TL;DR

# Do inference simulations
$ sh simul.sh

# Make Heatmaps about $E_I$, $d$, and $Q_C$
$ sh simul_heatmap.sh

---

🛜 Run Server

Run chatbot, use "samsum" dataset.

$ CUDA_VISIBLE_DEVICES=0,1 \
python server.py --dataset "samsum" --port 30327

or

$ CUDA_VISIBLE_DEVICES=0 \
python server.py --dataset "samsum" --port 30327

$ CUDA_VISIBLE_DEVICES=1 \
python server.py --dataset "samsum" --port 30328

• dataset: "samsum" as default
• port: "30327" as default

Inference

$ curl -X POST -H 'Content-Type: application/json' http://<IP>:<PORT>/chat -d '{"prompt": "Hello, World!", "version": "f9d7889c151fbcb863ef4c770fbcca33cd960cf23e05ddee113d09eb01aa1be8", "max_length": 128, "seed": 950327}'

{"result":"I am not the only AI in the world. There are many AI."}

$ curl -X POST -H 'Content-Type: application/json' http://<IP>:<PORT>/chat -d '{"prompt": "Hello, World!", "max_length": 128, "seed":950328}'

{"result":"How are you?"}

• prompt: input
• version: [optional] which version (default: 0x00; latest)
• temperature: [optional] temperature (default: 0.9)
• max_length: [optional] max_length (default: 128)
• seed: [optional] seed (seed: 42; INIT_SEED)

Train

$ curl -X POST -H 'Content-Type: application/json' http://<IP>:<PORT>/chat -d '{"from": 0, "to": 10}'

• from: dataset pagination starting point
• to: dataset pagination end point

Returns:

• previous: Previous version of model. Unique key per NN — hash of the weights.
• current: Current version of model. Unique key per NN — hash of the weights.

---

📈 Simulation

Inference

Get inference results:

$ sh test_0.sh
$ sh test_1.sh

• (round, time, input tokens, output tokens, response) test results are saved in results.

Then refine them:

$ python simulate/refine.py

Run BRAIN simulation:

$ python simulate/nodes.py --help
usage: nodes.py [-h] [--seed S] [--repeat R] [--stop P] [--verbose V] [--path PATH] [--interval I] [--size S] [--freq F] [--latency L] [--nodes N] [--byz B] [--epoch E] [--d D] [--qc QC] [--qto O]

Hyperparameters

options:
  -h, --help    show this help message and exit
  --seed S      Seed for simulation
  --repeat R    # simulation
  --stop P      Stop this round when for loop reaching P
  --verbose V   0: silent, 1: speak, 2: speak all
  --path PATH   path for saving pkl files
  --interval I  Average Block Time
  --size S      Average # of Transaction per block
  --freq F      Inference Request / Normal Tx
  --latency L   Latency of normal EVM execution (s)
  --nodes N     Number of nodes
  --byz B       The number of Byzantine nodes
  --epoch E     Epoch [blocks]
  --d D         difficulty (0, 2^256-1], but scaled in (0, 256]
  --qc QC       Quorum of Commitments
  --qto O       Inference Request's Timeout [blocks] \ 0 for infinity

or

$ sh simul.sh

Visualization

$ python visualization/heatmap.py

or

$ sh simul_heatmap.sh