[pditommaso]

Summary

Adds client support for the cloud-neutral GPU data that cloudinfo will serve once seqeralabs/cloudinfo#130 and seqeralabs/cloudinfo#132 are merged, expected in cloudinfo 0.30.0. With it, a consumer such as sched can ask for the same GPU on every cloud and filter by CUDA support or memory, without its own per-provider tables.

Changes

• CloudProduct, new fields. All are null on older backends and wherever cloudinfo omits them (CPU-only, TPU, Inferentia and Trainium types, unmapped GPUs). They're included in equals, hashCode and toString.

  Field	Example	Meaning
  gpuModel	nvidia-a10g	Canonical GPU model ID, the same on every provider
  gpuVendor	nvidia	nvidia or amd
  gpuCapability	8.6	NVIDIA CUDA compute capability
  gpuArch	ampere	Architecture of the capability
  gpuCudaMin / gpuCudaMax	11.1 / null	CUDA toolkit releases that can target the GPU; gpuCudaMax is null while still supported
  gpuMemoryGB	24	Memory of one whole GPU, as the vendor states it

  The CUDA fields are null for AMD GPUs and for parts whose chip isn't disclosed. gpuMemoryMiB keeps its meaning (total memory for the instance); its docs now note that cloudinfo derives it for Azure.

• ProductsQuery, new GPU filters, combined with the existing ones (a product must match all):

  • gpuModels(List<String>), sent as ?gpuModel=a,b, matches any of the given IDs.
  • minGpuCapability(String), sent as ?minGpuCapability=8.0.
  • cuda(String), sent as ?cuda=12.4.
  • minGpuMemoryGB(Long), sent as ?minGpuMemoryGB=40. A fractional GPU counts only its share of the card.

  Values are URL-encoded. The client doesn't validate them, because cloudinfo answers 400 to a malformed value. Older backends ignore the filters.

• Docs: README section "GPU Model, Capability and Memory", changelog entry, and VERSION bumped to 2.3.0.

Notes for review

• No [release] marker yet. It's left out on purpose, so merging doesn't publish 2.3.0 before cloudinfo 0.30.0 is deployed. Add the marker, or release separately, once it's out.
• 0.30.0 is an assumption: it's the next minor after the current 0.29.0. The docs need adjusting if the release is numbered differently.
• Field and parameter names follow cloudinfo feat(lib-data-range-redis): add size() (ZCARD) for queue-depth metrics #130 and chore: update dependency java-jdk to v27 #132 exactly. A rename there before merge needs a matching change here.

Testing

• ./gradlew :lib-cloudinfo:test: 121 tests, 0 failures.
• New CloudProductGpuFieldsTest: its test data (products-gpu-models.json) is real prod products from all four providers, fetched on 23 Oct 2026, with the new fields set to the values the cloudinfo branch computes. It covers:
  • A10G, with the AWS-reported 22 GiB kept separate from gpuMemoryGB: 24
  • K80, with both CUDA bounds and no memory
  • a ⅙ A10 on Azure, with derived gpuMemoryMiB: 4096
  • H100 on GCP
  • AMD MI300X and Alibaba GPU-H, with no CUDA fields
  • a CPU-only type, and a payload from an older backend
• Other tests: CloudInfoClientMockTest checks how the four filters appear in the request URL, including encoding. ProductsQueryTest covers equality and that the gpuModels list is copied.

🤖 Generated with Claude Code