Skip to content

Reapply "[Dispatch Creation] Rework dispatch formation logic (#21854)"#22065

Merged
IanWood1 merged 2 commits into
iree-org:mainfrom
IanWood1:fix_dispatch_creation_rework
Oct 10, 2025
Merged

Reapply "[Dispatch Creation] Rework dispatch formation logic (#21854)"#22065
IanWood1 merged 2 commits into
iree-org:mainfrom
IanWood1:fix_dispatch_creation_rework

Conversation

@IanWood1
Copy link
Copy Markdown
Member

@IanWood1 IanWood1 commented Sep 23, 2025

Re-land dispatch creation changes with an additional check to make sure that the number of parallel/reduction iterators match (only when fusing reductions together) and that they are not permuted. This fixes the issues in #22053 and #22055 that caused the revert.

…g#21854)" (iree-org#22058)

This reverts commit 087d5b9.

Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
@IanWood1 IanWood1 force-pushed the fix_dispatch_creation_rework branch 2 times, most recently from c152295 to 5e5113a Compare September 25, 2025 21:58
@IanWood1 IanWood1 marked this pull request as ready for review September 25, 2025 21:59
Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
@MaheshRavishankar
Copy link
Copy Markdown
Collaborator

Does this fix the two issues that were linked with the original issue ? #22053 and #22055 I think.

@IanWood1
Copy link
Copy Markdown
Member Author

Does this fix the two issues that were linked with the original issue ? #22053 and #22055 I think.

Yes

@MaheshRavishankar
Copy link
Copy Markdown
Collaborator

Yes

Should add that to the commit message.

hanhanW
hanhanW previously approved these changes Sep 26, 2025
@IanWood1
Copy link
Copy Markdown
Member Author

IanWood1 commented Oct 1, 2025

This causes correctness issues with llama-70b-fp8 which need to be resolved before landing this.

@hanhanW hanhanW dismissed their stale review October 2, 2025 22:53

Remove my approval, since I'll be out and the codebase is not ready for this yet.

@IanWood1
Copy link
Copy Markdown
Member Author

IanWood1 commented Oct 6, 2025

The correctness issues are resolved with #22223. Also, the runtime benchmarks were about equal when comparing with/without this change (minor prefill improvement) and there was a slight decrease to the number of dispatches.

This should be good to merge after review + #22223

@IanWood1 IanWood1 merged commit 941241b into iree-org:main Oct 10, 2025
67 of 69 checks passed
weidel-p pushed a commit to weidel-p/iree that referenced this pull request Oct 21, 2025
…g#21854)" (iree-org#22065)

Re-land dispatch creation changes with an additional check to make sure
that the number of parallel/reduction iterators match (only when fusing
reductions together) and that they are not permuted. This fixes the
issues in iree-org#22053 and iree-org#22055 that caused the revert.

---------

Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
Signed-off-by: Philipp <philipp.weidel@intel.com>
pstarkcdpr pushed a commit to pstarkcdpr/iree that referenced this pull request Nov 28, 2025
…g#21854)" (iree-org#22065)

Re-land dispatch creation changes with an additional check to make sure
that the number of parallel/reduction iterators match (only when fusing
reductions together) and that they are not permuted. This fixes the
issues in iree-org#22053 and iree-org#22055 that caused the revert.

---------

Signed-off-by: Ian Wood <ianwood@u.northwestern.edu>
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.

3 participants