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

convert-hf : display tensor shape

Browse source
This commit is contained in:
Francis Couture-Harpin 2024-05-01 16:59:21 -04:00
parent dcd8dfa1b5
commit 21068b6bdf
1 changed files with 6 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

8
convert-hf-to-gguf.py
View file

@ -191,6 +191,7 @@ class Model(Protocol):
break break
for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)): for new_name, data in ((n, d.squeeze().numpy()) for n, d in self.modify_tensors(data_torch, name, bid)):
data: np.ndarray = data # type hint
n_dims = len(data.shape) n_dims = len(data.shape)
data_dtype = data.dtype data_dtype = data.dtype
@ -211,7 +212,11 @@ class Model(Protocol):
if self.ftype == 1 and data_dtype == np.float32 and (name.endswith(".weight") and n_dims >= 2 or extra_f16) and not extra_f32: if self.ftype == 1 and data_dtype == np.float32 and (name.endswith(".weight") and n_dims >= 2 or extra_f16) and not extra_f32:
data = data.astype(np.float16) data = data.astype(np.float16)
print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}") # reverse shape to make it similar to the internal ggml dimension order
shape_str = f"{{{', '.join(str(n) for n in reversed(data.shape))}}}"
# n_dims is implicit in the shape
print(f"{new_name}, shape = {shape_str}, {old_dtype} --> {data.dtype}")
self.gguf_writer.add_tensor(new_name, data) self.gguf_writer.add_tensor(new_name, data)
@ -1774,7 +1779,6 @@ class InternLM2Model(Model):
tokenizer = SentencePieceProcessor() tokenizer = SentencePieceProcessor()
tokenizer.LoadFromFile(str(tokenizer_path)) tokenizer.LoadFromFile(str(tokenizer_path))
tokenizer.serialized_model_proto
vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size()) vocab_size = self.hparams.get('vocab_size', tokenizer.vocab_size())