sgl-flash-attn3: switch to torch stable ABI - #1054
Conversation
|
/kernel-bot build-and-stage sgl-flash-attn3 |
|
Build request processed. Command: Dispatched (1):
|
|
@adarshxs could you look into the build failures? |
ptxas 13.2 hits an internal compiler error (C7907) when building the
sm90 bf16 aggregators (flash_fwd_hdimall_bf16{,_softcap}_sm90.cu).
CUDA 12.8 and 13.0 build cleanly and the 13.0 build is backwards
compatible, so cap the matrix at 13.0.
|
/kernel-bot build-and-stage sgl-flash-attn3 |
|
Build request processed. Command: Dispatched (1):
|
|
/kernel-bot build-and-stage sgl-flash-attn3 |
|
I can only run |
|
/kernel-bot build-and-stage sgl-flash-attn3 |
|
Build request processed. Command: Dispatched (1):
|
|
@sayakpaul already done in this |
|
/kernel-bot build-and-stage sgl-flash-attn3 |
|
Build request processed. Command: Dispatched (1):
|
What does this PR do?
Ports
sgl-flash-attn3to the Torch stable ABI.Motivation
Changes
Testing
Checklist