fix(prism): rent 1x B200 on Lium 8x split hosts

[echobt]

Summary

• Prod no_capacity was a false sold-out: Lium only lists 8× B200, and the 1× pin rejected every row.
• Parse available_gpu_count / min_gpu_count_for_rental and rent one GPU (including idle hosts that omit min).
• Never buy the whole 8-pack; still never fall through to 5090 / 6000.

Test plan

• cargo test -p prism-lium-types -p prism-lium
• Replay live /executors → 4 B200 hosts become 1× candidates
• CI green on main
• Prod prism-challenge no longer floods no_capacity when B200s are listed
• A queued miner job reaches provisioning / rent with gpu_count=1

Summary by CodeRabbit

• New Features
  • Supports renting fewer GPUs from eligible split-capable hosts, including individual GPUs from larger B200 instances.
  • Marketplace offers now include rental minimums and available GPU counts for more accurate matching.
• Bug Fixes
  • Improved GPU offer selection while preventing unintended whole-host rentals or fallback to unsuitable GPU types.
  • Preserved existing whole-host rental behavior for non-splittable offers.
• Documentation
  • Updated GPU pinning guidance and refreshed the Lium marketplace inventory snapshot and pricing details.

[coderabbitai]

Caution

Review failed

The pull request is closed.

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro Plus

Run ID: 27efcca8-225c-4291-83da-d165b24df8fe

📥 Commits

Reviewing files that changed from the base of the PR and between 286f5e5 and b1208f7.

📒 Files selected for processing (6)

• crates/prism-lium-types/src/json.rs
• crates/prism-lium-types/src/types.rs
• crates/prism-lium/src/client.rs
• crates/prism-lium/src/sim.rs
• crates/prism-verda/src/lib.rs
• docs/PRISM.md

---

📝 Walkthrough

Walkthrough

Lium offers now parse rental metadata and support split GPU matching. Provisioning uses the calculated rental count, including one-GPU rentals from eligible 8× B200 hosts. Provider defaults, tests, and pin guidance were updated.

Changes

Split GPU rental

Layer / File(s)	Summary
Offer metadata and parsing crates/prism-lium-types/src/json.rs, crates/prism-lium-types/src/types.rs, crates/prism-lium/src/sim.rs, crates/prism-verda/src/lib.rs	Offer now stores optional rental minimum and available GPU counts. JSON parsing accepts valid bounded values. Provider constructors use Offer::default() for unspecified fields.
Split matching and rent calculation crates/prism-lium-types/src/types.rs	Offer now validates split rentals, calculates rent_count, and preserves whole-host rejection rules. Tests cover native, split-capable, idle, and minimum-four-GPU 8× offers.
Provisioning integration and pin guidance crates/prism-lium/src/client.rs, docs/PRISM.md	Provisioning uses selected.rent_count(spec.gpu_count). Tests cover one-GPU rental from an idle 8× B200 host. Documentation updates the dense GPU pin guidance and inventory snapshot.

Estimated code review effort: 4 (Complex) | ~45 minutes

Sequence Diagram(s)

sequenceDiagram
  participant LiumMarketplace
  participant Offer
  participant LiumClient
  participant RentalAPI
  LiumMarketplace->>Offer: provide GPU rental metadata
  LiumClient->>Offer: match requested GPU count
  Offer-->>LiumClient: return eligible offer and rent_count
  LiumClient->>RentalAPI: provision with rent_count
  RentalAPI-->>LiumClient: return provisioned pod

Loading

✨ Finishing Touches

📝 Generate docstrings

• Create stacked PR
• Commit on current branch

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix/lium-b200-split-rent

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}