ed
/
rwkv.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Update file format documentation

This commit is contained in:
saharNooby 2023-04-14 18:59:16 +04:00
parent e84c446d95
commit e4268a36c8
1 changed files with 9 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

14
rwkv/convert_pytorch_to_ggml.py
View File

@ -12,7 +12,7 @@
# int32 n_vocab; # int32 n_vocab;
# int32 n_embed; # int32 n_embed;
# int32 n_layer; # int32 n_layer;
# // 0 if float32, 1 if float16. # // 0 if float32, 1 if float16, 2 if Q4_0, 3 if Q4_1, 4 if Q4_1_O.
# int32 data_type; # int32 data_type;
# // Read until EOF. # // Read until EOF.
# Parameter[] parameters; # Parameter[] parameters;
@ -21,14 +21,18 @@
# Parameter { # Parameter {
# int32 dim_count; # int32 dim_count;
# int32 key_length; # int32 key_length;
# // 0 if float32, 1 if float16. # // 0 if float32, 1 if float16, 2 if Q4_0, 3 if Q4_1, 4 if Q4_1_O.
# int32 data_type; # int32 data_type;
# // Same values and order as in PyTorch's tensor.shape # // Compared to PyTorch's tensor.shape, dimension order is reversed here!
# int32[dim_count] shape; # int32[dim_count] shape;
# // Keys are like "emb.weight", "block.0.ln1.weight". # // Keys are like "emb.weight", "block.0.ln1.weight".
# uint8[key_length] key_utf8; # uint8[key_length] key_utf8;
# // Can be either float32 or float16. # // float32: 4 * element_count bytes.
# float[product(shape)] data; # // float16: 2 * element_count bytes.
# // Q4_0: element_count / 32 * 20 bytes.
# // Q4_1: element_count / 32 * 24 bytes.
# // Q4_1_O: element_count / 32 * 24 bytes.
# byte[] data;
# } # }
import os import os