Conversation
f72924c to
523d0ed
Compare
0521d1b to
77cc7c4
Compare
We use --exit-code instead of --quite to show the diffs causing the check-vendor make target to fail. Signed-off-by: Evan Lezar <elezar@nvidia.com>
Signed-off-by: Evan Lezar <elezar@nvidia.com>
Signed-off-by: Evan Lezar <elezar@nvidia.com>
77cc7c4 to
daf5eb7
Compare
| return "Hopper", nil | ||
| case nvml.DEVICE_ARCH_BLACKWELL: | ||
| return "Blackwell", nil | ||
| case nvml.DEVICE_ARCH_T23X: |
There was a problem hiding this comment.
This is required since the DEVICE_ARCH_T23X definition was removed from nvml.h as part of the CUDA 13.0 release.
ArangoGutierrez
left a comment
There was a problem hiding this comment.
only a nit comment
lgtm
pkg/nvlib/device/device.go
Outdated
| } | ||
|
|
||
| // IsCoherent returns whether the device is capable of coherent access to cpu | ||
| // and gpu memory. |
There was a problem hiding this comment.
nit: GPU in upper case
|
|
||
| check-vendor: vendor | ||
| git diff --quiet HEAD -- go.mod go.sum vendor | ||
| git diff --exit-code HEAD -- go.mod go.sum vendor |
There was a problem hiding this comment.
for the curious reader: exits with non-zero code if there are differences
(was this a bug?)
There was a problem hiding this comment.
I had some vendoring mismatches that CI fails on. Switching to --exit-code allows the diff to actually be displayed in CI so that it's easier to see what caused the failure without requiring that one try to reproduce it locally.
jgehrcke
left a comment
There was a problem hiding this comment.
Take this with a grain of salt but LGTM
This change adds an IsCoherent function to the Device type that can be used to check wither a device has coherent access to system memory based on the supported addressing mode. Note that this requires a CUDA 13.0 driver. Signed-off-by: Evan Lezar <elezar@nvidia.com>
daf5eb7 to
a308b4a
Compare
This change adds an
IsCoherentAPI to theDeviceinterface. This returns whether the selected device has coherent access to system memory and is determined by the NVML Addressing mode.This is blocked by:
go-nvmlto a version that includes the new APIs.