`Config`

Config holds every optimization setting. Pass it to wrap() or optimize(). With no Config, MemScale uses Config() — balanced mode with sensible defaults.

from memscale import Config, OptimizationMode
 
config = Config(
    mode=OptimizationMode.BALANCED,
    enable_offloading=True,
    max_cpu_offload_gb=64,
)

`OptimizationMode`

An enum with three values — CONSERVATIVE, BALANCED, AGGRESSIVE. See Optimization Modes.

Fields

Core

Field	Type	Default	Description
`mode`	`OptimizationMode`	`BALANCED`	Optimization aggressiveness.
`enable_checkpointing`	`bool`	`True`	Apply gradient checkpointing.
`enable_offloading`	`bool`	`True`	Apply CPU offloading.
`enable_tiling`	`bool`	`False`	Apply activation tiling (more experimental).
`use_mixed_precision`	`bool`	`False`	Use BF16/FP16 mixed precision.
`use_8bit_optimizer`	`bool`	`False`	Use an 8-bit optimizer (requires `bitsandbytes`).

Memory budget

Field	Type	Default	Description
`max_cpu_offload_gb`	`float \| None`	`None`	Cap on CPU RAM used for offloading. `None` = use available CPU RAM.
`target_gpu_utilization`	`float`	`0.85`	Fraction of GPU memory the engine aims to use.
`force_optimize`	`bool`	`False`	Optimize even when the model fits comfortably (bypasses the smart-skip).

Profiling

Field	Type	Default	Description
`use_static_profiling`	`bool`	`True`	Try `torch.fx` static analysis first.
`use_empirical_fallback`	`bool`	`True`	Fall back to runtime profiling if static fails.
`warmup_steps`	`int`	`2`	Warmup steps for empirical profiling.

Observability

Field	Type	Default	Description
`enable_logging`	`bool`	`True`	Log hardware detection and the optimization plan.
`log_file`	`str \| None`	`None`	Log file path. `None` = stdout only.
`observability_port`	`int \| None`	`None`	If set, expose Prometheus metrics on this port.
`verify_correctness`	`bool`	`False`	Run a baseline comparison to verify correctness (slow).

Experimental

Field	Type	Default	Description
`async_offload`	`bool`	`False`	Enable the tier-aware async CPU offload engine.
`async_config`	`AsyncOffloadConfig`	`None`	Power-user tuning for the async path.
`auto_policy`	`bool`	`False`	Opt in to the v1.2 ML policy Stage 1.

auto_policy defaults to False — the v1.2 ML policy is off unless you opt in. With auto_policy=False, MemScale uses your Config exactly as given, which keeps explicit-config runs bit-for-bit reproducible.

Example: aggressive with an explicit budget

config = Config(
    mode=OptimizationMode.AGGRESSIVE,
    use_mixed_precision=True,
    use_8bit_optimizer=True,
    max_cpu_offload_gb=32,
)
model = wrap(model, config)

detach()Hugging Face