Commit 2e69c3d
committed
Adds dbias accumulation flag to bwd params
Introduces a toggle to optionally accumulate bias gradients during backward attention.
Enables skipping unnecessary dbias work when unused and provides clearer control for kernels, aiding performance and configurability.1 parent 623b75d commit 2e69c3d
1 file changed
+2
-0
lines changed| Original file line number | Diff line number | Diff line change | |
|---|---|---|---|
| |||
195 | 195 | | |
196 | 196 | | |
197 | 197 | | |
| 198 | + | |
| 199 | + | |
198 | 200 | | |
199 | 201 | | |
200 | 202 | | |
| |||
0 commit comments