[placement groups] fix gpu ids for bundles

[richardliaw]

Why are these changes needed?

Closes #14542

Related issue number

Checks

• I've run scripts/format.sh to lint the changes in this PR.
• I've included any doc changes needed for https://docs.ray.io/en/master/.
• I've made sure the tests are passing. Note that there might be a few flaky tests, see the recent failures at https://flakey-tests.ray.io/
• Testing Strategy
  • Unit tests
  • Release tests
  • This PR is not tested :(

[tgaddair]

@richardliaw, kind of an edge case, but the order of the devices in CUDA_VISIBLE_DEVICES actually does have some implications. Namely, since CUDA assigns IDs based on the fastest device first (see CUDA_DEVICE_ORDER), you usually want to use device 0 first when available.

One alternative might be to use an order-preserving dedup instead: https://stackoverflow.com/questions/480214/how-do-you-remove-duplicates-from-a-list-whilst-preserving-order/39835527#39835527

[richardliaw]

hmm, what if we just sort the returned list at 434? I think Ray doesn't have any ordering guarantees from the detected CUDA devices, though maybe we can support that later

[tgaddair]

Yeah, that works! May want to double check if the ids here are strings or ints for that, though (would want to sort based on numerical value for strings).