Guillaume Wenzek
|
1756897d23
simplify _finalize_hypothesis
|
1 жил өмнө |
Guillaume Wenzek
|
f49763de86
fix beam size and scores
|
1 жил өмнө |
Guillaume Wenzek
|
c28db8c8ac
wip beam-size=2
|
1 жил өмнө |
Guillaume Wenzek
|
b24dbe3030
batching -> TransformerEmbeddingFrontend_forward
|
1 жил өмнө |
Guillaume Wenzek
|
86993cbd00
fix StandardTransformerEncoder
|
1 жил өмнө |
Guillaume Wenzek
|
28ed039370
fix MultiheadAttention_forward
|
1 жил өмнө |
Guillaume Wenzek
|
81cdf80eb9
WIP: MultiheadAttention_forward
|
1 жил өмнө |
Guillaume Wenzek
|
eb80195345
use ggml_diag_mask_inf
|
1 жил өмнө |
Guillaume Wenzek
|
88b0690a72
split tests files
|
1 жил өмнө |
Guillaume Wenzek
|
bfbafd9603
fix generation with beam_size=1
|
1 жил өмнө |
Guillaume Wenzek
|
45f986055a
add naive tweaking of lprobs
|
1 жил өмнө |
Guillaume Wenzek
|
7c9b2a1b95
pass correct prefix sequence in test
|
1 жил өмнө |
Guillaume Wenzek
|
c7b89f32f4
disable flash attn because of cross attention
|
1 жил өмнө |
Guillaume Wenzek
|
dcb9535666
wip: generate_sequence
|
1 жил өмнө |
Guillaume Wenzek
|
78e7c9a311
fix TransformerEmbeddingFrontend
|
1 жил өмнө |
Guillaume Wenzek
|
2238cea072
SinusoidalPositionEncoder + WIP: TransformerEmbeddingFrontend
|
1 жил өмнө |
Guillaume Wenzek
|
2fb09f34fb
generate fairseq2.cpp
|
1 жил өмнө |
Guillaume Wenzek
|
f1f33dbec1
has_layer + transformer decoder
|
1 жил өмнө |
Guillaume Wenzek
|
e1faba0c5d
text encoder
|
1 жил өмнө |
Guillaume Wenzek
|
b81061704c
StandardTransformerEncoderLayer_forward
|
1 жил өмнө |
Guillaume Wenzek
|
e7c3b7a4ba
allow flash attn
|
1 жил өмнө |
Guillaume Wenzek
|
b07a08102a
MultiheadAttention_forward
|
1 жил өмнө |
Guillaume Wenzek
|
be236b8584
remove size/init functions
|
1 жил өмнө |
Guillaume Wenzek
|
fa85f05545
test flash_attn
|
1 жил өмнө |
Guillaume Wenzek
|
06d4ed1475
ggml.ne != np.shape
|
1 жил өмнө |
Guillaume Wenzek
|
c2e6384e29
nb
|
1 жил өмнө |
Guillaume Wenzek
|
3f5912b973
forward
|
1 жил өмнө |
Guillaume Wenzek
|
772f90dfdc
load_fairseq2_ggml_file
|
1 жил өмнө |
Guillaume Wenzek
|
506dee42d8
move layers to fairseq2.cpp
|
1 жил өмнө |