Skip to content

Universal Continuous GPU Benchmark for Gonka Hosts (vLLM v0.19+, Anti-Sybil)

🔴 Expired

Author: Evgenii Maksimenkov Created: 2026-04-25 06:31 UTC Closes: 2026-05-25 06:31 UTC Language: EN Votes: 0 Avg. Bid: 0.00 GNK

Develop a vLLM v0.19+ universal GPU benchmark that runs continuously during the epoch without interrupting inference, fairly scores any model size, and is hardened against hot-plug and Sybil attacks.


Full Proposal

Context

Gonka measures Host compute capacity via Proof of Compute (PoC) Sprints. The current Sprint model concentrates measurement into a short synchronized window, which (a) creates a measurable inference-revenue gap during the Sprint and (b) opens a surface for "burst attacks" — Hosts attaching extra GPUs only for the Sprint window and detaching them afterwards.

This tender procures a universal, continuous GPU benchmark that runs throughout the epoch on the Host's actual production inference stack (vLLM v0.19+), produces a verifiable per-GPU compute score, and is robust against hot-plug, Sybil, and replay attacks.

The deliverable replaces (or supplements, subject to governance) burst-style Sprint scoring with a continuously-sampled metric that cannot be gamed by transient hardware, while preserving existing PoC validation properties (randomized verification, on-chain artifacts, slashing).

2. Scope of Work

The Contractor shall design, implement, test, document, and upstream a benchmark subsystem (hereinafter — the Benchmark) consisting of:

  1. A measurement agent integrated into the ML Node, running alongside vLLM serving.
  2. An on-chain attestation protocol for submitting and verifying samples.
  3. A scoring function producing per-Host weight inputs consumable by the existing PoC pipeline.
  4. A reference verifier runnable by any Host to independently re-check published artifacts.
  5. Operator tooling, dashboards, and full test/CI coverage.

3. Functional Requirements

3.1. Continuous, Non-Interrupting Operation

  • The Benchmark MUST run as a persistent workload throughout the epoch (17280 blocks, ~24h), not as a discrete Sprint window.
  • The Benchmark MUST NOT pause, drain, or evict active inference requests. p99 inter-token latency degradation on the served production model MUST NOT exceed 3% versus a clean baseline measured under identical concurrency.
  • The Benchmark MUST integrate with vLLM v0.19+ scheduler hooks (continuous batching, chunked prefill, prefix caching) and submit its workload as first-class requests through the same engine — not via a parallel CUDA context that would distort accounting.
  • KV cache pressure introduced by the Benchmark MUST be bounded by an operator-configurable ceiling (default: ≤5% of gpu_memory_utilization).
  • The Benchmark MUST yield CPU/GPU resources to user-paid inference under load (configurable QoS class; default: lower priority than paid traffic, higher than idle).

3.2. Universality Across Models

  • The Benchmark MUST produce comparable scores independent of the model the Host currently serves, including but not limited to: Qwen3-235B-A22B-Instruct-2507-FP8, dense models in the 7B–70B range, and MoE models with active-parameter counts from 7B to 50B+.
  • Scoring MUST normalize across: dense vs. MoE architectures, FP8 / BF16 / FP16 / INT8 precisions, single-GPU and tensor-parallel (TP=2/4/8) and pipeline-parallel deployments, NVLink and PCIe interconnects.
  • The Benchmark MUST define a canonical reference workload: a fixed set of synthetic transformer micro-tasks (attention, MLP/MoE routing, all-reduce, prefill, decode) executed on a canonical Sprint-style transformer with randomized layer transformations seeded by on-chain randomness, identical for every Host in a given measurement window.
  • Score units MUST be expressed in AI Tokens per second per GPU-equivalent, calibrated so that hardware tiers (e.g., A100 80GB, H100, H200, RTX PRO 6000 Blackwell, MI300X) map to a stable, monotonic ranking across the supported model spectrum.

