 Jay Shah
|
b3d60fa3a5
prune more dead code
|
4 달 전 |
|
Jay Shah
|
c06cc0ba9f
change cu_seqlens_k to seqused_k for kv cache api
|
4 달 전 |
|
Jay Shah
|
7c1473e0e5
remove Is_batch_dynamic from seqlen traits and handle fp8 perf regression using smem boolean
|
4 달 전 |
|
Jay Shah
|
8374e1fa78
remove test code
|
4 달 전 |
|
Jay Shah
|
35f3542442
refactor names
|
4 달 전 |
|
Jay Shah
|
d437d3dd5c
remove smem usage for when rmem -> gmem epilogue is used
|
4 달 전 |
|
Jay Shah
|
81d402463e
prune unused code
|
5 달 전 |
|
Jay Shah
|
6111666130
consolidate nblock min max methods
|
5 달 전 |
|
Jay Shah
|
b5cac6d586
rebase with Is_local disabled temporarily
|
5 달 전 |
|
Jay Shah
|
2472e5e0b4
add 'in principle' fp8 kv cache support
|
5 달 전 |
|
Jay Shah
|
e36e004cb3
change fp8 code path to allow for split kernel and kv cache without perf regression
|
5 달 전 |
|
Jay Shah
|
f3e5bd47cb
update parameters
|
5 달 전 |
|
Jay Shah
|
68ff3f7bee
add 1 mma warpgroup option, enable splitkv for hdim 256
|
6 달 전 |
|
Jay Shah
|
0a1a0c22b6
refactor for split kv
|
6 달 전 |
|
Jay Shah
|
0c4cea947b
add in lse writeout and store zero
|
6 달 전 |
|
Jay Shah
|
64a9cfb0fe
add causal logic
|
6 달 전 |
|
Jay Shah
|
5704a1f424
fix some errors
|
6 달 전 |
|
Jay Shah
|
535b8279bb
complete gqa parallel changes for non-causal
|
6 달 전 |
|
Jay Shah
|
fc8f704f28
decouple types of seqlen traits q and k
|
6 달 전 |
|
Jay Shah
|
74f160ba43
change template parameter for SeqLenTraits for ease of further extension
|
6 달 전 |
 Ying Zhang
|
7b4e68e04f
hopper local attention
|
6 달 전 |
|
jayhshah
|
c92ca63268
FA3 FP8 qkv descales + restore max offset for h128 causal + added sync for producer WG (#1173)
|
6 달 전 |
|
Ying Zhang
|
a3a257c71d
Fix out-of-bound writes for var-seq-len zero-length KVs
|
6 달 전 |
|
jayhshah
|
5018ac6ac5
Fp8 kernel with "in-kernel" transpose of V in producer (#1100)
|
7 달 전 |
 Ying Zhang
|
dfe1a59e4b
Add var-seq-len to FA3 fp16 / bf16 fwd (#1072)
|
7 달 전 |
 Tri Dao
|
74b0761ff7
[FA3] BF16 forward
|
7 달 전 |
|
Tri Dao
|
7f67966cc7
FA3 initial code release
|
7 달 전 |
