[Engine] reject duplicate GPUs in automatic replica groups

[j-luwierski]

Problem and result

On the current upstream main, gpu_groups() accepts overlapping GPUs when a numeric replica count derives groups from the configured GPU list. Reproduced before the fix: {"gpu":[0,0],"replicas":2} returns [[0],[0]]; {"gpu":[0,1,1,2],"replicas":2} returns [[0,1],[1,2]]. The server can then start independent replica engines on the same card. After this change, these configurations raise ReplicaConfigError before replica engines start.

Related discussion

PR #2953 was closed during the 0.2.0-beta tracker reset, not declined. The original change was revalidated against current main. Search of open and closed issues and PRs for gpu_groups, duplicate cards and replica GPU grouping found no other fix for this case.

Approach

Build numeric groups and pass them through the same non-empty and duplicate-card validation used for explicitly named groups. Add regression coverage for duplicate GPU IDs in list and comma-separated forms, normalized duplicate IDs, explicit groups, preserved card order, and disabled replica counts. Valid group ordering and the no-replica path remain unchanged. No engine backend or defaults are changed.

Validation

Passed:

• Reproduced the issue on upstream main at e88ee39d1d6be6e3364183af3b39b62a7ed6f211, Linux Debian 13, Python 3.13.5: duplicate GPU configurations returned overlapping groups instead of raising an error.
• uv run --with-requirements requirements.txt python -m unittest serve.test_replicas serve.test_parallel -q, on the change based on e88ee39d1d6be6e3364183af3b39b62a7ed6f211, Linux Debian 13, Python 3.13.5: 64 tests passed in 71.163 seconds. These tests use fake engines; no model or GPU was used.
• git diff --check: passed.

Failed:

• The initial python3 -m unittest serve.test_replicas serve.test_parallel -q attempt could not import jinja2 from the system Python. The same suite then passed when run with the pinned repository requirements in a temporary uv environment.

Not run:

• Multi-GPU inference and CUDA, HIP, SYCL or Metal builds; this change touches only Python server code and was verified with fake engines.

Limits

This validates configuration handling and fake-engine server behavior. It does not establish GPU runtime behavior or performance on hardware.