CUDA: fixed peer access toggle synchronization

JohannesGaessler · JohannesGaessler · commit 9474bab5a242 · 2023-12-22T15:21:36.000+01:00
diff --git a/ggml-cuda.cu b/ggml-cuda.cu
@@ -8,6 +8,7 @@
 #include <limits>
 #include <stdint.h>
 #include <stdio.h>
+#include <unistd.h>
 #include <vector>
 
 
@@ -7844,6 +7845,9 @@ static void ggml_cuda_set_peer_access(const int n_tokens) {
             }
         }
     }
+
+    usleep(100000); // 0.1 s
+
 #endif // NDEBUG
 
     peer_access_enabled = enable_peer_access;

Original file line number	Diff line number	Diff line change
`@@ -8,6 +8,7 @@`
`8`	`8`	`#include <limits>`
`9`	`9`	`#include <stdint.h>`
`10`	`10`	`#include <stdio.h>`
	`11`	`+#include <unistd.h>`
`11`	`12`	`#include <vector>`
`12`	`13`
`13`	`14`
`@@ -7844,6 +7845,9 @@ static void ggml_cuda_set_peer_access(const int n_tokens) {`
`7844`	`7845`	`}`
`7845`	`7846`	`}`
`7846`	`7847`	`}`
	`7848`	`+`
	`7849`	`+ usleep(100000); // 0.1 s`
	`7850`	`+`
`7847`	`7851`	`#endif // NDEBUG`
`7848`	`7852`
`7849`	`7853`	`peer_access_enabled = enable_peer_access;`