Qualcomm AI Engine Direct - fix sliding attention update bug

haowhsu-quic · web-flow · commit 5fd66ee788a6 · 2025-09-18T16:27:46.000-07:00
Differential Revision: D82745889 Pull Request resolved: pytorch#14411
diff --git a/examples/qualcomm/oss_scripts/llama/runner/kv_manager.cpp b/examples/qualcomm/oss_scripts/llama/runner/kv_manager.cpp
@@ -242,9 +242,8 @@ void KVManager<T>::update_attention_mask(
         std::fill_n(
             cur_ptr, std::abs(n_past + ar_len) - avalible_cache_len, neg_val);
       }
-
-      cur_ptr += metadata_.context_len;
     }
+    cur_ptr += metadata_.context_len;
   }
 }
 

Original file line number	Diff line number	Diff line change
`@@ -242,9 +242,8 @@ void KVManager<T>::update_attention_mask(`
`242`	`242`	`std::fill_n(`
`243`	`243`	`cur_ptr, std::abs(n_past + ar_len) - avalible_cache_len, neg_val);`
`244`	`244`	`}`
`245`		`-`
`246`		`- cur_ptr += metadata_.context_len;`
`247`	`245`	`}`
	`246`	`+ cur_ptr += metadata_.context_len;`
`248`	`247`	`}`
`249`	`248`	`}`
`250`	`249`