Repository navigation
[cuda] Score autotune picks with one event tick of tolerance, on longer kernels - #23635
Merged
Merged
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/23635
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 0dadcb0 with merge base 7f4dac6 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
This PR needs a
|
Gasoonjia
force-pushed
the
gasoonjia/autotune-regret-timer-resolution
branch
from
October 9, 2026 23:48
95031ed to
6d128a8
Compare
Gasoonjia
changed the title
Gasoonjia
force-pushed
the
gasoonjia/autotune-regret-timer-resolution
branch
from
October 9, 2026 23:50
6d128a8 to
febf90e
Compare
Gasoonjia
added a commit
that referenced
this pull request
Oct 12, 2026
unittest-cuda is cancelled at its 90-minute timeout on every recent main commit. #23605 moved the quantized GEMM kernel suites into unittest-cuda-quantized-gemm, but the same push added test_int4_large_quantized_gemm.py (~7.7 min on the A10G) and test_int6_large_quantized_gemm.py (~5.1 min), which still run in unittest-cuda. On #23635 its backends/cuda/tests pytest alone took 62 min and the job was cancelled during the Gemma 4 tests that follow. Run both in the second quantized-GEMM shard, renamed large-int5-int6-int8-graph-timing (that shard took 44 min; the int4 shard already takes ~60). unittest-cuda ignores them. No test is dropped.
Gasoonjia
added a commit
that referenced
this pull request
Oct 12, 2026
#23676) `unittest-cuda` is cancelled at its 90-minute timeout on every recent main commit (e.g. https://github.com/pytorch/executorch/actions/runs/37991426749/job/114026586806). #23605 moved the quantized GEMM kernel suites into `unittest-cuda-quantized-gemm`, but the same push also added two new suites that still run in `unittest-cuda`: - `test_int4_large_quantized_gemm.py` (`711dd81e23`, ~7.7 min on the A10G) - `test_int6_large_quantized_gemm.py` (`f19360763f`, ~5.1 min) On #23635 (https://github.com/pytorch/executorch/actions/runs/38006402026/job/114076255669), `unittest-cuda`'s `backends/cuda/tests` pytest alone took 62 min, and the job was cancelled during the Gemma 4 tests that follow. ### Change - Run both large suites in the second quantized-GEMM shard, renamed `large-int5-int6-int8-graph-timing`. That shard took 44 min on #23635; the int4 shard already takes ~60 min, so they don't go there. - `unittest-cuda` ignores them. - No test is dropped. ### Test plan CI: - `unittest-cuda` finishes within 90 min. - `unittest-cuda-quantized-gemm (large-int5-int6-int8-graph-timing)` runs both large suites and also stays within 90 min.
Gasoonjia
marked this pull request as ready for review
October 13, 2026 00:58
rascani
approved these changes
Oct 13, 2026
…er kernels SaturatedCpuExportTest and Int4AutotunePickTest score each autotune decision by the regret of the pick: its reference time over the best candidate's. Event timing is quantized (cudaEventElapsedTime steps by 0.512 us on the A10G CI GPU, hipEventElapsedTime by about 1 us), and a reading is the difference of two quantized timestamps, so it can fall either side of the true time by about a tick. On a kernel of a few microseconds that is a large ratio. The A10G unittest-cuda job failed on AssertionError: 1.2000000727595779 not less than or equal to 1.2 : Pick(kind='inductor', name='triton_poi_fused__unsafe_view_add_transpose_ copy_view_copy_9', candidates=2, regret=1.2000000727595779) which is exactly fp32(6 * 0.512 us) / fp32(5 * 0.512 us). Two changes: - _regret() credits the best candidate one tick: pick / (best + 1 tick), floored at 1.0, with margins for the float32 event times. A difference the timer cannot resolve no longer counts as a worse pick; anything larger is charged in proportion, so the limits stay meaningful at every scale. - The tests' kernels run longer, so a tick is a small fraction of them: SaturatedCpuExportTest's model uses widths (512, 1536, 3072) and 4096 tokens (was (256, 768, 1536) and 64), and Int4AutotunePickTest's smallest weight is 8192x2048 (was 2048x2048). summarize() now also prints the fastest kernel's reference time, so CI shows how far above a tick it is. RegretTest checks, on CPU, for both the CUDA and the HIP tick: one tick apart is 1.0 (up to 31250-tick kernels), two ticks apart is charged (up to ~1 s), and pick / (best + 1) beyond that.
Gasoonjia
force-pushed
the
gasoonjia/autotune-regret-timer-resolution
branch
from
October 13, 2026 05:26
e290977 to
0dadcb0
Compare
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
SaturatedCpuExportTest(test_cuda_graph_timing.py) andInt4AutotunePickTestscore each autotune decision by its regret: the pick's reference time divided by the best candidate's. Event timing is quantized:cudaEventElapsedTimesteps by 0.512 µs on the A10G CI GPU, andhipEventElapsedTimeby about 1 µs. A reading is the difference of two quantized timestamps, so it can fall on either side of the true time by about a tick. On a kernel of a few microseconds, that is a large ratio. The A10G job failed on (https://github.com/pytorch/executorch/actions/runs/37835437673/job/113511485001):That is exactly
fp32(6 × 0.512 µs) / fp32(5 × 0.512 µs).Changes
_regret()credits the best candidate one tick,pick / (best + 1 tick), floored at 1.0, with margins for the float32 event times.SaturatedCpuExportTest's model now uses widths (512, 1536, 3072) and 4096 tokens (was (256, 768, 1536) and 64).Int4AutotunePickTest's smallest weight is now 8192×2048 (was 2048×2048).summarize()now also prints the fastest kernel's reference time, so CI shows how far above one tick the shortest kernel runs.Tests
RegretTestruns on CPU and covers both the CUDA and the HIP tick:pick / (best + 1).Test plan
CI:
unittest-cuda/unittest-cuda-quantized-gemmshould showRegretTest,SaturatedCpuExportTestandInt4AutotunePickTestpassing, the "fastest kernel" line in their output, and how long the two pick tests now take.