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

move changes from local to BertModel

Browse source
This commit is contained in:
wheelspawn 2024-06-13 21:59:23 -05:00
parent a0c5a0e82f
commit 7a5d932eaf
1 changed files with 7 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

14
convert-hf-to-gguf.py
View file

@ -177,7 +177,7 @@ class Model:
return False
return name == (key_name + suffix)
def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias", ".beta", ".gamma")) -> str:
def map_tensor_name(self, name: str, try_suffixes: Sequence[str] = (".weight", ".bias")) -> str:
new_name = self.tensor_map.get_name(key=name, try_suffixes=try_suffixes)
if new_name is None:
raise ValueError(f"Can not map tensor {name!r}")
@ -249,9 +249,6 @@ class Model:
if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
continue
if name.startswith("bert."):
name = name.removeprefix("bert.")
old_dtype = data_torch.dtype
# convert any unsupported data types to float32
@ -2194,12 +2191,15 @@ class BertModel(Model):
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
del bid # unused
name = name.removeprefix("bert.")
# we are only using BERT for embeddings so we don't need the pooling layer
if name in ("embeddings.position_ids", "pooler.dense.weight", "pooler.dense.bias") or "cls." in name:
return [] # we don't need these
return [(self.map_tensor_name(name), data_torch)]
try_suffixes = (".weight", ".bias", ".beta", ".gamma")
return [(self.map_tensor_name(name, try_suffixes), data_torch)]
@Model.register("NomicBertModel")