Change SPV_KHR_bfloat16 workaround patch to work for different drivers #5585
Conversation
|
Do you know why there are no artifacts in the second attempts? Only the first attempt run have the artifacts of the triton pacakge, but what within is |
|
Please don’t try yet, noticed from CI failures that still need more changes. |
whitneywhtsang
force-pushed
the
whitneywhtsang/wi5574
branch
from
1c260fb to
43b020f
Compare
whitneywhtsang
force-pushed
the
whitneywhtsang/wi5574
branch
from
43b020f to
7d3d898
Compare
|
@Stonepia It should be good now, please give it another try. |
| - return mapType(T, BM->addFloatType(16, FPEncodingBFloat16KHR)); | ||
| + // Workaround for LTS2 driver. | ||
| + const char *driverVersion = std::getenv("INTEL_XPU_BACKEND_DRIVER_VERSION"); | ||
| + if (driverVersion && std::string(driverVersion) != "1.6.33578+38") { |
There was a problem hiding this comment.
If the LTS2 driver is updated, this code will break. Is it enough to use some substring?
There was a problem hiding this comment.
Do you know what's the driver version pattern? Are you suggesting to omit +38?
There was a problem hiding this comment.
Unfortunately no, my guess is that only this part will not change: 1.6.*
There was a problem hiding this comment.
I saw 1.6.35096+9 on https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/19873342424/job/56954531877 which uses Rolling.
Check the driver version at runtime instead of build time.
Fixes #5574