Merge branch 'transformers' of github.com:0cc4m/exllama into transfor…

…mers
0cc4m · Jun 19, 2023 · ecfea95 · ecfea95
2 parents ac19f97 + a09cf75
commit ecfea95
Showing 1 changed file with 7 additions and 0 deletions.
diff --git a/exllama_ext/cuda_func/column_remap.cu b/exllama_ext/cuda_func/column_remap.cu
@@ -1,7 +1,14 @@
 #include "column_remap.cuh"
 #include "../util.cuh"
 
+// Using 1024 make me crash with "Memory access fault by GPU node-1 (Agent
+// handle: 0x012345678912) on address 0x012345678912. Reason: Page not present
+// or supervisor privilege."
+#if defined(USE_ROCM)
 const int SHUF_BLOCKSIZE_X = 256;
+#else
+const int SHUF_BLOCKSIZE_X = 1024;
+#endif
 const int SHUF_BLOCKSIZE_Y = 16;
 
 __global__ void column_remap_kernel