3.3. Sampling and Randomization

  • Each epoch MUST contain N ≥ 64 micro-measurements per GPU, scheduled at unpredictable times derived from a VRF over (epoch_seed, host_address, gpu_uuid).
  • The exact start time, duration, seed, model permutation, and input nonces of each micro-measurement MUST NOT be predictable by the Host more than one block (~5s) in advance.
  • Micro-measurement duration MUST be bounded (target: 200ms–2s) so that no single measurement materially affects served latency, while N samples aggregate to a low-variance epoch score (target: per-Host score CoV ≤ 2% across 3 consecutive honest epochs on identical hardware).

4. Security Requirements

4.1. Anti-Hot-Plug / Anti-Burst Hardware Attack

  • The Benchmark MUST detect and penalize Hosts that add GPUs only for measurement windows. Required mechanisms:
  • Continuous GPU presence attestation: each declared GPU MUST report a signed heartbeat at least once per 30s, including gpu_uuid, pci_bus_id, nvml_serial, vbios_version, driver_version, current power draw, current SM clock, current memory clock, ECC error counters, and a monotonic uptime counter sourced from the device.
  • GPUs that miss more than 0.5% of heartbeats per epoch MUST be excluded from the epoch score for that Host.
  • Hardware fingerprint binding: a GPU's contribution to the epoch score MUST be discounted by min(1, observed_uptime_in_epoch / required_uptime) where required_uptime ≥ 95% of the epoch.
  • Cross-Host fingerprint deduplication: the same nvml_serial / vbios / pci triple appearing on two Hosts within a rolling 7-epoch window MUST trigger automatic disqualification of both Hosts pending governance review.

4.2. Anti-Sybil and Anti-Replay

  • All measurement inputs MUST be derived from on-chain VRF output; a Host cannot precompute outputs.
  • All measurement outputs MUST be committed via (commit, reveal) with a commit deadline ≤1 block after measurement and reveal ≤2 blocks after commit. Late reveals MUST be rejected.
  • Replay of a previous epoch's outputs MUST be detectable via VRF binding to (epoch_id, block_height, gpu_uuid).
  • No-mock guarantee: the Benchmark workload MUST share weights, KV cache, and CUDA streams with the production vLLM engine, so substituting a synthetic high-throughput kernel for the real one is detectable via cache-coherency and timing-side-channel signatures.

4.3. Anti-Outsourcing

  • Sample latency MUST be measured end-to-end inside the Chain Node, not self-reported by the ML Node, with RTT bounds enforced (default: ≤50ms intra-Host, configurable). This makes it economically irrational to forward measurement workloads to a remote GPU farm.
  • Optional TEE attestation (SEV-SNP, TDX, or NVIDIA H100/H200 Confidential Computing) SHOULD be supported as a verification booster (higher Confirmation Ratio for TEE-attested Hosts).

4.4. Statistical Validation Compatible with Existing PoC

  • Verification MUST follow Gonka's existing statistical (non-bitwise) validation model: a Validator Host re-runs a random subset (1–10% per Reputation tier) of the Executor's micro-measurements and accepts results within a defined tolerance band (e.g., logits L2 distance ≤ ε, throughput within ±5% of declared).
  • Cheating detection MUST trigger the existing slashing path: full epoch reward forfeiture for the offending Host.

5. Technical Requirements

Area Requirement
Language Python 3.11+ for agent; Go for Chain Node integration; Rust acceptable for hot-path kernels
Runtime vLLM v0.19+, CUDA 12.4+, NCCL 2.21+, PyTorch 2.4+
GPUs NVIDIA Ampere/Hopper/Blackwell (A100, H100, H200, B200, RTX PRO 6000 Blackwell); AMD MI300X support as stretch goal
Interconnects NVLink 3/4/5, NVSwitch, PCIe Gen4/Gen5, InfiniBand HDR/NDR
Orchestration Docker + docker-compose parity with existing inference-ignite deployment; optional Kubernetes Helm chart
Observability Prometheus metrics, structured JSON logs, OpenTelemetry traces; Grafana dashboard with per-GPU score, heartbeat status, sample histogram
Config All thresholds (heartbeat interval, uptime floor, KV ceiling, QoS class, sample count) MUST be governance-tunable parameters, not compile-time constants
API OpenAI-compatible inference path MUST remain unchanged externally; new /poc/continuous/* endpoints exposed on the API Node
Determinism Reference workload MUST be reproducible with seed, top_logprobs, and batch-invariant kernels enabled; tolerance bands documented per kernel

