fix: stabilize Gemma4 MoE inference — dynamic attention mask slicing

solderzzc · solderzzc · commit 32ce0e25bb26 · 2026-04-05T10:41:20.000-07:00
- Update mlx-swift-lm submodule with safeMask fix for broadcast_shapes
  crash during sliding window attention on long prompts (1370+ tokens)
- Add tmp/ to .gitignore

HomeSec-Bench: 87/96 (90%) — zero server crashes across full 96-test suite
diff --git a/.gitignore b/.gitignore
@@ -20,3 +20,4 @@ DerivedData/
 *.pid
 curl_out.txt
 sample.txt
+tmp/
diff --git a/mlx-swift-lm b/mlx-swift-lm
@@ -1 +1 @@
-Subproject commit 9a9575be3acdb5f4c9f74629a680b103981f030c
+Subproject commit e48243f5b2a12701650ae930ed4adc8bfc0df38a
