 Tri Dao
|
5458c78e6d
Remove sink token
|
4 napja |
|
Tri Dao
|
6752d62aa4
Add dynamic splits
|
1 hete |
 Ted Zadouri
|
06e34f62d1
Enable MLA flag in FA3 (rope=64, latent=512) (#1504)
|
1 hete |
|
Tri Dao
|
15cf7ee435
Rename collective_mainloop -> mainloop, move tile_scheduler variable
|
3 hete |
|
Tri Dao
|
4e8496a781
Support Qv
|
3 hete |
|
Tri Dao
|
ed53b5fc4c
Fix compilation for FP8 when kHeadDimV != kHeadDim
|
3 hete |
|
Tri Dao
|
4c8819d8c6
Rename Mma0->MmaQK, Mma1->MmaPV, use Cluster only if hdimV >= 192
|
3 hete |
|
Tri Dao
|
6d199aa207
Fix shape_O in epilogue params when kHeadDimV != kHeadDim
|
1 hónapja |
|
Tri Dao
|
2a204125ae
Implement HeadDim_V != HeadDim_QK, support hdimQK=192, hdimV=128
|
1 hónapja |
|
Tri Dao
|
454ce31594
[FA3] Compile with nvcc 12.8 instead of 12.3
|
1 hónapja |
|
Tri Dao
|
1e3208566a
Tune tile sizes for compilation
|
1 hónapja |
|
Tri Dao
|
df5fe55264
Change tile sizes for Sm8x to reduce stack frame
|
1 hónapja |
|
Tri Dao
|
180ff782dd
Template for Sm86
|
1 hónapja |
|
Tri Dao
|
7bc3f031a4
Compile for both Sm80 and Sm90
|
1 hónapja |
|
Tri Dao
|
7a802796e1
Big refactor and update
|
1 hónapja |
 Kai Londenberg
|
b443207c1f
Paged Attention support for FA3 (#1268)
|
3 hónapja |
 jayhshah
|
a5a75274bc
FA3 kvcache + split kv + gqa parallelization (#1236)
|
4 hónapja |
 Ying Zhang
|
1c9717d699
address comments
|
5 hónapja |
|
Ying Zhang
|
dff976a84a
fixes
|
6 hónapja |
|
Ying Zhang
|
7b4e68e04f
hopper local attention
|
6 hónapja |
|
Ying Zhang
|
db80387343
Add seqused_q in fwd / bwd and seqused_k in bwd.
|
6 hónapja |
|
jayhshah
|
c92ca63268
FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173)
|
6 hónapja |
|
Tri Dao
|
bafe253042
[FA3] Bwd
|
7 hónapja |
|
jayhshah
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
7 hónapja |
|
Tri Dao
|
3aae9c18c1
Revert "Changes For FP8 (#1075)"
|
7 hónapja |
 ganeshcolfax
|
1899c970c8
Changes For FP8 (#1075)
|
7 hónapja |
 Ying Zhang
|
dfe1a59e4b
Add var-seq-len to FA3 fp16 / bf16 fwd (#1072)
|
7 hónapja |
 Cameron Shinn
|
cb516f855b
Remove torchlib dependency from cpp files (#1083)
|
7 hónapja |
|
Tri Dao
|
74b0761ff7
[FA3] BF16 forward
|
7 hónapja |
|
Tri Dao
|
7f67966cc7
FA3 initial code release
|
7 hónapja |
