root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 19:04:35 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

fix convert-lora-to-ggml.py (#2738)

Browse Source
This commit is contained in:
slaren 2023-08-23 16:46:54 +02:00 committed by GitHub
parent 5290c38e6e
commit 335acd2ffd
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
1 changed files with 17 additions and 16 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

33
convert-lora-to-ggml.py
View File

@ -6,23 +6,22 @@ import struct
import sys import sys
from typing import Any, Dict, Sequence, TextIO from typing import Any, Dict, Sequence, TextIO
import numpy as np
import torch import torch
from convert import DATA_TYPE_TO_FTYPE, NUMPY_TYPE_TO_DATA_TYPE, DataType NUMPY_TYPE_TO_FTYPE: Dict[str, int] = {"float32": 0, "float16": 1}
HF_SUBLAYER_TO_GGML = { HF_SUBLAYER_TO_GGML = {
"self_attn.q_proj": "attention.wq", "self_attn.q_proj": "attn_q",
"self_attn.k_proj": "attention.wk", "self_attn.k_proj": "attn_k",
"self_attn.v_proj": "attention.wv", "self_attn.v_proj": "attn_v",
"self_attn.o_proj": "attention.wo", "self_attn.o_proj": "attn_output",
"mlp.gate_proj": "feed_forward.w1", "mlp.gate_proj": "ffn_gate",
"mlp.down_proj": "feed_forward.w2", "mlp.down_proj": "ffn_down",
"mlp.up_proj": "feed_forward.w3", "mlp.up_proj": "ffn_up",
"input_layernorm": "attention_norm", "input_layernorm": "attn_norm",
"post_attention_layernorm": "ffn_norm", "post_attention_layernorm": "ffn_norm",
# "norm": "norm",
# "embed_tokens": "tok_embeddings",
# "lm_head": "output",
} }
@ -39,7 +38,7 @@ def translate_tensor_name(t: str) -> str:
sys.exit(1) sys.exit(1)
output_string = ( output_string = (
f"layers.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}" f"blk.{nn}.{HF_SUBLAYER_TO_GGML[sub_layer]}.weight.lora{lora_type}"
) )
return output_string return output_string
else: else:
@ -54,12 +53,14 @@ def write_file_header(fout: TextIO, params: Dict[str, Any]) -> None:
# https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int # https://opendelta.readthedocs.io/en/latest/modules/deltas.html says that `lora_alpha` is an int
# but some models ship a float value instead # but some models ship a float value instead
# let's convert to int, but fail if lossless conversion is not possible # let's convert to int, but fail if lossless conversion is not possible
assert int(params["lora_alpha"]) == params["lora_alpha"], "cannot convert float to int losslessly" assert (
int(params["lora_alpha"]) == params["lora_alpha"]
), "cannot convert float to int losslessly"
fout.write(struct.pack("i", int(params["lora_alpha"]))) fout.write(struct.pack("i", int(params["lora_alpha"])))
def write_tensor_header( def write_tensor_header(
self, name: str, shape: Sequence[int], data_type: DataType self, name: str, shape: Sequence[int], data_type: np.dtype
) -> None: ) -> None:
sname = name.encode("utf-8") sname = name.encode("utf-8")
fout.write( fout.write(
@ -67,7 +68,7 @@ def write_tensor_header(
"iii", "iii",
len(shape), len(shape),
len(sname), len(sname),
DATA_TYPE_TO_FTYPE[NUMPY_TYPE_TO_DATA_TYPE[data_type]], NUMPY_TYPE_TO_FTYPE[data_type.name],
) )
) )
fout.write(struct.pack("i" * len(shape), *shape[::-1])) fout.write(struct.pack("i" * len(shape), *shape[::-1]))