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

clean up and simplify XLMRoberta conversion

Browse source
This commit is contained in:
Douglas Hanley 2024-07-24 01:50:16 -05:00
parent 1c7996a17c
commit 013a08f725
1 changed files with 3 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

9
convert_hf_to_gguf.py
View file

@ -2518,7 +2518,7 @@ class XLMRobertaModel(BertModel):
tokens: list[bytes] = [f"[PAD{i}]".encode("utf-8") for i in range(vocab_size)]
scores: list[float] = [-10000.0] * vocab_size
toktypes: list[int] = [SentencePieceTokenTypes.UNKNOWN] * vocab_size
toktypes: list[int] = [SentencePieceTokenTypes.UNUSED] * vocab_size
for token_id in range(tokenizer.vocab_size()):
piece = tokenizer.IdToPiece(token_id)
@ -2549,7 +2549,7 @@ class XLMRobertaModel(BertModel):
# realign tokens (see HF tokenizer code)
tokens = [b'<s>', b'<pad>', b'</s>', b'<unk>'] + tokens[3:-1]
scores = [0.0, -10000.0, 0.0, -10000.0] + scores[3:-1]
scores = [0.0, 0.0, 0.0, 0.0] + scores[3:-1]
toktypes = [
SentencePieceTokenTypes.CONTROL,
SentencePieceTokenTypes.CONTROL,
@ -2577,15 +2577,12 @@ class XLMRobertaModel(BertModel):
def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
# position embeddings start at pad_token_id + 1, so just chop down the weight tensor
if name == "embeddings.position_embeddings.weight":
del bid # unused
if self._position_offset is not None:
data_torch = data_torch[self._position_offset:,:]
return [(self.map_tensor_name(name), data_torch)]
return super().modify_tensors(data_torch, name, bid)
@Model.register("GemmaForCausalLM")
class GemmaModel(Model):
model_arch = gguf.MODEL_ARCH.GEMMA