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 →
4.8 KiB109 linesraw
1from __future__ import annotations
2
3from typing import Iterable, TYPE_CHECKING
4
5import torch
6
7if TYPE_CHECKING:
8    from torch import Tensor
9
10from .base import ModelBase, TextModel, gguf, logger
11
12
13@ModelBase.register("GroveMoeForCausalLM", "modeling_grove_moe.GroveMoeForCausalLM")
14@ModelBase.example("inclusionAI/GroveMoE-Inst")
15class GroveMoeModel(TextModel):
16    model_arch = gguf.MODEL_ARCH.GROVEMOE
17
18    def set_gguf_parameters(self):
19        super().set_gguf_parameters()
20        if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
21            self.gguf_writer.add_expert_feed_forward_length(moe_intermediate_size)
22            logger.info(f"gguf: expert feed forward length = {moe_intermediate_size}")
23        # FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L299
24        self.gguf_writer.add_expert_chunk_feed_forward_length(self.hparams.get("head_dim") or 128)
25        # FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L298
26        self.gguf_writer.add_experts_per_group(2)
27        # FIXME?: Hardcoded https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L376
28        self.gguf_writer.add_expert_group_scale(0.05)
29
30    _experts: list[dict[str, Tensor]] | None = None
31    _chunk_experts: list[dict[str, Tensor]] | None = None
32
33    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
34        if name.endswith(".expert_bias"):
35            # FIXME?: Unused https://huggingface.co/inclusionAI/GroveMoE-Inst/blob/c4c69e5970d18907b5e6ddccdfd55176fe292df1/modeling_grove_moe.py#L303
36            return
37
38        # process the experts separately
39        if name.find("chunk_experts") != -1:
40            n_experts = self.find_hparam(["num_local_experts", "num_experts"]) // 2 # see add_experts_per_group
41            assert bid is not None
42
43            if self._chunk_experts is None:
44                self._chunk_experts = [{} for _ in range(self.block_count)]
45
46            self._chunk_experts[bid][name] = data_torch
47
48            if len(self._chunk_experts[bid]) >= n_experts * 3:
49                # merge the experts into a single 3d tensor
50                for w_name in ["down_proj", "gate_proj", "up_proj"]:
51                    datas: list[Tensor] = []
52
53                    for xid in range(n_experts):
54                        ename = f"model.layers.{bid}.mlp.chunk_experts.{xid}.{w_name}.weight"
55                        datas.append(self._chunk_experts[bid][ename])
56                        del self._chunk_experts[bid][ename]
57
58                    data_torch = torch.stack(datas, dim=0)
59
60                    merged_name = f"model.layers.{bid}.mlp.chunk_experts.{w_name}.weight"
61
62                    yield from super().modify_tensors(data_torch, merged_name, bid)
63                return
64            else:
65                return
66        elif name.find("experts") != -1:
67            n_experts = self.find_hparam(["num_local_experts", "num_experts"])
68            assert bid is not None
69
70            if self._experts is None:
71                self._experts = [{} for _ in range(self.block_count)]
72
73            self._experts[bid][name] = data_torch
74
75            if len(self._experts[bid]) >= n_experts * 3:
76                # merge the experts into a single 3d tensor
77                for w_name in ["down_proj", "gate_proj", "up_proj"]:
78                    datas: list[Tensor] = []
79
80                    for xid in range(n_experts):
81                        ename = f"model.layers.{bid}.mlp.experts.{xid}.{w_name}.weight"
82                        datas.append(self._experts[bid][ename])
83                        del self._experts[bid][ename]
84
85                    data_torch = torch.stack(datas, dim=0)
86
87                    merged_name = f"model.layers.{bid}.mlp.experts.{w_name}.weight"
88
89                    yield from super().modify_tensors(data_torch, merged_name, bid)
90                return
91            else:
92                return
93
94        yield from super().modify_tensors(data_torch, name, bid)
95
96    def prepare_tensors(self):
97        super().prepare_tensors()
98
99        if self._chunk_experts is not None:
100            # flatten `list[dict[str, Tensor]]` into `list[str]`
101            chunk_experts = [k for d in self._chunk_experts for k in d.keys()]
102            if len(chunk_experts) > 0:
103                raise ValueError(f"Unprocessed adjugate experts: {chunk_experts}")
104
105        if self._experts is not None:
106            # flatten `list[dict[str, Tensor]]` into `list[str]`
107            experts = [k for d in self._experts for k in d.keys()]
108            if len(experts) > 0:
109                raise ValueError(f"Unprocessed experts: {experts}")