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 →
2.4 KiB60 linesraw
1from __future__ import annotations
2
3from typing import Callable, Iterable, TYPE_CHECKING
4
5if TYPE_CHECKING:
6    from torch import Tensor
7
8from .base import ModelBase, TextModel, gguf
9
10from .llama import LlamaModel
11
12
13@ModelBase.register("ChameleonForConditionalGeneration")
14@ModelBase.register("ChameleonForCausalLM")  # obsolete
15# [TAG_HF_EXAMPLE_GATED] facebook/chameleon-7b is gated
16# [TAG_HF_EXAMPLE_MISSING]
17class ChameleonModel(TextModel):
18    model_arch = gguf.MODEL_ARCH.CHAMELEON
19
20    def set_gguf_parameters(self):
21        super().set_gguf_parameters()
22        self.gguf_writer.add_swin_norm(self.hparams.get("swin_norm", False))
23
24    def set_vocab(self):
25        self._set_vocab_gpt2()
26
27    @classmethod
28    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
29        name, gen = item
30
31        # ignore image tokenizer for now
32        # TODO: image support for Chameleon
33        if name.startswith("model.vqmodel"):
34            return None
35
36        return super().filter_tensors(item)
37
38    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
39        n_head = self.hparams["num_attention_heads"]
40        n_kv_head = self.hparams.get("num_key_value_heads")
41        hidden_dim = self.hparams.get("hidden_size")
42
43        if name.endswith(("q_proj.weight", "q_proj.bias")):
44            data_torch = LlamaModel.permute(data_torch, n_head, n_head)
45        if name.endswith(("k_proj.weight", "k_proj.bias")):
46            data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
47        if name.endswith(("q_norm.weight", "q_norm.bias")):
48            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_head, hidden_dim)
49        if name.endswith(("k_norm.weight", "k_norm.bias")):
50            data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
51
52        yield from super().modify_tensors(data_torch, name, bid)
53
54    # see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
55    @staticmethod
56    def _reverse_hf_permute(data_torch, n_heads, hidden_dim):
57        head_dim = hidden_dim // n_heads
58        data_torch = data_torch[0].view(2, head_dim // 2).t().reshape(1, -1)
59        data_torch = data_torch.repeat_interleave(n_heads, 0)
60        return data_torch