Skip to content

Bounty for open-source PoC throughput contributions

🔴 Expired

Author: Serhii Hovorov Created: 2026-05-01 21:04 UTC Closes: 2026-05-08 21:04 UTC Language: EN Votes: 0 Avg. Bid: 0.00 GNK

Pay a fixed GNK/USDT bounty to anyone who upstreams a reproducible ≥5% PoC throughput optimization. Lifts the network floor; rewards contribution over hoarding.

Full Proposal

Problem

PoC throughput optimizations are network-wide value, but today they stay private. A pool operator who finds a +5% kernel rewrite has every incentive to keep it secret — they get 5% more weight relative to the rest of the network for as long as nobody else has it. The result:

  • Big pools capture the gains; smaller miners stay slow
  • Optimizations don't flow back into the public gonka-ai/vllm and mlnode repos
  • New miners join a network whose floor is lower than it could be
  • The chain's actual realized GNK/GPU-hour is below what's technically achievable

This is a coordination failure, not a malice problem. There's no mechanism today
that pays contributors for raising the network floor.

## Proposal

Establish a standing bounty in GNK for any PoC-related throughput optimization
that meets all of these criteria:

  1. Merged upstream into a public Gonka repo — gonka-ai/vllm,
    gonka-ai/mlnode, gonka-ai/inference-chain, or a designated successor.
  2. Reproducible benchmark — a public A/B harness against a reference build
    (currently Qwen3-235B-A22B-Instruct-2507-FP8) on at least one supported GPU
    class (B200 / H100 / A100). Includes "before" and "after" commit SHAs.
  3. Net positive impact — at least 5% sustained nonces/min uplift on the
    reference setup.
  4. Independently verified — reproduced by ≥ 2 unrelated miners on their
    own hardware, attested on-chain.

### Reward structure (for community discussion)

  • Flat bounty per qualifying contribution: 2,000 GNK
  • Bonus tier for ≥ 10% uplift: +3,000 GNK
  • Funded from a dedicated optimization-bounty pool topped up via routine
    community-pool spend proposals tied to confirmed merges — no extra emissions
  • Contributor names the recipient address at PR submission
  • One person can receive multiple bounties for distinct optimizations

### Why this is fair to the rest of the network

The bounty is one-time per contribution and capped. The throughput gain is permanent and accrues to every miner forever. The community-pool trades a
small fixed cost now for a permanent uplift in the network's earning capacity. Positive-sum by construction.

## Why now

A reward mechanism works best when there's already a verifiable example of
someone choosing to upstream rather than hoard. There is one in flight today:

  • gonka-ai/vllm#36 —
    @torch.compile decorator on apply_householder. Measured +10.2% on 8×B200 and +12.5% on 8×H100 for Qwen3-235B-FP8. One-line change.
    Reproducible A/B harness in the PR.

If this tender passes, that PR becomes the first concrete test of the bounty
workflow: independent reproduction on community hardware, on-chain attestation,
payout via a follow-up community-pool spend. Every future optimization — from anyone — would follow the same path.

Every week this stays unaddressed is a week where private wins outpace public
ones. There are pool operators sitting on similar optimizations right now. A clear, public reward path flips the math from "hoard" to "ship."

## What this is not

  • Not a proposal to pay any specific person — it's a standing policy
    anyone can claim.
  • Not a vote on vllm#36 itself — that PR lands or not on its technical
    merits regardless of this tender.
  • Not a binding tx — tenders are indicative polls. If this passes, a separate formal community-pool spend funds the first round.

## Open questions for the community

  • Right uplift threshold? (5% suggested; could be 3% or 10%)
  • Right bounty size? (2k + 3k GNK suggested; could scale with measured
    network-wide impact)
  • Verification process? (2 independent miners suggested)
  • Scope: PoC-only, or also general inference throughput? (broader = more impact, harder to verify)
  • Eligibility for retroactive contributions? (e.g. anything merged in the last 30 days)

Comments (1)

💬 Mikhail Chudinov

2026-05-04 14:34 · 👍 1 · 👎 0

Идея правильная, считаю что давать нужно минимум эквивалент 1000$ за каждый % прироста веса. Дала оптимизация +10% = выдать 10к$ Курс GNK может сильно плавать, поэтому такой бонус я бы привязывал к USD. Это будет хорошее, более понятное предложение для программистов, которые впервые услышали про гонку. Что тут можно пооптимизировать, понятным образом протестировать результат в нонсах в минуту скриптом и получить разовую приличную выплату в случае успеха.

Создавать отдельный пулл под финансирование не вижу смысла. Считаю что достаточно каждое такое улучшение проводить как пропозал отдельный, из комьюнити пула. В качестве верификации что эта оптимизация полезна выступают майнеры, своим голосованием.

View on gonka.vote