Edge Inference: ONNX + Jetson + MediaTek NPU + Qualcomm AI Engine

# === HF → ONNX → Edge Deploy ===
# 1. HF model'i ONNX'e dönüştür
optimum-cli export onnx \
    --model meta-llama/Llama-3.2-1B-Instruct \
    --opset 17 \
    --dtype fp16 \
    llama-3.2-1b-onnx/
 
# 2. Quantize (W4A8 mixed)
python -m onnxruntime.quantization.matmul_4bits_quantizer \
    --input llama-3.2-1b-onnx/ \
    --output llama-3.2-1b-onnx-int4/ \
    --block_size 32
 
# 3. Edge deploy — ONNX Runtime Mobile (Android/iOS)
# Kotlin (Android):
val session = OrtEnvironment.getEnvironment().createSession(
    "llama-3.2-1b-onnx-int4/model.onnx"
)
 
# Swift (iOS):
let session = try ORTSession(env: env, modelPath: path, sessionOptions: options)
let outputs = try session.run(withInputs: inputs)
 
# 4. Edge benchmark
# Pixel 8 Pro: SmolLM3 1.7B Q4 → 12 tok/s, first-token 150 ms
# Snapdragon 8 Gen 3: Llama 3.2 3B Q4 → 18 tok/s, first-token 95 ms