ejkernel.ops.config.persistent

ejkernel.ops.config.persistent#

Persistent disk-based configuration cache for kernel optimization.

This module provides persistent storage for optimal kernel configurations, allowing optimization results to be preserved across program runs. The cache uses JSON files for storage with automatic serialization/deserialization.

Key Features:

  • Atomic file operations to prevent corruption

  • Automatic serialization for dataclasses and Pydantic models

  • Custom loader/dumper support for complex types

  • Device and operation-specific configuration storage

  • Thread-safe atomic updates using temporary files

The persistent cache complements the in-memory cache by providing:

  • Long-term storage of optimization results

  • Sharing of configurations across program runs

  • Reduced autotuning overhead for repeated operations

  • Backup storage for critical optimization data

Example Usage:
>>>
>>> cache = PersistentCache('/path/to/cache.json')
>>> cache.put('gpu:0', 'matmul_v1', 'key123', my_config)
>>> config = cache.get('gpu:0', 'matmul_v1', 'key123')
>>>
>>>
>>> def custom_dumper(cfg): return cfg.to_dict()
>>> def custom_loader(data): return MyConfig.from_dict(data)
>>> cache = PersistentCache('/path/to/cache.json', custom_loader, custom_dumper)
class ejkernel.ops.config.persistent.PersistentCache(opname: str, path: str | None = None, loader: collections.abc.Callable[[Any], ejkernel.ops.config.persistent.Cfg] | None = None, dumper: collections.abc.Callable[[ejkernel.ops.config.persistent.Cfg], Any] | None = None, cfg_type: type[Cfg] | None = None)[source]#

Bases: Generic[Cfg]

Persistent disk-based cache for kernel configurations.

Provides thread-safe storage and retrieval of optimization configurations using JSON files. Supports automatic serialization for common types and custom serialization via loader/dumper functions.

The cache uses atomic file operations to ensure data consistency and prevent corruption during concurrent access.

Type Parameters:

Cfg: Configuration type to be cached

path#

File path for the JSON cache

loader#

Optional function to deserialize configurations

dumper#

Optional function to serialize configurations

get(device: str, op_id: str, call_key: str) → Optional[Cfg][source]#

Retrieve cached configuration for the given coordinates.

Parameters

  • device – Device fingerprint for the configuration

  • op_id – Operation identifier

  • call_key – Function call signature hash

Returns

Cached configuration if found, None otherwise

Note

If a custom loader was provided, it will be used to deserialize the stored data. Otherwise, the raw JSON data is returned.

put(device: str, op_id: str, call_key: str, cfg: Cfg)[source]#

Store configuration in the cache with atomic file update.

Parameters

  • device – Device fingerprint for the configuration

  • op_id – Operation identifier

  • call_key – Function call signature hash

  • cfg – Configuration to store

Note

Uses atomic file operations (write to temporary file, then replace) to ensure data consistency. If a custom dumper was provided, it will be used for serialization. Otherwise, automatic serialization is attempted for dataclasses and Pydantic models.

Serialization Priority:

  1. Custom dumper function (if provided)

  2. Dataclass.asdict() for dataclass objects

  3. model_dump() for Pydantic v2 models

  4. Raw value (must be JSON-serializable)

Contents