Skip to content
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.

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

[Backend] Assert if we hit known-broken case with small #mma shapes. #3549

Merged
merged 1 commit into from
Apr 3, 2024

Conversation

jlebar
Copy link
Collaborator

@jlebar jlebar commented Apr 3, 2024

[Backend] Assert if we hit known-broken case with small #mma shapes.

PR chain

  1. 👉 [Backend] Assert if we hit known-broken case with small #mma shapes. #3549 👈 YOU ARE HERE

@jlebar jlebar marked this pull request as draft April 3, 2024 17:10
@ThomasRaoux
Copy link
Collaborator

To be honest, I strongly dislike having to propagate FailureOr everywhere. I think asserting is the right thing in this kind of cases as it is an internal error.

@jlebar
Copy link
Collaborator Author

jlebar commented Apr 3, 2024

I will make the change you're asking for, but note that unlike other assertions, for this one we don't actually know how to avoid hitting it. That is, we're not asserting an invariant that we believe we maintain throughout the compiler. So it's quite possible that real users will hit this assertion.

@jlebar jlebar force-pushed the dev-jlebar/mma-error-check branch from fba1224 to 7d32f42 Compare April 3, 2024 17:57
@jlebar jlebar changed the title [Backend] Raise an error if we hit known-broken local_load of a small #mma shape. [Backend] Assert if we hit known-broken case with small #mma shapes. Apr 3, 2024
@jlebar jlebar marked this pull request as ready for review April 3, 2024 18:12
@jlebar jlebar requested review from ThomasRaoux and removed request for Superjomn, ptillet and goostavz April 3, 2024 18:12
@jlebar jlebar enabled auto-merge (squash) April 3, 2024 18:12
Copy link
Collaborator

@ThomasRaoux ThomasRaoux left a comment

Choose a reason for hiding this comment

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

LGTM, thanks

@jlebar jlebar merged commit ca09c49 into main Apr 3, 2024
5 checks passed
@jlebar jlebar deleted the dev-jlebar/mma-error-check branch April 3, 2024 18:23
jlebar added a commit that referenced this pull request Apr 3, 2024
PR #3472 partially rolled back PR #3415 due to internal test failures.  This PR
rolls forward the change as much as we currently can, allowing *most* but not
all relevant loads to be pipelined.

There is still a TritonGPU -> LLVM codegen bug in Triton that we have not been
able to fix, but now we catch it with asserts, PR #3549.

GPC: dot-pipelining
jlebar added a commit that referenced this pull request Apr 3, 2024
PR #3472 partially rolled back PR #3415 due to internal test failures.  This PR
rolls forward the change as much as we currently can, allowing *most* but not
all relevant loads to be pipelined.

There is still a TritonGPU -> LLVM codegen bug in Triton that we have not been
able to fix, but now we catch it with asserts, PR #3549.

GPC: dot-pipelining
jlebar added a commit that referenced this pull request Apr 4, 2024
PR #3472 partially rolled back PR #3415 due to internal test failures.  This PR
rolls forward the change as much as we currently can, allowing *most* but not
all relevant loads to be pipelined.

There is still a TritonGPU -> LLVM codegen bug in Triton that we have not been
able to fix, but now we catch it with asserts, PR #3549.

GPC: dot-pipelining
jlebar added a commit that referenced this pull request Apr 4, 2024
PR #3472 partially rolled back PR #3415 due to internal test failures.  This PR
rolls forward the change as much as we currently can, allowing *most* but not
all relevant loads to be pipelined.

There is still a TritonGPU -> LLVM codegen bug in Triton that we have not been
able to fix, but now we catch it with asserts, PR #3549.

GPC: dot-pipelining
jlebar added a commit that referenced this pull request Apr 4, 2024
PR #3472 partially rolled back PR #3415 due to internal test failures.  This PR
rolls forward the change as much as we currently can, allowing *most* but not
all relevant loads to be pipelined.

There is still a TritonGPU -> LLVM codegen bug in Triton that we have not been
able to fix, but now we catch it with asserts, PR #3549.

GPC: dot-pipelining
jlebar added a commit that referenced this pull request Apr 4, 2024
PR #3472 partially rolled back PR #3415 due to internal test failures.
This PR rolls forward the change as much as we currently can, allowing *most*
but not all relevant loads to be pipelined.

There is still a TritonGPU -> LLVM codegen bug in Triton that we have
not been able to fix, but now we catch it with asserts, PR #3549.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

2 participants