"""Execution policy: PERFORMANCE | BALANCED | EFFICIENCY. A policy is a set of decision thresholds, not marketing words. Every knob below has a consumer in the scheduler/monitor and changes real behavior: - probe_iterations: how many real micro-benchmark iterations run when choosing a strategy / precision. - strategy aggressiveness: PERFORMANCE may test model_parallel for large models; EFFICIENCY never leaves single-GPU unless VRAM forces it. - monitoring frequency: PERFORMANCE samples GPU state less often during training (less overhead); EFFICIENCY idles at the slowest safe rate. """ from __future__ import annotations from dataclasses import dataclass from enum import Enum from typing import Dict class PerformanceMode(str, Enum): PERFORMANCE = "performance" BALANCED = "balanced" EFFICIENCY = "efficiency" @dataclass(frozen=True) class Policy: mode: PerformanceMode probe_iterations: int model_parallel_min_param_mb: float # below this, model split is refused allow_strategy_probe: bool # measure DP vs single at startup monitor_active_seconds: float monitor_idle_seconds: float min_dp_batch_per_gpu: int # DP pointless below this per-GPU batch allow_ddp_auto: bool # DDP may win a probe and be selected POLICIES: Dict[PerformanceMode, Policy] = { PerformanceMode.PERFORMANCE: Policy( mode=PerformanceMode.PERFORMANCE, probe_iterations=12, model_parallel_min_param_mb=200.0, allow_strategy_probe=True, monitor_active_seconds=1.0, monitor_idle_seconds=5.0, min_dp_batch_per_gpu=2, allow_ddp_auto=False, # Windows gloo reliability: opt-in only, always ), PerformanceMode.BALANCED: Policy( mode=PerformanceMode.BALANCED, probe_iterations=6, model_parallel_min_param_mb=400.0, allow_strategy_probe=True, monitor_active_seconds=2.0, monitor_idle_seconds=10.0, min_dp_batch_per_gpu=4, allow_ddp_auto=False, ), PerformanceMode.EFFICIENCY: Policy( mode=PerformanceMode.EFFICIENCY, probe_iterations=0, # no probing overhead at all model_parallel_min_param_mb=1e9, allow_strategy_probe=False, monitor_active_seconds=5.0, monitor_idle_seconds=30.0, min_dp_batch_per_gpu=8, allow_ddp_auto=False, ), } def get_policy(mode: str) -> Policy: try: return POLICIES[PerformanceMode(str(mode).lower())] except KeyError: raise ValueError( f"unknown performance mode {mode!r}; valid: {[m.value for m in PerformanceMode]}" )