Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
PR #11478: [XLA:GPU] add guards for flash attention graph with cuDNN …
…>= 8.9.4 Imported from GitHub PR openxla/xla#11478 * Build XLA with cuDNN 8.6 will cause compilation error because flash attention graph will use `CudnnfMHAUid` which is only defined with cuDNN > 8.8. * Add a guard for flash attention graph with at least cuDNN 8.9.4. So the current logic would be only compile FMHA with cuDNN > 8.8 and only compile flash attention with cuDNN > 8.9.4. Copybara import of the project: -- a1aa585f4e6ce42c7486336549447151cd5f7690 by cjkkkk <ske@nvidia.com>: add guards for flash attention graph with at least 8.9.4 Merging this change closes #11478 PiperOrigin-RevId: 625425491
- Loading branch information