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 12from .qwen import QwenModel13 14 15@ModelBase.register("KimiLinearModel", "KimiLinearForCausalLM")16@ModelBase.example("moonshotai/Kimi-Linear-48B-A3B-Instruct")17class KimiLinearModel(TextModel):18    """Kimi-Linear model with hybrid MLA+KDA architecture"""19    model_arch = gguf.MODEL_ARCH.KIMI_LINEAR20 21    _experts: list[dict[str, Tensor]] | None = None22 23    def set_vocab(self):24        try:25            self._set_vocab_gpt2()26            return27        except Exception:28            pass29 30        from transformers import AutoTokenizer31        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)32        tokpre = self.get_vocab_base_pre(tokenizer)33 34        if tokpre == "kimi-k2":35            # Build merges list using the approach similar to HunYuanMoE36            merges = []37            vocab = {}38            mergeable_ranks = tokenizer.model._mergeable_ranks  # ty: ignore[unresolved-attribute]39            for token, rank in mergeable_ranks.items():40                vocab[QwenModel.token_bytes_to_string(token)] = rank41                if len(token) == 1:42                    continue43                merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)44                if len(merged) == 2:45                    merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))46            # Build token list47            vocab_size = self.hparams["vocab_size"]48            special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]49            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}50            tokens: list[str] = []51            toktypes: list[int] = []52 53            for i in range(vocab_size):54                if i not in reverse_vocab:55                    tokens.append(f"[PAD{i}]")56                    toktypes.append(gguf.TokenType.UNUSED)57                else:58                    token = reverse_vocab[i]59                    tokens.append(token)60                    if i in special_tokens.values():61                        toktypes.append(gguf.TokenType.CONTROL)62                    else:63                        toktypes.append(gguf.TokenType.NORMAL)64 65            self.gguf_writer.add_tokenizer_model("gpt2")66            self.gguf_writer.add_tokenizer_pre(tokpre)67            self.gguf_writer.add_token_list(tokens)68            self.gguf_writer.add_token_types(toktypes)69            self.gguf_writer.add_token_merges(merges)70 71            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)72            special_vocab.add_to_gguf(self.gguf_writer)73            # override eos id in config.json with tiktoken eos id74            self.gguf_writer.add_eos_token_id(tokenizer.eos_id)  # ty: ignore[unresolved-attribute]75        else:76            raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")77 78    def set_gguf_parameters(self):79        # note: To enable MLA KV cache, attention needs to be converted into MQA (ie: GQA with 1 group)80        self.hparams["num_key_value_heads"] = 181 82        super().set_gguf_parameters()83        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])84 85        # KDA & MLA params86        # Get ssm_d_conv from linear_attn_config.short_conv_kernel_size or ssm_d_conv87        linear_attn_config = self.hparams["linear_attn_config"]88        # n_head == 0 for KDA layers, n_head > 0 for MLA layers89        # full_attention_layers list will be used to distinguish layer type90        _num_kv_heads = list()91        _full_attn_layers = linear_attn_config["full_attn_layers"]92        for il in range(self.hparams["num_hidden_layers"]):93            if il + 1 in _full_attn_layers:94                _num_kv_heads.append(self.hparams["num_key_value_heads"])95            else:96                _num_kv_heads.append(0)97        assert len(_num_kv_heads) == self.hparams["num_hidden_layers"]98        self.gguf_writer.add_head_count_kv(_num_kv_heads)99 100        if (ssm_d_conv := linear_attn_config.get("short_conv_kernel_size")) is not None:101            self.gguf_writer.add_ssm_conv_kernel(ssm_d_conv)102        if (kda_head_dim := linear_attn_config.get("head_dim")) is not None:103            self.gguf_writer.add_kda_head_dim(kda_head_dim)104 105        # MLA params - use add_* methods that handle arch substitution106        # Support both HuggingFace naming (q_lora_rank, kv_lora_rank) and internal naming (n_lora_q, n_lora_kv)107        if (q_lora_rank := self.find_hparam(["q_lora_rank", "n_lora_q"], optional=True)) is not None:108            self.gguf_writer.add_q_lora_rank(q_lora_rank)109        # To enable MLA KV cache, MLA needs to be converted into MQA with larger heads, then decompresses to MHA110        kv_lora_rank = self.find_hparam(["kv_lora_rank", "n_lora_kv"], optional=False)111        self.gguf_writer.add_kv_lora_rank(kv_lora_rank)112 113        # MLA head dimensions114        # Support HuggingFace naming: qk_nope_head_dim, qk_rope_head_dim, v_head_dim115        qk_nope_head_dim = self.hparams.get("qk_nope_head_dim")116        # Rotation - use qk_rope_head_dim for Kimi117        qk_rope_head_dim = self.find_hparam(["qk_rope_head_dim", "n_rot"], optional=False)118        self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)119        self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)120        v_head_dim = self.hparams.get("v_head_dim")121 122        # Calculate n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim123        if (n_embd_head_k_mla := self.find_hparam(["n_embd_head_k_mla"], optional=True)) is not None:124            self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)125        elif qk_nope_head_dim is