 Guillaume Wenzek
|
b7b31a3978
rm python 3.11 idiom
|
1 year ago |
|
Guillaume Wenzek
|
81cdf80eb9
WIP: MultiheadAttention_forward
|
1 year ago |
|
Guillaume Wenzek
|
f2b5007277
start testing with batch size > 1
|
1 year ago |
|
Guillaume Wenzek
|
f7d2c90ceb
_almost_contiguous
|
1 year ago |
|
Guillaume Wenzek
|
88b0690a72
split tests files
|
1 year ago |
|
Guillaume Wenzek
|
bfbafd9603
fix generation with beam_size=1
|
1 year ago |
|
Guillaume Wenzek
|
45f986055a
add naive tweaking of lprobs
|
1 year ago |
|
Guillaume Wenzek
|
2238cea072
SinusoidalPositionEncoder + WIP: TransformerEmbeddingFrontend
|
1 year ago |
|
Guillaume Wenzek
|
f1f33dbec1
has_layer + transformer decoder
|
1 year ago |
|
Guillaume Wenzek
|
3f1d6992f3
fix to_numpy for transposed tensors
|
1 year ago |
|
Guillaume Wenzek
|
a80a3b49f3
remove un-needed code
|
1 year ago |
|
Guillaume Wenzek
|
fa85f05545
test flash_attn
|
1 year ago |
|
Guillaume Wenzek
|
06d4ed1475
ggml.ne != np.shape
|
1 year ago |
|
Guillaume Wenzek
|
c2e6384e29
nb
|
1 year ago |
|
Guillaume Wenzek
|
3f5912b973
forward
|
1 year ago |
|
Guillaume Wenzek
|
c0bec21155
export model size in hparams
|
1 year ago |
|
Guillaume Wenzek
|
772f90dfdc
load_fairseq2_ggml_file
|
1 year ago |
|
Guillaume Wenzek
|
506dee42d8
move layers to fairseq2.cpp
|
1 year ago |
|
Guillaume Wenzek
|
6f32f3c06f
unity_graph -> unity_audio_encoder_graph
|
1 year ago |
|
Guillaume Wenzek
|
6cf3dfeb05
pass the input tensor explicitly
|
1 year ago |
|
Guillaume Wenzek
|
22f8430903
clearly split between vendored ggml.py and our utilities
|
1 year ago |
|
Guillaume Wenzek
|
9021fad301
use same convention than ggml.c
|
1 year ago |
|
Guillaume Wenzek
|
e7dc2b86fc
add ggml python bindings
|
1 year ago |
