Brunobkr/llama.cpp_AlgMor24_github

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")16class KimiLinearModel(TextModel):17    """Kimi-Linear model with hybrid MLA+KDA architecture"""18    model_arch = gguf.MODEL_ARCH.KIMI_LINEAR19 20    _experts: list[dict[str, Tensor]] | None = None21 22    def set_vocab(self):23        try:24            self._set_vocab_gpt2()25            return26        except Exception:27            pass28 29        from transformers import AutoTokenizer30        tokenizer = AutoTokenizer.from_pretrained(self.dir_model, trust_remote_code=True)31        tokpre = self.get_vocab_base_pre(tokenizer)32 33        if tokpre == "kimi-k2":34            # Build merges list using the approach similar to HunYuanMoE35            merges = []36            vocab = {}37            mergeable_ranks = tokenizer.model._mergeable_ranks  # ty: ignore[unresolved-attribute]38            for token, rank in mergeable_ranks.items():39                vocab[QwenModel.token_bytes_to_string(token)] = rank40                if len(token) == 1:41                    continue42                merged = QwenModel.bpe(mergeable_ranks, token, max_rank=rank)43                if len(merged) == 2:44                    merges.append(' '.join(map(QwenModel.token_bytes_to_string, merged)))45            # Build token list46            vocab_size = self.hparams["vocab_size"]47            special_tokens = tokenizer.special_tokens  # ty: ignore[unresolved-attribute]48            reverse_vocab = {id_ : encoded_tok for encoded_tok, id_ in {**vocab, **special_tokens}.items()}49            tokens: list[str] = []50            toktypes: list[int] = []51 52            for i in range(vocab_size):53                if i not in reverse_vocab:54                    tokens.append(f"[PAD{i}]")55                    toktypes.append(gguf.TokenType.UNUSED)56                else:57                    token = reverse_vocab[i]58                    tokens.append(token)59                    if i in special_tokens.values():60                        toktypes.append(gguf.TokenType.CONTROL)61                    else:62                        toktypes.append(gguf.TokenType.NORMAL)63 64            self.gguf_writer.add_tokenizer_model("gpt2")65            self.gguf_writer.add_tokenizer_pre(tokpre)66            self.gguf_writer.add_token_list(tokens)67            self.gguf_writer.add_token_types(toktypes)68            self.gguf_writer.add_token_merges(merges)69 70            special_vocab = gguf.SpecialVocab(self.dir_model, load_merges=False)71            special_vocab.add_to_gguf(self.gguf_writer)72            # override eos id in config.json with tiktoken eos id73            self.gguf_writer.add_eos_token_id(tokenizer.eos_id)  # ty: ignore[unresolved-attribute]74        else:75            raise NotImplementedError(f"Deepseek pre-tokenizer {tokpre!r} is not supported yet!")76 77    def set_gguf_parameters(self):78        # note: To enable MLA KV cache, attention needs to be converted into MQA (ie: GQA with 1 group)79        self.hparams["num_key_value_heads"] = 180 81        super().set_gguf_parameters()82        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])83 84        # KDA & MLA params85        # Get ssm_d_conv from linear_attn_config.short_conv_kernel_size or ssm_d_conv86        linear_attn_config = self.hparams["linear_attn_config"]87        # n_head == 0 for KDA layers, n_head > 0 for MLA layers88        # full_attention_layers list will be used to distinguish layer type89        _num_kv_heads = list()90        _full_attn_layers = linear_attn_config["full_attn_layers"]91        for il in range(self.hparams["num_hidden_layers"]):92            if il + 1 in _full_attn_layers:93                _num_kv_heads.append(self.hparams["num_key_value_heads"])94            else:95                _num_kv_heads.append(0)96        assert len(_num_kv_heads) == self.hparams["num_hidden_layers"]97        self.gguf_writer.add_head_count_kv(_num_kv_heads)98 99        if (ssm_d_conv := linear_attn_config.get("short_conv_kernel_size")) is not None:100            self.gguf_writer.add_ssm_conv_kernel(ssm_d_conv)101        if (kda_head_dim := linear_attn_config.get("head_dim")) is not None:102            self.gguf_writer.add_kda_head_dim(kda_head_dim)103 104        # MLA params - use add_* methods that handle arch substitution105        # Support both HuggingFace naming (q_lora_rank, kv_lora_rank) and internal naming (n_lora_q, n_lora_kv)106        if (q_lora_rank := self.find_hparam(["q_lora_rank", "n_lora_q"], optional=True)) is not None:107            self.gguf_writer.add_q_lora_rank(q_lora_rank)108        # To enable MLA KV cache, MLA needs to be converted into MQA with larger heads, then decompresses to MHA109        kv_lora_rank = self.find_hparam(["kv_lora_rank", "n_lora_kv"], optional=False)110        self.gguf_writer.add_kv_lora_rank(kv_lora_rank)111 112        # MLA head dimensions113        # Support HuggingFace naming: qk_nope_head_dim, qk_rope_head_dim, v_head_dim114        qk_nope_head_dim = self.hparams.get("qk_nope_head_dim")115        # Rotation - use qk_rope_head_dim for Kimi116        qk_rope_head_dim = self.find_hparam(["qk_rope_head_dim", "n_rot"], optional=False)117        self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)118        self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)119        v_head_dim = self.hparams.get("v_head_dim")120 121        # Calculate n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim122        if (n_embd_head_k_mla := self.find_hparam(["n_embd_head_k_mla"], optional=True)) is not None:123            self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)124        