Skip to content

FAQ.md

Latest commit

 

History

History
116 lines (79 loc) · 3.51 KB

FAQ.md

File metadata and controls

116 lines (79 loc) · 3.51 KB
Raw

GRAIL Miner FAQ

Frequently asked questions for GRAIL miners on Bittensor subnet 81.

1. What GPU should I use?

Minimum requirements (compute.min.yaml:1):

  • GPU: NVIDIA A100 40GB/80GB
  • VRAM: 24GB minimum, 40GB recommended
  • Compute Capability: 9.0+ (Hopper architecture)
  • CPU: 16+ cores @ 3.2GHz
  • RAM: 64GB minimum
  • Storage: 100GB SSD (50K IOPS)

For competitive mining: Multiple GPUs (2-8+ A100s) required to generate enough rollouts.

See compute.min.yaml for complete specifications.

2. How do I debug and optimize my miner?

Don't debug on mainnet! Run a local validator instead.

Quick steps:

  1. Run local validator: grail -vv validate > validate.log 2>&1 &
  2. Run your miner: grail -vv mine > mine.log 2>&1 &
  3. Monitor: tail -f validate.log | grep "uid=<your-uid>"
  4. Reduce ban time: Set FAILURE_LOOKBACK_WINDOWS = 1 in grail/shared/constants.py:223

Full guide: miner-debugging.md

Optimization libraries: vLLM, TensorRT-LLM, DeepSpeed, Flash Attention

See vllm-setup.md for integration.

3. Is running the base code enough to earn incentives?

No. Base code is 10-20x too slow.

Requirements (grail/shared/constants.py:187, grail/shared/constants.py:204):

  • Target: ~5,120 unique rollouts per window (6 minutes)
  • Rate needed: ~14 rollouts/second sustained
  • Base code: ~1-2 rollouts/second (insufficient)

To compete, you need:

  • Multiple GPUs (2-8+)
  • Optimized inference (vLLM, SGLang)
  • Parallelization and batching
  • Profiling and tuning

Base code is a starting point, not production.

4. How are incentives structured?

Based on unique rollouts over 12 windows (grail/scoring/weights.py:34):

Scoring formula:

  1. Count unique rollouts (cap: 61,440 total across 12 windows)
  2. Apply superlinear: score = unique^4.0 (grail/shared/constants.py:179)
  3. Normalize against cap
  4. Apply 80% burn to UID 0 (grail/shared/constants.py:197)
  5. Distribute remaining 20% to miners

Key insight: Superlinear exponent (4.0) means 2x rollouts = 16x score. Volume dominates.

Failure penalties (grail/scoring/weights.py:75):

  • Any failure in last 14 windows → zero weight
  • Zero activity → zero weight
  • Never validated → zero weight

Example scores:

Rollouts (12 windows) Relative Score
5,000 1x
10,000 16x
30,000 1,296x

Top miners near the 61,440 cap dominate weight distribution.

5. How many rollouts are top miners generating?

Check the public dashboards:

WandB (README.md:27): https://wandb.ai/tplr/grail

  • Filter by validator runs
  • Look for metrics: miner/{uid}/rollouts_per_window, miner/{uid}/unique_rollouts

Grafana (README.md:30): https://grail-grafana.tplr.ai/

  • Real-time logs and network stats

Typical top miner performance:

  • 3,000-8,000 rollouts per window
  • 30,000-60,000+ unique rollouts per 12 windows
  • 90%+ validation success rate

To compete: Match or exceed top miner throughput using multi-GPU setups and optimized inference.

Additional Resources