Add approved G4 GPU qualification and eight-device accounting - #241
Conversation
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_a03d70d2-137d-492e-a2d2-956678e5302a) |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_94e21d75-0eac-44f1-86ef-e1b69785a0e5) |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 961671bba5
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| if time.monotonic() >= deadline: | ||
| raise IndependentLiveError("GPU round deadline exceeded before complete deduplication") |
There was a problem hiding this comment.
When G4 qualification encounters two unreachable primary axons, the sequential fleet probes can each consume the configured 150-second timeout and exhaust the shared 300-second deadline. On the next miner, this check raises outside the per-miner exception handler, so main() discards every previously collected provider and publishes the total qualification_unavailable document instead of the intended partial directory. Record the remaining miners as unavailable or return the accumulated round when the deadline is reached.
Useful? React with 👍 / 👎.
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_2eb18a54-09cc-4710-b6f4-c9080145afda) |
1 participant
Add
cathedral-gpu-qualifyto discover GPU fleets, verify signed admission and completed CUDA work, and produce a local score plan plus a sanitized public directory. The command cannot construct or submit chain transactions.The first approved profile requires eight separate Google Spot G4 instances under one miner hotkey, each with one RTX PRO 6000 GPU and distinct instance, GPU, TLS and worker-signing identities. Configured operator roots authorize the guest and its reports; AMD SEV CPU attestation remains unverified. Incomplete fleets and duplicate claimants receive zero local test credit. Public G4 eligibility stays false.
Eight bounded concurrent requests fit the local NVIDIA verification/work deadlines. Partial discovery preserves known providers; preflight failure publishes explicit unavailability. Signed validator requests and existing CPU behavior remain intact.
The separate
gpuextra pins sandbox commit08438dd842d25c901bd4a908f306390421cc7052, the reviewed head merged in cathedralai/cathedral-sandbox#202 (identical tree to merge6689d00c5242925752b5f5ebb26e35a4d44094aa). Existing SNP/CPU pins are unchanged; GPU qualification installs into its own environment.Validation:
scripts/accept_gpu_prelaunch.shpassed against that exact installed dependency with PEP 610 provenance checked and no sandbox PYTHONPATH: 42 tests, CLI help and scoped Ruff passed. Tests cover signed real TLS transport, bindings, eight-device accounting, shared-key rejection and directory behavior with synthetic hardware evidence. The independent identity review and its one repair verification passed.No hardware was rented or qualified. Private customer work, GPU economic policy and live rewards are not activated. Part of cathedralai/cathedral-sandbox#73.
Note
Medium Risk
New validator-signed worker access and prelaunch scoring logic affect how miners are qualified and listed, though outputs are explicitly non-writing and gated by strict config and deduplication rules.
Overview
Adds
cathedral-gpu-qualify, a no-chain-write path that reads a finalized metagraph, discovers signed worker fleets, runs GPU admission plus fixed CUDA work verification, deduplicates claims, and emits a local weight plan (chain_write: false) and a sanitized public provider directory.Native TDX+GPU mode uses the signed registry and production verifiers; G4 approved-operator mode uses
cathedral_gpu_g4_prelaunch_v1with configured operator roots, requires eight distinct single-GPU VMs per miner (instance, GPU, TLS, worker signing key), and zeros incomplete or duplicate bundles. An external request signer is limited to GPU HTTP access documents only.Supporting changes extend signed validator requests and HTTPS response bounds to
/v1/gpu-capabilities,/v1/gpu-evidence, and/v1/gpu-work, with network/netuid on signed headers. A separategpuoptional dependency pins a distinct sandbox commit; docs, acceptance script, and tests cover protocol, real TLS wire, and G4 bundle rules.Reviewed by Cursor Bugbot for commit b1c359e. Bugbot is set up for automated code reviews on this repo. Configure here.