Test fixes for multi-GPU instances.

Author: maleadt

test/nvml.jl

@@ -5,7 +5,10 @@ macro maybe_unsupported(ex)
         try
             $(esc(ex))
         catch err
-            (isa(err, NVMLError) && err.code == NVML.ERROR_NOT_SUPPORTED) || rethrow()
+            # XXX: we get a permission error when trying to access an entire device
+            #      when we've actually been assigned a compute instance (MIG);
+            #      but how do we know if the current device is an instance?
+            (isa(err, NVML.NVMLError) && err.code in [NVML.ERROR_NOT_SUPPORTED, NVML.NVML_ERROR_NO_PERMISSION]) || rethrow()
         end
     end
 end
@@ -25,7 +28,7 @@ end
 
     @test NVML.uuid(nvml_dev) == uuid(cuda_dev)
     NVML.brand(nvml_dev)
-    @test NVML.name(nvml_dev) == name(cuda_dev)
+    @test occursin(NVML.name(nvml_dev), name(cuda_dev)) # entire device vs MIG
     @maybe_unsupported NVML.serial(nvml_dev)
 
     @maybe_unsupported NVML.power_usage(nvml_dev)

test/setup.jl

@@ -87,7 +87,10 @@ function runtests(f, name, time_source=:cuda, snoop=nothing)
                     missing
                 end
             catch err
-                (isa(err, NVML.NVMLError) && err.code == NVML.ERROR_NOT_SUPPORTED) || rethrow()
+                # XXX: we get a permission error when trying to access an entire device
+                #      when we've actually been assigned a compute instance (MIG);
+                #      but how do we know if the current device is an instance?
+                (isa(err, NVML.NVMLError) && err.code in [NVML.ERROR_NOT_SUPPORTED, NVML.NVML_ERROR_NO_PERMISSION]) || rethrow()
                 missing
             end
         else