Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension


Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
37 changes: 21 additions & 16 deletions Cargo.lock

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

2 changes: 1 addition & 1 deletion Cargo.toml
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@ members = [
]

[workspace.package]
version = "0.2.33"
version = "0.2.34"
edition = "2024"
license = "MIT"
repository = "https://github.com/profullstack/c0mpute"
Expand Down
2 changes: 1 addition & 1 deletion apps/tui/package.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "@c0mpute/tui",
"version": "0.2.33",
"version": "0.2.34",
"private": true,
"type": "module",
"bin": {
Expand Down
4 changes: 2 additions & 2 deletions apps/web/package.json
Original file line number Diff line number Diff line change
@@ -1,6 +1,6 @@
{
"name": "c0mpute",
"version": "0.2.33",
"version": "0.2.34",
"private": true,
"type": "module",
"scripts": {
Expand All @@ -15,7 +15,7 @@
"@profullstack/x402-gateway": "^0.11.0",
"@sqlitecloud/drivers": "^1.0.880",
"ioredis": "^5.11.1",
"next": "16.3.8",
"next": "16.4.1",
"posthog-js": "^1.381.0",
"react": "^19.0.0",
"react-dom": "^19.0.0"
Expand Down
2 changes: 1 addition & 1 deletion apps/web/src/app/releases/latest.json/route.ts
Original file line number Diff line number Diff line change
Expand Up @@ -28,7 +28,7 @@ export const dynamic = "force-dynamic";
* this in sync with the workspace version in `Cargo.toml` on each release so a
* freshly built CLI reports "already latest" rather than a phantom upgrade.
*/
const FALLBACK_VERSION = "0.2.33";
const FALLBACK_VERSION = "0.2.34";
const FALLBACK_MIN_REQUIRED = "0.0.1";

type Channel = "stable" | "beta" | "nightly";
Expand Down
30 changes: 15 additions & 15 deletions bun.lock

Some generated files are not rendered by default. Learn more about how customized files appear on GitHub.

95 changes: 95 additions & 0 deletions dips/0026-infernet-temporal-workloads.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,95 @@
---
dip: 0026
title: "Run Infernet temporal policies (liquid networks) as c0mpute workloads"
status: Accepted
authors:
- riotcoder@profullstack.com
created: 2026-10-26
updated: 2026-10-26
discussion:
implementation: node/crates/c0mpute-core/src/temporal/, node/crates/c0mpute-cli/src/temporal.rs
supersedes:
superseded-by:
---

## Summary

Workers can train and serve Infernet's temporal-policy models (CfC and LTC
liquid neural networks from ncps, with a GRU baseline) through two workloads,
`infernet.temporal.train` and `infernet.temporal.infer`, on the existing
offer, bid, accept and receipt protocol. c0mpute runs no model code: it asks
the local `infernet` what it can execute and drives `infernet model train`
and `infernet model serve --stdio`. There is no new scheduler and no new
marketplace plugin; `plugins/infernet/module.toml` lists the two workloads.

## Motivation

The W3BS Liquid Agent pipeline (Infernet `docs/liquid-agent/`) learns small
UI policies from consented desktop demonstrations. Training is cheap enough
for one CPU, so the network's job is to let a user borrow a trusted machine
and to keep a stateful policy's hidden state on one worker while it serves a
session.

## Detailed design

Opt in on a worker (nothing is installed implicitly):

```bash
infernet model runtime setup --yes # operator's choice
c0mpute infernet temporal serve --train # accept training jobs
c0mpute infernet temporal serve --model sha256:<hash> # serve a model
```

Capability tags (advertised only for what `infernet model capabilities --json`
reports): `infernet:temporal:train`, `infernet:temporal:infer`,
`infernet:temporal:runtime[:<version>]`, `infernet:temporal:device:{cpu,cuda,mps}`,
`c0mpute:ram:<n>gb+` buckets, and `infernet:temporal:model:<sha256-hex>` for
each loaded model. Buyers turn requirements (GPU, minimum RAM, runtime
version, model hash) into `required_capabilities`; the existing dispatcher
filter does the matching. An `OfferFilter` adds per-workload checks before a
bid: a train worker validates the spec and does not bid while it is busy
(one job per node); an infer worker bids only for models it holds.

Training (`c0mpute infernet temporal train`): one single-node job per family
(CfC plus baselines). The spec carries a liquid-agent/v1 `TrainJobV1`, a
token-gated dataset URL served by the buyer, the dataset hash and an upload
URL. Workers refuse a job unless `privacy.mode = trusted_worker` and
`upload_approved = true`; the CLI refuses to send without `--yes` after a
preview. `--worker <peer-id>` restricts bids to trusted workers. The worker
runs `infernet model train --job` (which re-verifies every hash and the
secret scan), uploads `manifest.json`, `model.safetensors` and `eval.json`,
and receipts with the model hash. The buyer accepts the artifacts only if
the weights hash equals the receipt's `output_hash` and the manifest's.
Repeated idempotency keys reuse the finished model; failed or timed-out jobs
are re-auctioned up to `--retries` times.

Inference (`c0mpute infernet temporal predict --session <id>`): the first call
auctions a session open; the winner opens the session in its runtime,
publishes its X25519 key half and receipts with the model hash. Later calls
are AES-256-GCM sealed with an HKDF key bound to the session and go only to
that worker, which answers only the peer that opened the session (gossipsub
signs the source). A call id cache gives exactly-once execution when the
buyer republishes. Sequence rules (duplicate, late, gap), TTL, capacity and
account scope are enforced by the Infernet runtime.

## Compatibility

Additive. Workers that do not run `temporal serve` never see these topics.
`c0mpute infernet <anything else>` still delegates to the infernet binary.

## Security and privacy

- A worker sees the redacted training data in plaintext; transport security
does not change that. Local training is the default upstream; remote
training is an explicit, previewed choice of a trusted worker.
- The dataset URL is gated by a random token; use `--public-url https://...`
off the LAN. Artifacts are verified by hash, not trusted.
- Session traffic is encrypted between buyer and worker; other subscribers of
the topic see only ciphertext and routing ids.

## Not done

- Cross-worker session failover (checkpoint transfer) over the network; the
runtime supports checkpoint/restore locally.
- Price discovery beyond the existing cooperative bid, and settlement wiring
for these receipts.
1 change: 1 addition & 0 deletions dips/README.md
Original file line number Diff line number Diff line change
Expand Up @@ -132,3 +132,4 @@ superseded-by: <DIP-NNNN if this gets replaced>
| 0016 | Abuse policy: CSAM hash blocklist, per-worker opt-in | Draft |
| 0024 | Separate the protocol from every hosted service that accelerates it | Draft |
| 0025 | Signed envelopes and native `did:c0mpute` identity | Draft |
| 0026 | Infernet temporal policies (liquid networks) as c0mpute workloads | Accepted |
Loading
Loading