Adds PyTorch patch to implement bf16 GELU via LUT.

nSircombe · nSircombe · commit 8e2511dbbf72 · 2025-08-11T10:55:04.000+01:00
Gives ~8x speedups.
diff --git a/ML-Frameworks/pytorch-aarch64/CHANGELOG.md b/ML-Frameworks/pytorch-aarch64/CHANGELOG.md
@@ -8,6 +8,8 @@ where `YY` is the year, and `MM` the month of the increment.
 ## [unreleased]
 
 ### Added
+- Adds https://github.com/pytorch/pytorch/pull/159859, a WIP LUT implmentation of bf15 GELU
+  ~8x speedup over existing oneDNN implementation
 
 ### Changed
 - Updates hashes for:
diff --git a/ML-Frameworks/pytorch-aarch64/get-source.sh b/ML-Frameworks/pytorch-aarch64/get-source.sh
@@ -39,6 +39,9 @@ git-shallow-clone https://github.com/pytorch/pytorch.git $PYTORCH_HASH
     # https://github.com/pytorch/pytorch/pull/160184 - Draft: separate reqs for manywheel build and pin
     apply-github-patch pytorch/pytorch 9a8b0df99eac62e7ec6199dd0223a80d26e2dee0
 
+    # https://github.com/pytorch/pytorch/pull/159859 - PoC LUT optimisation for GELU bf16 operators
+    apply-github-patch pytorch/pytorch 51626269d3730df1a6b465fa0191074fc31f7c29
+
     git submodule sync
     git submodule update --init --checkout --force --recursive --jobs=$(nproc)
     (

Original file line number	Diff line number	Diff line change
`@@ -39,6 +39,9 @@ git-shallow-clone https://github.com/pytorch/pytorch.git $PYTORCH_HASH`
`39`	`39`	`# https://github.com/pytorch/pytorch/pull/160184 - Draft: separate reqs for manywheel build and pin`
`40`	`40`	`apply-github-patch pytorch/pytorch 9a8b0df99eac62e7ec6199dd0223a80d26e2dee0`
`41`	`41`
	`42`	`+ # https://github.com/pytorch/pytorch/pull/159859 - PoC LUT optimisation for GELU bf16 operators`
	`43`	`+ apply-github-patch pytorch/pytorch 51626269d3730df1a6b465fa0191074fc31f7c29`
	`44`	`+`
`42`	`45`	`git submodule sync`
`43`	`46`	`git submodule update --init --checkout --force --recursive --jobs=$(nproc)`
`44`	`47`	`(`