Repository navigation
Infernet temporal policies (liquid networks) as c0mpute workloads; v0.2.34 - #61
Merged
Merged
Conversation
…rnet.temporal.infer; v0.2.34 Workers opt in with `c0mpute infernet temporal serve [--train] [--model ...]` and advertise only what `infernet model capabilities` reports (Infernet >= 0.1.66): runtime version, CPU/CUDA/MPS, RAM buckets, served model hashes. Buyers turn requirements into required_capabilities; a new per-workload OfferFilter validates temporal specs and keeps a training worker from bidding while busy (one job per node). - train: one single-node job per family over the existing auction. Spec carries a liquid-agent/v1 TrainJobV1; workers refuse unless privacy.mode=trusted_worker and upload_approved. Buyer serves the dataset from a token-gated transfer server, the worker runs `infernet model train --job`, uploads the artifacts, receipts with the model hash; the buyer verifies weights against the receipt. --worker allowlist (trusted workers), --yes consent after a preview, retries, idempotency reuse. - infer: a session open is auctioned; the winner holds the session's hidden state in `infernet model serve --stdio`. Calls are X25519/HKDF/AES-GCM sealed, answered only for the opener's peer id, exactly-once by call id. - c0mpute infernet temporal serve|train|predict|models; module.toml lists the workloads; DIP 0026. - Tests: unit (spec validation, tag filtering, crypto, transfer server, family selection) and a loopback e2e with a fake infernet covering the allowlist, GPU filtering, receipt-verified training, sticky sessions, a forged peer refused and late replays. Also run by hand end to end with the real Infernet CLI and CfC/GRU runtime. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
ThreatCrush Security Scan9 finding(s) MEDIUM: 9
Snippets are redacted; ThreatCrush never prints matched credential material. |
…ries GHSA-2464-fhr4-w9xw and GHSA-673q-mjm6-r5xr (Next.js unauthenticated RCE), GHSA-cg2g-xh64-v4q7 (DOMPurify). Published after master's last green bun audit; next build verified locally. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Review the following changes in direct dependencies. Learn more about Socket for GitHub.
|
…poral topics; e2e dials the infer worker (no mDNS on CI) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Phase 2 of the Liquid Agent PRDs on the c0mpute side (DIP 0026). Infernet v0.1.66 (InfernetProtocol/infernet-protocol#60, released) does the actual execution. c0mpute only advertises it, dispatches it and carries the receipts. There is no new scheduler and no new marketplace plugin.
c0mpute infernet temporal serve [--train] [--model sha256:..]advertises only whatinfernet model capabilitiesreports: runtime version, CPU/CUDA/MPS, RAM buckets and the model hashes it serves.required_capabilities. A new per-workloadOfferFiltervalidates temporal specs and stops a busy training worker from bidding, so each node runs one job at a time.c0mpute infernet temporal trainruns one single-node job per family on the existing auction.trusted_worker+upload_approvedare set.--yesafter a preview.--workerrestricts bids to an allowlist.c0mpute infernet temporal predict --sessionopens a session on one worker and keeps it there.plugins/infernet/module.tomllists the two workloads withmin_version = 0.1.66.Verified
cargo test -p c0mpute-core: 20 unit tests plus 2 e2e tests, all passing. The new loopback e2e uses a fakeinfernetthat speaks the same CLI and JSON as the real one.cargo test -p c0mpute-cli: 13 tests passing.E_SEQ_LATE), reset and close.Not done: moving a session to another worker over the network (checkpoint transfer), and settlement wiring for these receipts.
🤖 Generated with Claude Code