 Guillaume Wenzek
|
d5b035f230
fix PositionalEmbedding with cache
|
1 год назад |
|
Guillaume Wenzek
|
8c074387a9
s2tt generate the sample if needed
|
1 год назад |
|
Guillaume Wenzek
|
ac4a2eada9
job.num_threads
|
2 лет назад |
|
Guillaume Wenzek
|
f6d810543d
fix incremental decoding !
|
1 год назад |
|
Guillaume Wenzek
|
522b97234e
WIP: simple failing test case
|
1 год назад |
|
Guillaume Wenzek
|
2c543185e2
add kv_cache to fairseq2_model
|
1 год назад |
|
Guillaume Wenzek
|
44a4ca129a
read model config from "layer_config"
|
2 лет назад |
|
Guillaume Wenzek
|
eb7810b81f
force little-endian
|
2 лет назад |
|
Guillaume Wenzek
|
c31926c1a8
working out new way of saving hparams
|
2 лет назад |
 Ning
|
1589c529bf
unity.cpp speech_encoder_frontend+speech_encoder+adaptor (#81)
|
2 лет назад |
|
Guillaume Wenzek
|
e38442d1f0
rename mask to attn_mask in MultiheadAttention_forward
|
2 лет назад |
|
Guillaume Wenzek
|
6fbb465f2b
generate_sequence return full results
|
2 лет назад |
|
Guillaume Wenzek
|
1756897d23
simplify _finalize_hypothesis
|
2 лет назад |
|
Guillaume Wenzek
|
f49763de86
fix beam size and scores
|
2 лет назад |
|
Guillaume Wenzek
|
c28db8c8ac
wip beam-size=2
|
2 лет назад |
|
Guillaume Wenzek
|
b24dbe3030
batching -> TransformerEmbeddingFrontend_forward
|
2 лет назад |
|
Guillaume Wenzek
|
86993cbd00
fix StandardTransformerEncoder
|
2 лет назад |
|
Guillaume Wenzek
|
28ed039370
fix MultiheadAttention_forward
|
2 лет назад |
|
Guillaume Wenzek
|
81cdf80eb9
WIP: MultiheadAttention_forward
|
2 лет назад |
|
Guillaume Wenzek
|
eb80195345
use ggml_diag_mask_inf
|
2 лет назад |
|
Guillaume Wenzek
|
88b0690a72
split tests files
|
2 лет назад |
|
Guillaume Wenzek
|
bfbafd9603
fix generation with beam_size=1
|
2 лет назад |
|
Guillaume Wenzek
|
45f986055a
add naive tweaking of lprobs
|
2 лет назад |
|
Guillaume Wenzek
|
7c9b2a1b95
pass correct prefix sequence in test
|
2 лет назад |
|
Guillaume Wenzek
|
c7b89f32f4
disable flash attn because of cross attention
|
2 лет назад |
|
Guillaume Wenzek
|
dcb9535666
wip: generate_sequence
|
2 лет назад |
|
Guillaume Wenzek
|
78e7c9a311
fix TransformerEmbeddingFrontend
|
2 лет назад |
|
Guillaume Wenzek
|
2238cea072
SinusoidalPositionEncoder + WIP: TransformerEmbeddingFrontend
|
2 лет назад |
|
Guillaume Wenzek
|
2fb09f34fb
generate fairseq2.cpp
|
2 лет назад |
|
Guillaume Wenzek
|
f1f33dbec1
has_layer + transformer decoder
|
2 лет назад |
