LLM Edge Deployment: Mobile, IoT, and Edge Inference

from dataclasses import dataclass
from typing import List, Dict
from enum import Enum

class TargetPlatform(Enum):
    ANDROID = "android"
    IOS = "ios"
    RASPBERRY_PI = "raspberry_pi"
    ESP32 = "esp32"
    BROWSER = "browser"

@dataclass
class EdgeProfile:
    platform: TargetPlatform
    max_model_size_mb: float
    max_ram_mb: float
    supports_fp16: bool
    supports_int8: bool
    supports_int4: bool

class EdgeOptimizer:
    PLATFORM_PROFILES = {
        TargetPlatform.ANDROID: EdgeProfile(
            platform=TargetPlatform.ANDROID,
            max_model_size_mb=200, max_ram_mb=512,
            supports_fp16=True, supports_int8=True, supports_int4=True
        ),
        TargetPlatform.IOS: EdgeProfile(
            platform=TargetPlatform.IOS,
            max_model_size_mb=150, max_ram_mb=400,
            supports_fp16=True, supports_int8=True, supports_int4=False
        ),
        TargetPlatform.RASPBERRY_PI: EdgeProfile(
            platform=TargetPlatform.RASPBERRY_PI,
            max_model_size_mb=50, max_ram_mb=256,
            supports_fp16=True, supports_int8=True, supports_int4=True
        ),
        TargetPlatform.BROWSER: EdgeProfile(
            platform=TargetPlatform.BROWSER,
            max_model_size_mb=50, max_ram_mb=200,
            supports_fp16=True, supports_int8=False, supports_int4=False
        ),
    }

    def get_optimization_plan(self, platform: TargetPlatform,
                               model_size_mb: float) -> Dict:
        profile = self.PLATFORM_PROFILES[platform]
        plan = {"target": platform.value, "steps": []}
        if model_size_mb > profile.max_model_size_mb:
            reduction_needed = model_size_mb / profile.max_model_size_mb
            if profile.supports_int4 and reduction_needed > 4:
                plan["steps"].append({"type": "quantize", "bits": 4})
            elif profile.supports_int8:
                plan["steps"].append({"type": "quantize", "bits": 8})
            plan["steps"].append({"type": "prune", "sparsity": 0.3})
        if profile.supports_fp16:
            plan["steps"].append({"type": "convert_fp16": True})
        plan["steps"].append({"type": "graph_optimize"})
        return plan

    def estimate_memory(self, params: int, platform: TargetPlatform) -> Dict:
        profile = self.PLATFORM_PROFILES[platform]
        fp32_gb = params * 4 / (1024**3)
        fp16_gb = params * 2 / (1024**3)
        int8_gb = params * 1 / (1024**3)
        int4_gb = params * 0.5 / (1024**3)
        return {
            "fp32_gb": round(fp32_gb, 3),
            "fp16_gb": round(fp16_gb, 3),
            "int8_gb": round(int8_gb, 3),
            "int4_gb": round(int4_gb, 3),
            "fits_device": fp16_gb <= profile.max_ram_mb / 1024
        }

import time
from dataclasses import dataclass
from typing import Optional, Dict

@dataclass
class InferenceResult:
    text: str
    tokens_generated: int
    latency_ms: float
    memory_used_mb: float
    tokens_per_second: float

class EdgeInferenceEngine:
    def __init__(self, model_path: str, max_tokens: int = 256):
        self.model_path = model_path
        self.max_tokens = max_tokens
        self.model_loaded = False

    def load_model(self) -> bool:
        self.model_loaded = True
        return True

    def generate(self, prompt: str, max_new_tokens: int = 100) -> InferenceResult:
        if not self.model_loaded:
            raise RuntimeError("Model not loaded")
        start = time.time()
        generated = self._run_inference(prompt, max_new_tokens)
        latency = (time.time() - start) * 1000
        tokens = len(generated.split())
        return InferenceResult(
            text=generated,
            tokens_generated=tokens,
            latency_ms=latency,
            memory_used_mb=50.0,
            tokens_per_second=tokens / max(latency / 1000, 0.001)
        )

    def _run_inference(self, prompt: str, max_tokens: int) -> str:
        return f"Generated response for: {prompt[:50]}..."

    def benchmark(self, prompts: List[str], iterations: int = 10) -> Dict:
        latencies = []
        tps_values = []
        for prompt in prompts:
            for _ in range(iterations):
                result = self.generate(prompt)
                latencies.append(result.latency_ms)
                tps_values.append(result.tokens_per_second)
        return {
            "avg_latency_ms": sum(latencies) / len(latencies),
            "p95_latency_ms": sorted(latencies)[int(0.95 * len(latencies))],
            "avg_tps": sum(tps_values) / len(tps_values)
        }