Repository navigation
[inductor] Clamp explicit mix-order split size to a power of two on npu - #201600
zhoushaowen69 wants to merge 1 commit into
Conversation
pytorch-bot
Bot
added
ciflow/inductor
ciflow/torchtitan
|
The following ciflow label(s) have been added but CI has not been triggered yet because the workflows are awaiting approval:
Once a maintainer approves the workflows (scroll to the bottom of the PR page), the corresponding CI jobs will be triggered automatically. Please ping one of the reviewers if you do not have access to approve and run workflows. |
1 similar comment
|
The following ciflow label(s) have been added but CI has not been triggered yet because the workflows are awaiting approval:
Once a maintainer approves the workflows (scroll to the bottom of the PR page), the corresponding CI jobs will be triggered automatically. Please ping one of the reviewers if you do not have access to approve and run workflows. |
|
One or more co-authors of this pull request were not found. You must specify co-authors in commit message trailer via: Supported
Alternatively, if the co-author should not be included, remove the Please update your commit message(s) by doing |
github-actions
Bot
added
bot-triaged
Probed on 910B4 / triton-ascend 3.6 with (40961, 129) bf16 sum(dim=-1)+sum(dim=0): RSPLIT_SIZE values that are powers of two (16/32/64) produce correct results; non-powers of two (18/24/48) miscompile the tail-block mask of the mix-order kernel and over-count 1-3 rows. The upstream heuristic branch already yields a power of two (last_power_of_2, clamped to [16, 128]); only the explicit config.triton.mix_order_reduction_split_size override can produce a non-power-of-two, so clamp it on npu to the largest power of two <= the requested value. Co-Authored-By: Claude Code <noreply@anthropic.com>
5a26acc to
e80e5e4
Compare
This branch was successfully deployed
2 participants
On npu (triton-ascend on Ascend 910B4), a non-power-of-two
config.triton.mix_order_reduction_split_sizemiscompiles the tail-blockmask of the mix-order kernel: with a (40961, 129) bf16
sum(dim=-1)+sum(dim=0)probe, split sizes 18/24/48 over-count 1-3 rows of the tail block (each
column's sum is off by an exact multiple of 0.25), while 16/32/64 are correct.
The heuristic branch below already yields a power of two (
last_power_of_2clamped to [16, 128]); only the explicit config override can inject a
non-power-of-two value. Clamp it on npu to the largest power of two <= the
requested value (18→16, 24→16, 48→32), all within the verified-correct set.
Other devices are untouched.
With this clamp, split sizes 16/18/20/24/32/48/64 all produce bit-correct
results on 910B4, and
test/inductor/test_mix_order_reduction.py::MixOrderReductionNumericTest(passing an explicit split size of 18) runs clean.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo
🤖 Generated with Claude Code