NVIDIA / holodeck / 22025328511

48%

Build Type

push

github

Committed by web-flow

Commit Message

fix(aws): infer AMI architecture from instance type for arm64 support (#669)

* fix(aws): infer AMI architecture from instance type when unset

When image.architecture is not explicitly set, holodeck defaults to
x86_64 regardless of the instance type. This causes EC2 RunInstances
to fail with "Unsupported configuration" when an arm64-only instance
type (e.g., g5g, m7g, c7g) is paired with an x86_64 AMI.

Add inferArchFromInstanceType() that queries DescribeInstanceTypes to
determine the instance type's supported architectures. When the type
only supports arm64, architecture is automatically set to "arm64"
before AMI resolution. This is wired into all three resolution paths:
resolveOSToAMI (single-node), setLegacyAMI (legacy default), and
createInstances (cluster mode).

Backward compatible: dual-arch or x86_64-only instance types still
default to x86_64. Explicit image.architecture always takes precedence.

Ref: https://github.com/NVIDIA/gpu-driver-container/actions/runs/22012665274/job/63611032634
Signed-off-by: Carlos Eduardo Arango Gutierrez <eduardoa@nvidia.com>

* fix(aws): address copilot review feedback on arch inference

- Normalize *_mac architecture variants (x86_64_mac, arm64_mac) using
  strings.HasPrefix instead of exact match, so Mac instance types like
  mac2-m2.metal are correctly classified
- Surface inference errors in resolveOSToAMI, setLegacyAMI, and
  createInstances instead of silently falling back to x86_64, which
  would mask IAM/API failures and reproduce the original mismatch
- Use synthetic instance type name in dual-arch test case to avoid
  confusion with real AWS instance families
- Add test coverage for arm64_mac and x86_64_mac architecture variants

Signed-off-by: Carlos Eduardo Arango Gutierrez <eduardoa@nvidia.com>

---------

Signed-off-by: Carlos Eduardo Arango Gutierrez <eduardoa@nvidia.com>

Run Details

30 of 54 new or added lines in 2 files covered. (55.56%)

2 existing lines in 1 file now uncovered.

2609 of 5409 relevant lines covered (48.23%)

0.54 hits per line