root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-28 12:24:35 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

gguf : upd gguf conversion script

Browse Source
This commit is contained in:
M. Yusuf Sarıgöz 2023-07-29 13:31:07 +03:00
parent cc3dd7f042
commit 0317c41d98
1 changed files with 16 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
convert-llama-h5-to-gguf.py
View File

@ -12,11 +12,14 @@ from sentencepiece import SentencePieceProcessor
NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]' NDArray: 'TypeAlias' = 'np.ndarray[Any, Any]'
def permute(weights: NDArray, n_head: int) -> NDArray: def permute(weights: NDArray, n_head: int) -> NDArray:
return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:]) return (weights.reshape(n_head, 2, weights.shape[0] // n_head // 2, *weights.shape[1:])
.swapaxes(1, 2) .swapaxes(1, 2)
.reshape(weights.shape)) .reshape(weights.shape))
if len(sys.argv) < 3: if len(sys.argv) < 3:
print("Usage: convert-h5-to-ggml.py dir-model ftype\n") print("Usage: convert-h5-to-ggml.py dir-model ftype\n")
print(" ftype == 0 -> float32") print(" ftype == 0 -> float32")
@ -56,7 +59,6 @@ for name in list_vars.keys():
continue continue
tensor_count += 1 tensor_count += 1
#fout = open(fname_out, "wb")
gguf_writer = gguf.GGUFWriter.open(fname_out) gguf_writer = gguf.GGUFWriter.open(fname_out)
with open(dir_model + "/config.json", "r", encoding="utf-8") as f: with open(dir_model + "/config.json", "r", encoding="utf-8") as f:
@ -118,14 +120,14 @@ if Path( dir_model + "/tokenizer.model").is_file():
text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8") text = tokenizer.id_to_piece(i).replace("\u2581", " ").encode("utf-8")
score: float = tokenizer.get_score(i) score: float = tokenizer.get_score(i)
tokens.append( str(text) ); tokens.append(str(text))
scores.append( score ); scores.append(score)
print("write gguf tokens") print("write gguf tokens")
gguf_writer.write_string("tokenizer.ggml.model", "llama") gguf_writer.write_tokenizer_model("llama")
gguf_writer.write_array("tokenizer.ggml.tokens",tokens) gguf_writer.write_token_list(tokens)
gguf_writer.write_array("tokenizer.ggml.scores",scores) gguf_writer.write_token_scores(scores)
# TENSORS # TENSORS
@ -197,7 +199,7 @@ for name in list_vars.keys():
print(" Skip tensor: " + name) print(" Skip tensor: " + name)
continue continue
## permute these # permute these
if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"): if name.endswith(".q_proj.weight") or name.endswith(".k_proj.weight"):
print(" Permute tensor: " + name) print(" Permute tensor: " + name)
data = permute(data, hparams["num_attention_heads"]) data = permute(data, hparams["num_attention_heads"])
@ -221,7 +223,6 @@ for name in list_vars.keys():
data = data.astype(np.float32) data = data.astype(np.float32)
ftype_cur = 0 ftype_cur = 0
gguf_writer.write_tensor_padding()
gguf_writer.write_tensor(data) gguf_writer.write_tensor(data)
gguf_writer.close() gguf_writer.close()