 Tri Dao
|
a157cc8c9b
[FT] Implement MQA/GQA
|
1 year ago |
|
Tri Dao
|
2800efc71f
[FT] rotary_cos/sin should have batch_size dimension
|
1 year ago |
|
Tri Dao
|
62e9814466
[Rotary] Make sure frequency calculation is in fp32
|
1 year ago |
|
Tri Dao
|
48bc6eacd6
[Gen] Add rotary base as an argument to FT attention kernel
|
1 year ago |
|
Tri Dao
|
be1afaa276
[Gen, FT] Use fp32 accum for FMA
|
1 year ago |
|
Tri Dao
|
f266fc7262
[Gen, FT] Use tlength instead of params.timestep for rotary
|
1 year ago |
|
Tri Dao
|
a01d1213d7
[Gen] Add kernel from FasterTransformer for benchmarking
|
1 year ago |
