Add check for INVALID_ARGUMENT in NvLink checks

Unfortunately, the check to see if a link is active throws an error if
an invalid linkID is passed in (instead of simply saying that the link
is inactive). This causes problems since the newest nvml.h is for CUDA
11 (which has an NVML_NVLINK_MAX_LINKS of 12) and older versions had an
NVML_NVLINK_MAX_LINKS of 6.

This patch adds a check to see if the various calls that take a linkID
fail with INVALID_ARGUMENT, and if so, silently ignore the error. This
hould be OK since we are fairly confident all other arguments are valid.
It would have been nice to avoid this (somewhat hacky) solution though.

Signed-off-by: Kevin Klues <[email protected]>

bindings/go/nvml/bindings.go

@@ -319,7 +319,7 @@ func (h handle) deviceGetNvLinkState(link uint) (*uint, error) {
 	var isActive C.nvmlEnableState_t
 
 	r := C.nvmlDeviceGetNvLinkState(h.dev, C.uint(link), &isActive)
-	if r == C.NVML_ERROR_NOT_SUPPORTED {
+	if r == C.NVML_ERROR_NOT_SUPPORTED || r == C.NVML_ERROR_INVALID_ARGUMENT {
 		return nil, nil
 	}
 
@@ -330,7 +330,7 @@ func (h handle) deviceGetNvLinkRemotePciInfo(link uint) (*string, error) {
 	var pci C.nvmlPciInfo_t
 
 	r := C.nvmlDeviceGetNvLinkRemotePciInfo(h.dev, C.uint(link), &pci)
-	if r == C.NVML_ERROR_NOT_SUPPORTED {
+	if r == C.NVML_ERROR_NOT_SUPPORTED || r == C.NVML_ERROR_INVALID_ARGUMENT {
 		return nil, nil
 	}