Enable triton kernels matmul tests #5128
Conversation
|
Is this for testing? If it is not ready for review, please convert to draft. |
wdziurdz
force-pushed
the
dev/wdziurdz/test-matmul-1
branch
2 times, most recently
from
84862d8 to
aafbe1a
Compare
wdziurdz
force-pushed
the
dev/wdziurdz/test-matmul-1
branch
3 times, most recently
from
5b2d42b to
990069a
Compare
| out_matmul_scale = out_matmul_scale.data.view(torch.uint8) | ||
| if has_scratchpad and "mx_out_scale" in memory["scratchpad"]: | ||
| out_matmul_scale = memory["scratchpad"]["mx_out_scale"] | ||
| out_matmul_has_mx = out_matmul_scale is not None and out_matmul.element_size() == 1 |
There was a problem hiding this comment.
What is the reason for changing this line of code. This currently matches the upstream version of this file.
There was a problem hiding this comment.
Because, in my opinion, it is a logical mistake. In this line (https://github.com/intel/intel-xpu-backend-for-triton/blob/c2a39f4396fcd5297a57b9326ffd788a29e82ef2/python/triton_kernels/triton_kernels/matmul_ogs.py#L294C10-L294C25 ), it always initializes with float32 if split_k > 2, and then element_size == 4. I think element_size is not the most suitable condition. It would make more sense to check if out_dtype equals 8 bits. Maybe it’s a bit confusing because I didn’t update the skiplist or remove the tests that were already fixed by these changes.
There was a problem hiding this comment.
I updated the skiplist. Also, all batched test cases were fixed in this Triton commit when we reland this commit all batched test cases should be resolved. However, there are still 16 test cases that need to be fixed.
There was a problem hiding this comment.
can you please upstream the change?
There was a problem hiding this comment.
I added pr to upstream triton-lang/triton#8519
Signed-off-by: Witold Dziurdz <[email protected]>
wdziurdz
force-pushed
the
dev/wdziurdz/test-matmul-1
branch
from
990069a to
f51fe27
Compare
3 participants
Fixes #5074