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 →
3.2 KiB80 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("GPT2LMHeadModel")
14@ModelBase.example("openai-community/gpt2")
15class GPT2Model(TextModel):
16    model_arch = gguf.MODEL_ARCH.GPT2
17
18    def set_gguf_parameters(self):
19        self.gguf_writer.add_block_count(self.block_count)
20        self.gguf_writer.add_context_length(self.hparams["n_ctx"])
21        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
22        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
23        self.gguf_writer.add_head_count(self.hparams["n_head"])
24        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
25        self.gguf_writer.add_file_type(self.ftype)
26
27    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
28        # we don't need these
29        if name.endswith((".attn.bias", ".attn.masked_bias")):
30            yield from super().modify_tensors(data_torch, name, bid)
31            return
32
33        if name.endswith((".c_attn.weight", ".c_proj.weight", ".c_fc.weight", ".c_proj.weight")):
34            data_torch = data_torch.transpose(1, 0)
35
36        new_name = self.map_tensor_name(name)
37
38        yield from super().modify_tensors(data_torch, new_name, bid)
39
40
41@ModelBase.register("RuGPT3XLForCausalLM")
42@ModelBase.example("evilfreelancer/ruGPT3XL")
43class RuGPT3XLModel(TextModel):
44    model_arch = gguf.MODEL_ARCH.GPT2
45
46    _qkv_parts: list[dict[str, Tensor]] | None = None
47
48    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
49        # Fuse separate Q, K, V projections into a single QKV tensor
50        if ".self_attn.q_proj." in name or ".self_attn.k_proj." in name or ".self_attn.v_proj." in name:
51            suffix = "weight" if name.endswith(".weight") else "bias"
52            part = "q" if ".q_proj." in name else ("k" if ".k_proj." in name else "v")
53            key = f"{part}.{suffix}"
54
55            assert bid is not None
56            if self._qkv_parts is None:
57                self._qkv_parts = [{} for _ in range(self.block_count)]
58            self._qkv_parts[bid][key] = data_torch
59
60            q_key, k_key, v_key = f"q.{suffix}", f"k.{suffix}", f"v.{suffix}"
61            if all(k in self._qkv_parts[bid] for k in [q_key, k_key, v_key]):
62                q = self._qkv_parts[bid].pop(q_key)
63                k = self._qkv_parts[bid].pop(k_key)
64                v = self._qkv_parts[bid].pop(v_key)
65                data_torch = torch.cat([q, k, v], dim=0)
66                name = self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, bid, f".{suffix}")
67                logger.debug(f"Fused Q/K/V {suffix} for layer {bid} -> {name}")
68            else:
69                return
70
71        yield from super().modify_tensors(data_torch, name, bid)
72
73    def prepare_tensors(self):
74        super().prepare_tensors()
75
76        if self._qkv_parts is not None:
77            # flatten `list[dict[str, Tensor]]` into `list[str]`
78            parts = [f"({i}){k}" for i, d in enumerate(self._qkv_parts) for k in d.keys()]
79            if len(parts) > 0:
80                raise ValueError(f"Unprocessed Q/K/V parts: {parts}")