char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Xuan-Son Nguyenconvert: add @ModelBase.example (#27208)fa88ae936

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
8.9 KiB193 linesraw
1from __future__ import annotations
2
3import re
4
5from typing import Callable, Iterable, TYPE_CHECKING
6
7import torch
8
9if TYPE_CHECKING:
10    from torch import Tensor
11
12from .base import ModelBase, TextModel, gguf
13
14
15@ModelBase.register("BailingMoeV3ForCausalLM")
16@ModelBase.example("inclusionAI/Ling-3.0-tiny", "inclusionAI/Ling-3.0-flash")
17class BailingMoeV3Model(TextModel):
18    model_arch = gguf.MODEL_ARCH.BAILINGMOE3
19    supports_mtp_export = True
20
21    _experts: list[dict[str, Tensor]] | None = None
22    _main_layers: int | None = None
23
24    def __init__(self, *args, **kwargs):
25        super().__init__(*args, **kwargs)
26        nextn_layers = self.hparams.get("num_nextn_predict_layers", 0) or 0
27        if self.no_mtp:
28            nextn_layers = 0
29        self.block_count = self.hparams["num_hidden_layers"] + nextn_layers
30        self.tensor_map = gguf.get_tensor_name_map(self.model_arch, self.block_count)
31
32    def index_tensors(self, remote_hf_model_id: str | None = None):
33        type(self)._main_layers = self.hparams["num_hidden_layers"]
34        return super().index_tensors(remote_hf_model_id=remote_hf_model_id)
35
36    def set_vocab(self):
37        self._set_vocab_gpt2()
38
39    def is_full_attention(self, bid: int) -> bool:
40        n_layer = self.hparams["num_hidden_layers"]
41        layer_group_size = self.hparams["layer_group_size"]
42        return bid >= n_layer or (bid + 1) % layer_group_size == 0 or bid >= n_layer // layer_group_size * layer_group_size
43
44    def set_gguf_parameters(self):
45        if not self.hparams.get("no_kda_lora", False):
46            raise ValueError("BailingMoeV3 KDA LoRA projections are not supported")
47        if not self.hparams.get("kda_safe_gate", False):
48            raise ValueError("BailingMoeV3 non-safe KDA gates are not supported")
49        if self.hparams.get("gated_attention_proj_granularity_type") != "head_wise":
50            raise ValueError("BailingMoeV3 requires head-wise attention gates")
51
52        self.hparams["num_key_value_heads"] = 1
53        super().set_gguf_parameters()
54
55        n_head_kv = [1 if self.is_full_attention(il) else 0 for il in range(self.block_count)]
56        self.gguf_writer.add_head_count_kv(n_head_kv)
57
58        self.gguf_writer.add_vocab_size(self.hparams["vocab_size"])
59        self.gguf_writer.add_ssm_conv_kernel(self.hparams["short_conv_kernel_size"])
60        self.gguf_writer.add_kda_head_dim(self.hparams["head_dim"])
61        self.gguf_writer.add_kda_safe_gate(self.hparams["kda_safe_gate"])
62        self.gguf_writer.add_kda_gate_lower_bound(self.hparams["kda_lower_bound"])
63
64        kv_lora_rank = self.hparams["kv_lora_rank"]
65        qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
66        qk_rope_head_dim = self.hparams["qk_rope_head_dim"]
67        if (q_lora_rank := self.hparams.get("q_lora_rank")) is not None:
68            self.gguf_writer.add_q_lora_rank(q_lora_rank)
69        self.gguf_writer.add_kv_lora_rank(kv_lora_rank)
70        self.gguf_writer.add_rope_dimension_count(qk_rope_head_dim)
71        self.gguf_writer.add_key_length(kv_lora_rank + qk_rope_head_dim)
72        self.gguf_writer.add_key_length_mla(qk_nope_head_dim + qk_rope_head_dim)
73        self.gguf_writer.add_value_length_mla(self.hparams["v_head_dim"])
74
75        self.gguf_writer.add_expert_feed_forward_length(self.hparams["moe_intermediate_size"])
76        self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["moe_shared_expert_intermediate_size"])
77        self.gguf_writer.add_expert_shared_count(self.hparams["num_shared_experts"])
78        self.gguf_writer.add_leading_dense_block_count(self.hparams["first_k_dense_replace"])
79        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
80        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
81
82        def clamp_limits(key: str) -> list[float] | None:
83            values = self.hparams.get(key)
84            if values is None:
85                return None
86            values = [0.0 if value is None else float(value) for value in values[:self.block_count]]
87            return values + [0.0] * (self.block_count - len(values))
88
89        if (values := clamp_limits("expert_swiglu_limit_list")) is not None:
90            self.gguf_writer.add_swiglu_clamp_exp(values)
91        if (values := clamp_limits("share_expert_swiglu_limit_list")) is not None:
92            self.gguf_writer.add_swiglu_clamp_shexp(values)
93
94        if not self.no_mtp and (nextn_layers := self.hparams.get("num_nextn_predict_layers", 0)):