not None:126            n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim127            self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)128 129        # n_embd_head_v_mla = v_head_dim130        if (n_embd_head_v_mla := self.hparams.get("n_embd_head_v_mla")) is not None:131            self.gguf_writer.add_value_length_mla(n_embd_head_v_mla)132        elif v_head_dim is not None:133            self.gguf_writer.add_value_length_mla(v_head_dim)134 135        # moe_intermediate_size (1024 for Kimi)136        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])137        # num_shared_experts (1 for Kimi)138        self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])139        # first_k_dense_replace (1 for Kimi - first layer uses dense MLP)140        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])141        # Routed scaling factor (expert_weights_scale = 2.446 for Kimi)142        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])143 144    def prepare_tensors(self):145        super().prepare_tensors()146        if self._experts is not None:147            experts = [k for d in self._experts for k in d.keys()]148            if len(experts) > 0:149                raise ValueError(f"Unprocessed experts: {experts}")150 151    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:152        logger.info(f"Processing {name}: shape before = {tuple(data_torch.shape)}")153 154        # Handle KDA conv1d weights155        # HuggingFace/vLLM stores as [d_inner, d_conv] (2D), memory layout: conv_step changes fastest156        # llama.cpp expects ggml ne = [d_conv, 1, d_inner, 1], memory layout: ne[0]=d_conv changes fastest157        # GGUF reverses numpy shape when writing, so numpy (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]158        # Memory layouts match: both have conv_step (d_conv) changing fastest159        if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")):160            # HF shape: [d_inner, d_conv] e.g. [4096, 4]161            # Target numpy shape: (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]162            if data_torch.ndim == 2:163                d_inner, d_conv = data_torch.shape164                # Reshape to (1, d_inner, 1, d_conv) - memory layout preserved (d_conv fastest)165                data_torch = data_torch.reshape(1, d_inner, 1, d_conv)166                logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")167            elif data_torch.ndim == 3:168                # Already 3D [d_inner, 1, d_conv] from unsqueeze169                d_inner, _, d_conv = data_torch.shape170                data_torch = data_torch.reshape(1, d_inner, 1, d_conv)171                logger.info(f"Reshaped conv1d weight {name}: [d_inner={d_inner}, 1, d_conv={d_conv}] -> numpy {tuple(data_torch.shape)} -> ggml ne=[{d_conv}, 1, {d_inner}, 1]")172 173        # Handle A_log: iHF stores as [1, 1, num_heads, 1]174        # llama.cpp expects ggml ne = [1, num_heads, 1, 1]175        # GGUF reverses numpy shape: numpy (1, 1, num_heads, 1) -> ggml ne = [1, num_heads, 1, 1]176        if name.endswith(".A_log"):177            data_torch = -torch.exp(data_torch)178        if name.endswith(".dt_bias"):179            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"180            logger.info("Changed dt_bias to dt_proj.bias")181 182        # process the experts separately183        if name.find("block_sparse_moe.experts") != -1:184            n_experts = self.find_hparam(["num_local_experts", "num_experts"])185            assert bid is not None186 187            if self._experts is None:188                self._experts = [{} for _ in range(self.block_count)]189 190            self._experts[bid][name] = data_torch191 192            if len(self._experts[bid]) >= n_experts * 3:193                # merge the experts into a single 3d tensor194                # w1: gate, w2: down, w3: up195                for wid, tname in [("w1", gguf.MODEL_TENSOR.FFN_GATE_EXP),196                                   ("w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP),197                                   ("w3", gguf.MODEL_TENSOR.FFN_UP_EXP)]:198                    datas: list[Tensor] = []199                    for xid in range(n_experts):200                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"201                        datas.append(self._experts[bid][ename])202                        del self._experts[bid][ename]203                    data_torch = torch.stack(datas, dim=0)204                    new_name = self.format_tensor_name(tname, bid)205                    yield from super().modify_tensors(data_torch, new_name, bid)206            return207 208        # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed209        if name.endswith("kv_b_proj.weight"):210            name_kb = name.replace("kv_b_proj", "k_b_proj")211            name_vb = name.replace("kv_b_proj", "v_b_proj")212            n_head_kv = self.hparams["num_key_value_heads"]213            v_head_dim = self.find_hparam(["n_embd_head_v_mla", "v_head_dim"], optional=False)214            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]215            logger.info("Split kv_b n_head_kv %d\n" % n_head_kv)216            assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)217            kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])218            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)219            k_b = k_b.transpose(1, 2)220            yield from super().modify_tensors(k_b, name_kb, bid)221            yield from super().modify_tensors(v_b, name_vb, bid)222            return223 224        yield from super().modify_tensors(data_torch, name, bid)225