-
Notifications
You must be signed in to change notification settings - Fork 74.1k
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
Change Redzone space limit for XLA GPU #54860
Conversation
c5548ae
to
86fbb3a
Compare
Hi @kaixih , could you provide some more context on what is the desired goal? |
@cheshire Sure. Basically, we found that the max space limit of redzone allocator for the XLA conv is set to be 4GB, which is insufficient for some models that expect large input/output tensors. In addition, we also noticed that this limit is not adjustable during runtime. So, compared to the XLA gemm, which has already set the limit of the input/output redzone allocator to the numeric max of int, we think it might be appropriate to follow it for the XLA conv. Moreover, we introduced a new env var to control the scratch redzone allocator max limit as well in case it needs to be adjusted. |
86fbb3a
to
2886d6d
Compare
@cheshire Can you help check what blocks the merging? It seems some "Google internal checks" failed. Thanks. |
PiperOrigin-RevId: 434441627
There was a merge conflict in xla.proto, you used the same tag as was already used in a recent change. I fixed that and got your PR merged. |
Seems auto-merge is not happening but the changes are merged into master now, so we can close this. Thank you for the PR. |
This PR changes how the redzone space limit is set in the XLA gpu conv algorithm picker.
XLA_FLAGS=--xla_gpu_redzone_scratch_max_megabytes=6144
.cc. @nluehr