[https://nvbugs/5344910][fix] Corrected memory position when setting buffers to 0 in standalone_stable_radix_topk_ (NVIDIA#6712)

stnie · web-flow · commit 40f773658e4c · 2025-08-08T15:25:59.000+02:00
Signed-off-by: Stefan Niebler &lt;82932102+stnie@users.noreply.github.com&gt;
diff --git a/cpp/tensorrt_llm/kernels/topkLastDim.cu b/cpp/tensorrt_llm/kernels/topkLastDim.cu
@@ -1275,8 +1275,8 @@ void standalone_stable_radix_topk_(void* buf, size_t& buf_size, T const* in, Idx
             sort_in = static_cast<decltype(sort_in)>(aligned_pointers[9]);
             sort_in_idx = static_cast<decltype(sort_in_idx)>(aligned_pointers[10]);
         }
-        cudaMemsetAsync(
-            buf, 0, static_cast<char*>(aligned_pointers[2]) - static_cast<char*>(aligned_pointers[0]), stream);
+        cudaMemsetAsync(aligned_pointers[0], 0,
+            static_cast<char*>(aligned_pointers[2]) - static_cast<char*>(aligned_pointers[0]), stream);
     }
 
     T const* in_buf = nullptr;

Original file line number	Diff line number	Diff line change
`@@ -1275,8 +1275,8 @@ void standalone_stable_radix_topk_(void* buf, size_t& buf_size, T const* in, Idx`
`1275`	`1275`	`sort_in = static_cast<decltype(sort_in)>(aligned_pointers[9]);`
`1276`	`1276`	`sort_in_idx = static_cast<decltype(sort_in_idx)>(aligned_pointers[10]);`
`1277`	`1277`	`}`
`1278`		`- cudaMemsetAsync(`
`1279`		`- buf, 0, static_cast<char>(aligned_pointers[2]) - static_cast<char>(aligned_pointers[0]), stream);`
	`1278`	`+ cudaMemsetAsync(aligned_pointers[0], 0,`
	`1279`	`+ static_cast<char>(aligned_pointers[2]) - static_cast<char>(aligned_pointers[0]), stream);`
`1280`	`1280`	`}`
`1281`	`1281`
`1282`	`1282`	`T const* in_buf = nullptr;`