 Guillaume Wenzek
|
b7b31a3978
rm python 3.11 idiom
|
2 years ago |
|
Guillaume Wenzek
|
81cdf80eb9
WIP: MultiheadAttention_forward
|
2 years ago |
|
Guillaume Wenzek
|
f2b5007277
start testing with batch size > 1
|
2 years ago |
|
Guillaume Wenzek
|
eb80195345
use ggml_diag_mask_inf
|
2 years ago |
|
Guillaume Wenzek
|
f7d2c90ceb
_almost_contiguous
|
2 years ago |
|
Guillaume Wenzek
|
88b0690a72
split tests files
|
2 years ago |
|
Guillaume Wenzek
|
bfbafd9603
fix generation with beam_size=1
|
2 years ago |
|
Guillaume Wenzek
|
45f986055a
add naive tweaking of lprobs
|
2 years ago |
|
Guillaume Wenzek
|
7c9b2a1b95
pass correct prefix sequence in test
|
2 years ago |
|
Guillaume Wenzek
|
c7b89f32f4
disable flash attn because of cross attention
|
2 years ago |
|
Guillaume Wenzek
|
bf9b2510df
missing file
|
2 years ago |
|
Guillaume Wenzek
|
dcb9535666
wip: generate_sequence
|
2 years ago |
|
Guillaume Wenzek
|
78e7c9a311
fix TransformerEmbeddingFrontend
|
2 years ago |
|
Guillaume Wenzek
|
2238cea072
SinusoidalPositionEncoder + WIP: TransformerEmbeddingFrontend
|
2 years ago |
|
Guillaume Wenzek
|
2fb09f34fb
generate fairseq2.cpp
|
2 years ago |
|
Guillaume Wenzek
|
f1f33dbec1
has_layer + transformer decoder
|
2 years ago |
|
Guillaume Wenzek
|
9b157d0384
allclose
|
2 years ago |
|
Guillaume Wenzek
|
e1faba0c5d
text encoder
|
2 years ago |
|
Guillaume Wenzek
|
b81061704c
StandardTransformerEncoderLayer_forward
|
2 years ago |
|
Guillaume Wenzek
|
e7c3b7a4ba
allow flash attn
|
2 years ago |
|
Guillaume Wenzek
|
b07a08102a
MultiheadAttention_forward
|
2 years ago |
|
Guillaume Wenzek
|
3f1d6992f3
fix to_numpy for transposed tensors
|
2 years ago |
|
Guillaume Wenzek
|
a80a3b49f3
remove un-needed code
|
2 years ago |
 Ning
|
4aaca94e11
Self attention (SDPA) in unity.cpp (#34)
|
2 years ago |
|
Guillaume Wenzek
|
be236b8584
remove size/init functions
|
2 years ago |
|
Guillaume Wenzek
|
fa85f05545
test flash_attn
|
2 years ago |
|
Guillaume Wenzek
|
06d4ed1475
ggml.ne != np.shape
|
2 years ago |
|
Guillaume Wenzek
|
c2e6384e29
nb
|
2 years ago |
|
Guillaume Wenzek
|
3f5912b973
forward
|
2 years ago |
|
Guillaume Wenzek
|
2b6130855c
handle 4 dims tensors and eof
|
2 years ago |
