Make cache writes optional in InferenceCalculatorAdvancedGL

Previously, caches were always written, and an error would cause the graph to close abruptly. This prevented services with read-only access to the cache from using the calculator. The new behavior allows services to choose whether or not to write caches. PiperOrigin-RevId: 561866791
2023-08-31 23:31:27 -07:00
parent dea6ccba25
commit de0c7f2a30
3 changed files with 55 additions and 2 deletions
@@ -122,6 +122,8 @@ describe('TaskRunner', () => {
        allowPrecisionLoss: true,
        cachedKernelPath: undefined,
        serializedModelDir: undefined,
+        cacheWritingBehavior: InferenceCalculatorOptions.Delegate.Gpu
+                                  .CacheWritingBehavior.WRITE_OR_ERROR,
        modelToken: undefined,
        usage: InferenceCalculatorOptions.Delegate.Gpu.InferenceUsage
                   .SUSTAINED_SPEED,