Skip to content

SYCL: disable faulty fp16 exp kernel #14395

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 4 commits into from
Jun 29, 2025
+3 −1
Merged

SYCL: disable faulty fp16 exp kernel #14395

Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
ed0aab1
SYCL: disable faulty fp16 CPU exponent for now
qnixsynapse Jun 26, 2025
706ef76
Revert "SYCL: disable faulty fp16 CPU exponent for now"
qnixsynapse Jun 29, 2025
a65866b
SYCL: disable faulty fp16 CPU exponent for now
qnixsynapse Jun 29, 2025
b48a8c5
Fix logic of disabling exponent kernel
qnixsynapse Jun 29, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions ggml/src/ggml-sycl/ggml-sycl.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4216,6 +4216,8 @@ static bool ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, const g
case GGML_UNARY_OP_GELU_ERF:
case GGML_UNARY_OP_TANH:
case GGML_UNARY_OP_EXP:
// Disable FP16 until we find out the root cause of failing fp16 sycl::exp
return ggml_is_contiguous(op->src[0]) && (op->type == op->src[0]->type) && op->src[0]->type == GGML_TYPE_F32;
Copy link
Collaborator

@Rbiessy Rbiessy Jun 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'm having a quick look at this issue but I think the issue comes from the CPU backend. I don't think we should merge this change as part of this PR

Copy link
Collaborator Author

@qnixsynapse qnixsynapse Jun 26, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Is it confirmed? I will revert this change ASAP

I wish there was a way to print the tensor:

[EXP] inf mismatch: SYCL0=inf CPU=1.442383 FAIL

Copy link
Contributor

@Alcpz Alcpz Jun 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Rbiessy continued the conversation in #14317 (comment). Let's see what the maintainers say

Copy link
Collaborator

@Rbiessy Rbiessy Jun 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I'd still prefer to disable fp16 exp in a separate PR but yeah it seems ok to me. I'm hoping we can re-enable it once we switch to oneAPI 2025.2.

On a side note I'll be away tomorrow and early July so don't expect answers :)

Copy link
Collaborator Author

@qnixsynapse qnixsynapse Jun 26, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I see. Please take your time. :)

case GGML_UNARY_OP_SGN:
case GGML_UNARY_OP_ABS:
case GGML_UNARY_OP_ELU:
Expand Down
Loading