happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

fix: remove Q/K permutation (NEOX RoPE doesn't need it)

This commit is contained in:
alielfilali01 2026-02-12 08:22:44 +00:00
parent 56571c3238
commit d9a442f602
1 changed files with 0 additions and 12 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
convert_hf_to_gguf.py
View File

@ -8537,18 +8537,6 @@ class Jais2Model(TextModel):
head_dim = hparams.get("head_dim", hparams["hidden_size"] // hparams["num_attention_heads"])
self.gguf_writer.add_rope_dimension_count(head_dim)
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# Jais2 uses LLaMA-style RoPE (rotate_half), requiring Q/K permutation
n_head = self.hparams["num_attention_heads"]
n_kv_head = self.hparams.get("num_key_value_heads", n_head)
if name.endswith(("q_proj.weight", "q_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_head)
if name.endswith(("k_proj.weight", "k_proj.bias")):
data_torch = LlamaModel.permute(data_torch, n_head, n_kv_head)
return [(self.map_tensor_name(name), data_torch)]
@ModelBase.register("JAISLMHeadModel")
class JaisModel(TextModel):