 'chat_template.jinja').is_file():25            logger.error('Error: Missing vocab and chat template, download files from https://huggingface.co/alvarobartt/grok-2-tokenizer')26            sys.exit(1)27 28        self._set_vocab_gpt2()29 30    def __init__(self, *args, **kwargs):31        super().__init__(*args, **kwargs)32 33    def set_gguf_parameters(self):34        super().set_gguf_parameters()35 36        self.gguf_writer.add_attn_logit_softcapping(self.hparams.get("attn_logit_softcapping", 30.0))37        self.gguf_writer.add_router_logit_softcapping(self.hparams.get("router_logit_softcapping", 30.0))38        if (final_logit_softcap := self.hparams.get("final_logit_softcapping")):39            self.gguf_writer.add_final_logit_softcapping(final_logit_softcap)40 41        if (rope_dim := self.hparams.get("head_dim")) is None:42            rope_dim = self.hparams["hidden_size"] // self.hparams["num_attention_heads"]43 44        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:45            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)46 47        # Treat "original" as "yarn", seems to have been a mistake48        if self.hparams.get("rope_type") in ("yarn", "original"):49            self.gguf_writer.add_rope_scaling_type(gguf.RopeScalingType.YARN)50            self.gguf_writer.add_rope_scaling_factor(self.hparams["scaling_factor"])51            self.gguf_writer.add_rope_scaling_orig_ctx_len(self.hparams["original_max_position_embeddings"])52            self.gguf_writer.add_rope_scaling_yarn_ext_factor(self.hparams["extrapolation_factor"])53            self.gguf_writer.add_rope_scaling_yarn_attn_factor(self.hparams["attn_factor"])54            self.gguf_writer.add_rope_scaling_yarn_beta_fast(self.hparams["beta_fast"])55            self.gguf_writer.add_rope_scaling_yarn_beta_slow(self.hparams["beta_slow"])56 57        if temp_len := self.hparams.get("attn_temperature_len"):58            self.gguf_writer.add_attn_temperature_length(temp_len)59 60        self.gguf_writer.add_attn_output_scale(self.hparams.get("attn_output_multiplier", rope_dim**-0.5))61        self.gguf_writer.add_embedding_scale(self.hparams["embedding_multiplier_scale"])62        self.gguf_writer.add_logit_scale(self.hparams["output_multiplier_scale"])63 64    _experts: list[dict[str, list[Tensor]]] | None = None65    _cur_expert = ""66 67    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:68        deferred: list[tuple[Tensor, str, int | None]] = []69        is_expert = ".moe." in name or ".block_sparse_moe.experts." in name70 71        if not is_expert:72            deferred.append((data_torch, name, bid))73 74        # process the experts separately75        if is_expert or self._cur_expert:76            n_experts = self.hparams["num_local_experts"]77 78            assert bid is not None79 80            if self._experts is None:81                self._experts = [{} for _ in range(self.block_count)]82 83            # concatenate split tensors84            if name in self._experts[bid]:85                self._cur_expert = name86                self._experts[bid][name].append(data_torch)87                return88            elif is_expert:89                self._cur_expert = name90                self._experts[bid][name] = [data_torch]91                return92            else:93                self._cur_expert = ""94 95            for bid in range(self.block_count):96                if len(self._experts[bid]) >= n_experts * 3:97                    # merge the experts into a single 3d tensor98                    for wid in [("linear", "w1", 0), ("linear_1", "w2", 1), ("linear_v", "w3", 0)]:99                        datas: list[Tensor] = []100 101                        for xid in range(n_experts):102                            ename = f"transformer.decoder_layer.{bid}.moe.{xid}.{wid[0]}.weight"103                            if ename not in self._experts[bid]:104                                ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid[1]}.weight"105                            tensor_list = self._experts[bid][ename]106                            datas.append(torch.cat(tensor_list, dim=wid[2]) if len(tensor_list) > 1 else tensor_list[0])107                            del self._experts[bid][ename]108 109                        data_torch = torch.stack(datas, dim=0)110 111                        merged_name = f"transformer.decoder_layer.{bid}.moe.{wid[0]}.weight"112 113                        yield from super().modify_tensors(data_torch, merged_name, bid)114 115        for t in deferred:116            yield from super().modify_tensors(*t)117