6. Deliverables

  1. Source code in the gonka-ai/gonka GitHub organization, dual-licensed compatible with the Gonka Protocol License, with ≥85% unit test coverage on the agent and ≥70% on Chain Node integration.
  2. Reference Docker images published to a public registry, signed with cosign.
  3. Whitepaper update extending the existing pow-security-analysis.pdf, formally analyzing the continuous-measurement security model and providing a probabilistic bound on the expected cost-of-attack for hot-plug and Sybil scenarios.
  4. Migration guide for existing Hosts, including a parallel-run period (≥2 epochs) where both legacy Sprint and Continuous Benchmark scores are produced and reconciled.
  5. Governance proposal text ready for on-chain submission to activate the new scoring path.
  6. Operator runbook: deployment, troubleshooting, expected metric ranges per GPU SKU.
  7. Conformance test suite runnable by any Host as inferenced benchmark verify.

7. Acceptance Criteria

The Benchmark is accepted when, on a testnet of ≥20 heterogeneous Hosts running for ≥7 consecutive epochs:

  • Inference p99 latency degradation ≤3% versus pre-deployment baseline.
  • Per-Host score CoV ≤2% across 3 consecutive honest epochs on stable hardware.
  • All four reference attack scenarios (hot-plug burst, output replay, remote outsourcing, kernel substitution) are detected with ≥99% true-positive rate and ≤0.1% false-positive rate.
  • Score ranking across Hosts is monotonic in raw GPU FLOPS×memory-bandwidth product within ±5% noise, independent of which model the Host serves.
  • Verifier re-execution of a random 10% sample yields ≥99.5% agreement with Executor-published artifacts on honest Hosts.

Comments (12)

💬 Evgenii Maksimenkov

2026-04-25 18:41 · 👍 2 · 👎 0

What are your thoughts?


💬 Alex A.A.

2026-04-25 19:05 · 👍 2 · 👎 0

Начало положено! Язык работает.


💬 Alex A.A.

2026-04-25 19:09 · 👍 1 · 👎 0

Через телеграм можно зайти. Круто!


💬 Nikita

2026-04-25 19:45 · 👍 1 · 👎 0

🔥🔥🔥


💬 Alisa Chudinov

2026-04-25 19:47 · 👍 1 · 👎 0

Very intresting!


💬 Mikhail Chudinov

2026-04-25 19:50 · 👍 1 · 👎 0

Изучаю пока функционал сервиса. Проблема с мульти PoC точно есть, и над ней надо думать. Но такой бенчмарк не единственный подход. Думаю что надо думать о полностью рыночном механизме, который не может быть заменен бенчмарком.


💬 Alex Sharoiko Александр Шаройко

2026-04-26 19:57 · 👍 1 · 👎 0

Тест. Интересно, как работает ) Не обращайте внимание )


💬 Alex Sharoiko Александр Шаройко

2026-04-26 19:58 · 👍 0 · 👎 0

Тестирую функционал тоже ) Не обращай внимание )


💬 Alex Sharoiko Александр Шаройко

2026-04-26 19:59 · 👍 0 · 👎 0

Hi! How is it going?!


💬 Alex Sharoiko Александр Шаройко

2026-04-26 20:09 · 👍 0 · 👎 0

https://www.youtube.com/shorts/WTSRfhsLYwQ


💬 Alex Sharoiko Александр Шаройко

2026-04-29 04:45 · 👍 0 · 👎 0

А сколько проверка PoC занимает сейчас ресурсов? Предлагаемое решение будет забирать до 3%, верно?

  1. Когда у сети будет больше Инференса, не решит ли это проблему с читерами? Не будут ли они улетать в МисРейт тогда?

💬 Alex Sharoiko Александр Шаройко

2026-04-26 19:57 · 👍 0 · 👎 1

But what will happend if I reply in English?