[ROCm] Increase hipBLASLt default workspace size #139300
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/139300
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 8f5e23e with merge base ad637a4 ( This comment was automatically generated by Dr. CI and updates every 15 minutes. |
naromero77amd
changed the title
naromero77amd
force-pushed
the
increase_hipblaslt_default_workspace_size
branch
from
74cb4ce to
8f5e23e
Compare
jeffdaily
changed the title
eqy
left a comment
eqy
left a comment
There was a problem hiding this comment.
Approving but note that historically this setting is sticky as it would break e.g., Meta internal workloads that are on the edge of memory usage.
Going from 32MB to 76MB for the tunableop case is too much? For the non-tunable path, the 8MB default was too small for adequate performance. Should have also been at least 32MB. 76MB is necessary for highest-performing algorithms. |
|
We've had the 31MiB increase (1 -> 32 MiB) get reverted on the cuBLASLt side a few times, see e.g., #120925 |
Is there some Meta workload that can be run before this is merged in? I was planning on merging in once the CI checks out. |
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
This PR increases hipBLASLt default workspace size to 76 MB which is the recommended default. This PR does not contain any bug fixes. Pull Request resolved: pytorch#139300 Approved by: https://github.com/jeffdaily, https://github.com/eqy
5 participants
This PR increases hipBLASLt default workspace size to 76 MB which is the recommended default. This PR does not contain any bug fixes.
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang