Guard large unaligned Qwen expert gathers
This commit is contained in:
1 parent
57b8ac28a7
commit
990abb4755
22 files changed
+951
-3
No files matched your search
@@ -27,7 +27,7 @@ fn gather(
|
||||
streams: &Streams,
|
||||
stream: Stream,
|
||||
) -> Result<Array, String> {
|
||||
linear::gather_qmm_bf16(
|
||||
linear::guarded_gather_qmm_bf16(
|
||||
input,
|
||||
[&layer.weight, &layer.scales, &layer.biases],
|
||||
[None, Some(ids)],
|
||||
|
||||
Reference in new issue
Block a user