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

Merge 76543311ac into d7b31a9d84

Browse source
This commit is contained in:
fairydreaming 2025-02-10 17:34:14 +08:00 • committed by GitHub
commit 8bb3cb5e9c
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
10 changed files with 203 additions and 30 deletions
Download patch file Download diff file Expand all files Collapse all files

22
convert_hf_to_gguf.py
View file

@ -4136,6 +4136,28 @@ class DeepseekV2Model(Model):
else:
return []
if name.endswith("kv_b_proj.weight"):
name_kb = name.replace("kv_b_proj", "k_b_proj")
name_vb = name.replace("kv_b_proj", "v_b_proj")
n_head_kv = self.hparams["num_key_value_heads"]
v_head_dim = self.hparams["v_head_dim"]
qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
assert data_torch.shape[0] == n_head_kv * (v_head_dim + qk_nope_head_dim)
kv_b = data_torch.view(n_head_kv, v_head_dim + qk_nope_head_dim, data_torch.shape[-1])
k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
k_b = k_b.transpose(1, 2)
k_b = k_b.reshape(n_head_kv * data_torch.shape[-1], qk_nope_head_dim)
v_b = v_b.reshape(n_head_kv * v_head_dim, data_torch.shape[-1])
return [
(self.map_tensor_name(name), data_torch),
(self.map_tensor_name(name_kb), k_b),
(self.map_tensor_name(name_vb), v_b)
]
return [(self.map_tensor_name(name), data_torch)]
def prepare_tensors(self):