vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

return qk norm weights and biases to original format

Browse source
This commit is contained in:
nopperl 2024-07-17 12:53:56 +02:00
parent 654b1b35b3
commit c460d5c3bb
1 changed files with 14 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

14
convert_hf_to_gguf.py
View file

@ -3441,14 +3441,28 @@ class ChameleonModel(Model):
n_head = self.hparams["num_attention_heads"] n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams.get("num_key_value_heads") n_kv_head = self.hparams.get("num_key_value_heads")
hidden_dim = self.hparams.get("hidden_size")
if name.endswith(("q_proj.weight", "q_proj.bias")): if name.endswith(("q_proj.weight", "q_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_head) data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith(("k_proj.weight", "k_proj.bias")): if name.endswith(("k_proj.weight", "k_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head) data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
if name.endswith(("q_norm.weight", "q_norm.bias")):
data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_head, hidden_dim)
if name.endswith(("k_norm.weight", "k_norm.bias")):
data_torch = ChameleonModel._reverse_hf_permute(data_torch, n_kv_head, hidden_dim)
return [(self.map_tensor_name(name), data_torch)] return [(self.map_tensor_name(name), data_torch)]
# see: https://github.com/huggingface/transformers/blob/72fb02c47dbbe1999ae105319f24631cad6e2e00/src/transformers/models/chameleon/convert_chameleon_weights_to_hf.py#L176-L203
@staticmethod
def _reverse_hf_permute(data_torch, n_heads, hidden_dim):
head_dim = hidden_dim // n_heads
data_torch = data_torch[0].view(2, head_dim // 2).t().reshape(1, -1)
data_torch = data_torch.repeat_interleave(n_heads, 0)
return data_torch
###### CONVERSION LOGIC ###### ###### CONVERSION LOGIC ######