Support for Phi-2 #4490
ggerganov
merged
17 commits into
ggml-org:master
from
ebeyabraham:master
phi2 implementation
12cc80cb
fix breaking change
e2076553
phi-2 : various fixes
a2a3d2c8
phi-2 : use layer norm eps
aa5c881a
py : whitespaces
7500fa2f
llama : fix meta KV override bug
5469d82d
convert : phi don't add BOS token
a878be4c
ggerganov
commented
on 2023-12-16
convert : revert "added_tokens_decoder" change
0b6ffa58
phi-2 : scale Q instead of KQ for better precision
0644c3be
ggerganov
commented
on 2023-12-16
devic1
approved these changes
on 2023-12-17
ggml : fix NeoX rope to rotate just first n_dims
f703ca8a
cuda : less diff in the rope_neox kernel
42e95258
Merge branch 'master' into HEAD
a8d2a6f3
ggml : add ggml_mul_mat_set_prec
18c67bdd
Update ggml-cuda.cu
3c8d6b16
Update ggml-cuda.cu
30338c56
cuda : ggml_cuda_op_mul_mat_cublas support F32 precision
7ea427db
cuda : remove oboslete comment
c02412c3
ggerganov
approved these changes
on 2023-12-18
Login to write a write a comment.
Login via GitHub