Model Quantization & Deployment

This article introduces how to quantize KataGo models to reduce resource requirements, and deployment solutions for various platforms.

---

Quantization Techniques Overview

Why Quantization?

Precision	Size	Speed	Accuracy Loss
FP32	100%	Baseline	0%
FP16	50%	+50%	~0%
INT8	25%	+100%	<1%

Quantization Types

---

FP16 Half Precision

Concept

Convert 32-bit floating point to 16-bit:

# FP32 → FP16 conversion
model_fp16 = model.half()

# Inference
with torch.cuda.amp.autocast():
    output = model_fp16(input.half())

KataGo Settings

# config.cfg
useFP16 = true           # Enable FP16 inference
useFP16Storage = true    # FP16 intermediate storage

Performance Impact

GPU Series	FP16 Speedup
GTX 10xx	None (no Tensor Core)
RTX 20xx	+30-50%
RTX 30xx	+50-80%
RTX 40xx	+80-100%

---

INT8 Quantization

Quantization Process

import torch.quantization as quant

# 1. Prepare model
model.eval()
model.qconfig = quant.get_default_qconfig('fbgemm')

# 2. Prepare for quantization
model_prepared = quant.prepare(model)

# 3. Calibration (use representative data)
with torch.no_grad():
    for data in calibration_loader:
        model_prepared(data)

# 4. Convert to quantized model
model_quantized = quant.convert(model_prepared)

Calibration Data

def create_calibration_dataset(num_samples=1000):
    """Create calibration dataset"""
    samples = []

    # Sample from actual games
    for game in random_games(num_samples):
        position = random_position(game)
        features = encode_state(position)
        samples.append(features)

    return samples

Notes

• INT8 quantization requires calibration data
• Some layers may not be suitable for quantization
• Need to test accuracy loss

---

TensorRT Deployment

Conversion Process

import tensorrt as trt

def convert_to_tensorrt(onnx_path, engine_path):
    logger = trt.Logger(trt.Logger.WARNING)
    builder = trt.Builder(logger)
    network = builder.create_network(
        1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
    )
    parser = trt.OnnxParser(network, logger)

    # Parse ONNX model
    with open(onnx_path, 'rb') as f:
        parser.parse(f.read())

    # Set optimization options
    config = builder.create_builder_config()
    config.max_workspace_size = 1 << 30  # 1GB

    # Enable FP16
    config.set_flag(trt.BuilderFlag.FP16)

    # Build engine
    engine = builder.build_engine(network, config)

    # Save
    with open(engine_path, 'wb') as f:
        f.write(engine.serialize())

Using TensorRT Engine

def inference_with_tensorrt(engine_path, input_data):
    # Load engine
    with open(engine_path, 'rb') as f:
        engine = trt.Runtime(logger).deserialize_cuda_engine(f.read())

    context = engine.create_execution_context()

    # Allocate memory
    d_input = cuda.mem_alloc(input_data.nbytes)
    d_output = cuda.mem_alloc(output_size)

    # Copy input
    cuda.memcpy_htod(d_input, input_data)

    # Execute inference
    context.execute_v2([int(d_input), int(d_output)])

    # Get output
    output = np.empty(output_shape, dtype=np.float32)
    cuda.memcpy_dtoh(output, d_output)

    return output

---

ONNX Export

PyTorch → ONNX

import torch.onnx

def export_to_onnx(model, output_path):
    model.eval()

    # Create sample input
    dummy_input = torch.randn(1, 22, 19, 19)

    # Export
    torch.onnx.export(
        model,
        dummy_input,
        output_path,
        input_names=['input'],
        output_names=['policy', 'value', 'ownership'],
        dynamic_axes={
            'input': {0: 'batch_size'},
            'policy': {0: 'batch_size'},
            'value': {0: 'batch_size'},
            'ownership': {0: 'batch_size'}
        },
        opset_version=13
    )

Validate ONNX Model

import onnx
import onnxruntime as ort

# Validate model structure
model = onnx.load("model.onnx")
onnx.checker.check_model(model)

# Test inference
session = ort.InferenceSession("model.onnx")
output = session.run(None, {'input': input_data})

---

Cross-Platform Deployment

Server Deployment

# docker-compose.yml
version: '3'
services:
  katago:
    image: katago/katago:latest
    deploy:
      resources:
        reservations:
          devices:
            - driver: nvidia
              count: 1
              capabilities: [gpu]
    volumes:
      - ./models:/models
      - ./config:/config
    command: >
      katago analysis
      -model /models/kata-b18c384.bin.gz
      -config /config/analysis.cfg

Desktop Application Integration

# Embed KataGo in Python application
import subprocess
import json

class KataGoProcess:
    def __init__(self, katago_path, model_path):
        self.process = subprocess.Popen(
            [katago_path, 'analysis', '-model', model_path],
            stdin=subprocess.PIPE,
            stdout=subprocess.PIPE,
            text=True
        )

    def analyze(self, moves):
        query = {
            'id': 'query1',
            'moves': moves,
            'rules': 'chinese',
            'komi': 7.5,
            'boardXSize': 19,
            'boardYSize': 19
        }
        self.process.stdin.write(json.dumps(query) + '\n')
        self.process.stdin.flush()

        response = self.process.stdout.readline()
        return json.loads(response)

Mobile Deployment

iOS (Core ML)

import coremltools as ct

# Convert to Core ML
mlmodel = ct.convert(
    model,
    inputs=[ct.TensorType(shape=(1, 22, 19, 19))],
    minimum_deployment_target=ct.target.iOS15
)

mlmodel.save("KataGo.mlmodel")

Android (TensorFlow Lite)

import tensorflow as tf

# Convert to TFLite
converter = tf.lite.TFLiteConverter.from_saved_model(model_path)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.target_spec.supported_types = [tf.float16]

tflite_model = converter.convert()

with open('katago.tflite', 'wb') as f:
    f.write(tflite_model)

Embedded Systems

Raspberry Pi

# Use Eigen backend (CPU only)
./katago gtp -model kata-b10c128.bin.gz -config rpi.cfg

# rpi.cfg - Raspberry Pi optimized settings
numSearchThreads = 4
maxVisits = 100
nnMaxBatchSize = 1

NVIDIA Jetson

# Use CUDA backend
./katago gtp -model kata-b18c384.bin.gz -config jetson.cfg

---

Performance Comparison

Performance by Deployment Method

Deployment	Hardware	Playouts/sec
CUDA FP32	RTX 3080	~3000
CUDA FP16	RTX 3080	~5000
TensorRT FP16	RTX 3080	~6500
OpenCL	M1 Pro	~1500
Core ML	M1 Pro	~1800
TFLite	Pixel 7	~50
Eigen	RPi 4	~15

Model Size Comparison

Format	b18c384 Size
Original (.bin.gz)	~140 MB
ONNX FP32	~280 MB
ONNX FP16	~140 MB
TensorRT FP16	~100 MB
TFLite FP16	~140 MB

---

Deployment Checklist

• Select appropriate quantization precision
• Prepare calibration data (INT8)
• Export to target format
• Validate acceptable accuracy loss
• Test performance on target platform
• Optimize memory usage
• Set up automated deployment pipeline

---

Further Reading

• GPU Backend & Optimization — Basic performance optimization
• Evaluation & Benchmarking — Verify post-deployment performance
• Integrate into Your Project — API integration examples