95            self.gguf_writer.add_nextn_predict_layers(nextn_layers)
96
97    def prepare_metadata(self, vocab_only: bool):
98        from_dir = self.fname_out.is_dir()
99        super().prepare_metadata(vocab_only=vocab_only)
100
101        if not self.mtp_only or not from_dir:
102            return
103
104        output_type: str = self.ftype.name.partition("_")[2]
105        fname_default: str = gguf.naming_convention(
106            self.metadata.name, self.metadata.basename, self.metadata.finetune,
107            self.metadata.version, size_label=None, output_type=output_type, model_type=None)
108        self.fname_out = self.fname_out.parent / f"mtp-{fname_default}.gguf"
109
110    @classmethod
111    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
112        name, gen = item
113        if name.endswith(".expert_bias"):
114            name += ".bias"
115
116        if cls._main_layers is None:
117            return super().filter_tensors((name, gen))
118
119        m = re.match(r"model\.layers\.(\d+)\.", name)
120        is_mtp = m is not None and int(m.group(1)) >= cls._main_layers
121
122        if is_mtp and cls.no_mtp:
123            return None
124        if cls.mtp_only and not is_mtp and name not in (
125            "model.word_embeddings.weight", "model.norm.weight", "lm_head.weight",
126        ):
127            return None
128
129        return super().filter_tensors((name, gen))
130
131    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
132        if name.endswith((".q_conv1d.weight", ".k_conv1d.weight", ".v_conv1d.weight")) and data_torch.ndim in (2, 3):
133            d_inner = data_torch.shape[0]
134            d_conv = data_torch.shape[-1]
135            data_torch = data_torch.reshape(1, d_inner, 1, d_conv)
136
137        if name.endswith(".A_log"):
138            data_torch = torch.exp(data_torch).reshape(-1, 1)
139
140        if name.endswith(".dt_bias"):
141            name = name.rpartition(".dt_bias")[0] + ".dt_proj.bias"
142
143        if name.endswith(".attention.f_proj.weight"):
144            assert bid is not None
145            if self.is_full_attention(bid):
146                raise ValueError(f"unexpected f_proj on full-attention layer {bid}")
147            name = self.format_tensor_name(gguf.MODEL_TENSOR.SSM_F_A, bid)
148
149        if name.endswith(".attention.g_proj.weight"):
150            assert bid is not None
151            tensor = gguf.MODEL_TENSOR.ATTN_GATE if self.is_full_attention(bid) else gguf.MODEL_TENSOR.SSM_G_A
152            name = self.format_tensor_name(tensor, bid)
153
154        if ".mlp.experts." in name:
155            n_experts = self.hparams["num_experts"]
156            assert bid is not None
157
158            if self._experts is None:
159                self._experts = [{} for _ in range(self.block_count)]
160
161            self._experts[bid][name] = data_torch
162            if len(self._experts[bid]) >= n_experts * 3:
163                for weight_name in ("down_proj", "gate_proj", "up_proj"):
164                    tensors = []
165                    for expert_id in range(n_experts):
166                        expert_name = f"model.layers.{bid}.mlp.experts.{expert_id}.{weight_name}.weight"
167                        tensors.append(self._experts[bid].pop(expert_name))
168                    merged_name = f"model.layers.{bid}.mlp.experts.{weight_name}.weight"
169                    yield from super().modify_tensors(torch.stack(tensors, dim=0), merged_name, bid)
170            return
171
172        if name.endswith(".attention.kv_b_proj.weight"):
173            assert bid is not None
174            n_head = self.hparams["num_attention_heads"]
175            v_head_dim = self.hparams["v_head_dim"]
176            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
177            assert data_torch.shape[0] == n_head * (v_head_dim + qk_nope_head_dim)
178            kv_b = data_torch.view(n_head, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
179            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
180            name_k = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K_B, bid)
181            name_v = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V_B, bid)
182            yield from super().modify_tensors(k_b.transpose(1, 2), name_k, bid)
183            yield from super().modify_tensors(v_b, name_v, bid)
184            return
185
186        yield from super().modify_tensors(data_torch, name, bid)
187
188    def prepare_tensors(self):
189        super().prepare_tensors()
190        if self._experts is not None:
191            experts = [name for layer in self._experts for name in layer]
192            if experts:
193                raise ValueError(f"Unprocessed experts: {experts}")