elif qk_nope_head_dim is not None:125            n_embd_head_k_mla = qk_nope_head_dim + qk_rope_head_dim126            self.gguf_writer.add_key_length_mla(n_embd_head_k_mla)127 128        # n_embd_head_v_mla = v_head_dim129        if (n_embd_head_v_mla := self.hparams.get("n_embd_head_v_mla")) is not None:130            self.gguf_writer.add_value_length_mla(n_embd_head_v_mla)131        elif v_head_dim is not None:132            self.gguf_writer.add_value_length_mla(v_head_dim)133 134        # moe_intermediate_size (1024 for Kimi)135        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])136        # num_shared_experts (1 for Kimi)137        self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])138        # first_k_dense_replace (1 for Kimi - first layer uses dense MLP)139        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])140        # Routed scaling factor (expert_weights_scale = 2.446 for Kimi)141        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])142 143    def prepare_tensors(self):144        super().prepare_tensors()145        if self._experts is not None:146            experts = [k for d in self._experts for k in d.keys()]147            if len(experts) > 0:148                raise ValueError(f"Unprocessed experts: {experts}")149 150    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:151        logger.info(f"Processing {name}: shape before = {tuple(data_torch.shape)}")152 153        # Handle KDA conv1d weights154        # HuggingFace/vLLM stores as [d_inner, d_conv] (2D), memory layout: conv_step changes fastest155        # llama.cpp expects ggml ne = [d_conv, 1, d_inner, 1], memory layout: ne[0]=d_conv changes fastest156        # GGUF reverses numpy shape when writing, so numpy (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]157        # Memory layouts match: both have conv_step (d_conv) changing fastest158        if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")):159            # HF shape: [d_inner, d_conv] e.g. [4096, 4]160            # Target numpy shape: (1, d_inner, 1, d_conv) -> ggml ne = [d_conv, 1, d_inner, 1]161            if data_torch.ndim == 2:162                d_inner, d_conv = data_torch.shape163                # Reshape to (1, d_inner, 1, d_conv) - memory layout preserved (d_conv fastest)164                data_torch = data_torch.reshape(1, d_inner, 1, d_conv)165                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]")166            elif data_torch.ndim == 3:167                # Already 3D [d_inner, 1, d_conv] from unsqueeze168                d_inner, _, d_conv = data_torch.shape169                data_torch = data_torch.reshape(1, d_inner, 1, d_conv)170                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]")171 172        # Handle A_log: iHF stores as [1, 1, num_heads, 1]173        # llama.cpp expects ggml ne = [1, num_heads, 1, 1]174        # GGUF reverses numpy shape: numpy (1, 1, num_heads, 1) -> ggml ne = [1, num_heads, 1, 1]175        if name.endswith(".A_log"):176            data_torch = -torch.exp(data_torch)177        if name.endswith(".dt_bias"):178            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"179            logger.info("Changed dt_bias to dt_proj.bias")180 181        # process the experts separately182        if name.find("block_sparse_moe.experts") != -1:183            n_experts = self.find_hparam(["num_local_experts", "num_experts"])184            assert bid is not None185 186            if self._experts is None:187                self._experts = [{} for _ in range(self.block_count)]188 189            self._experts[bid][name] = data_torch190 191            if len(self._experts[bid]) >= n_experts * 3:192                # merge the experts into a single 3d tensor193                # w1: gate, w2: down, w3: up194                for wid, tname in [("w1", gguf.MODEL_TENSOR.FFN_GATE_EXP),195                                   ("w2", gguf.MODEL_TENSOR.FFN_DOWN_EXP),196                                   ("w3", gguf.MODEL_TENSOR.FFN_UP_EXP)]:197                    datas: list[Tensor] = []198                    for xid in range(n_experts):199                        ename = f"model.layers.{bid}.block_sparse_moe.experts.{xid}.{wid}.weight"200                        datas.append(self._experts[bid][ename])201                        del self._experts[bid][ename]202                    data_torch = torch.stack(datas, dim=0)203                    new_name = self.format_tensor_name(tname, bid)204                    yield from super().modify_tensors(data_torch, new_name, bid)205            return206 207        # note: MLA with the absorption optimization, needs these two split and k_b_proj transposed208        if name.endswith("kv_b_proj.weight"):209            name_kb = name.replace("kv_b_proj", "k_b_proj")210            name_vb = name.replace("kv_b_proj", "v_b_proj")211            n_head_kv = self.hparams["num_key_value_heads"]212            v_head_dim = self.find_hparam(["n_embd_head_v_mla", "v_head_dim"], optional=False)213            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]214            logger.info("Split kv_b n_head_kv %d\n" % n_head_kv)215            assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)216            kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])217            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)218            k_b = k_b.transpose(1, 2)219            yield from super().modify_tensors(k_b, name_kb, bid)220            yield from super().modify_tensors(v_b, name_vb, bid)221            return222 223        yield from super().modify_tensors(data_torch, name, bid)224