 Jay Shah
|
bb169e2a54
revert to single tile scheduler
|
1 mesiac pred |
|
Jay Shah
|
483b26e541
add working fp8 varlen
|
1 mesiac pred |
|
jayhshah
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
2 mesiacov pred |
 Ying Zhang
|
dff976a84a
fixes
|
3 mesiacov pred |
|
Ying Zhang
|
7b4e68e04f
hopper local attention
|
3 mesiacov pred |
|
jayhshah
|
c92ca63268
FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173)
|
3 mesiacov pred |
|
jayhshah
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
4 mesiacov pred |
 Tri Dao
|
3aae9c18c1
Revert "Changes For FP8 (#1075)"
|
4 mesiacov pred |
 ganeshcolfax
|
1899c970c8
Changes For FP8 (#1075)
|
4 mesiacov pred |
 Ying Zhang
|
dfe1a59e4b
Add var-seq-len to FA3 fp16 / bf16 fwd (#1072)
|
4 mesiacov pred |
|
Tri Dao
|
74b0761ff7
[FA3] BF16 forward
|
5 mesiacov pred |
|
Tri Dao
|
7f67966cc7
FA3 initial code release
|
5 mesiacov pred |
