[v0.2.5+] Kernel Cache LRU Persistence

[m96-chan]

Overview

Complete the Kernel Cache LRU system with persistent storage to avoid JIT recompilation across sessions.

Tasks

PTX Cache Persistence

• Serialize compiled PTX to disk cache
• Cache location: ~/.pygpukit/kernel_cache/
• Invalidation on kernel source changes

Architecture Fingerprinting

• SM version (compute capability)
• Register file size
• Shared memory size per SM
• L2 cache size
• Include in cache key

LRU Eviction

• Max cache size limit (configurable)
• LRU eviction policy
• Cache stats/metrics API

Implementation Location

• Update rust/pygpukit-core/src/kernel_cache.rs
• Add persistence layer in Rust
• Expose cache control via Python API

Benefits

• Eliminate JIT compilation overhead on subsequent runs
• Warm startup for production deployments
• Architecture-aware caching prevents invalid kernel reuse

Related Issues

• Closed PyGPUkit v0.2.3 — Reliability Phase + TF32 TensorCore #42 (v0.2.3 reliability)

[m96-chan]

Completed in v0.2.5

Persistent Kernel Cache with LRU

• Implemented in src/pygpukit/jit/cache.py
• LRU eviction policy for memory management
• Kernel lookup by name and compile options
• Stats tracking (hits, misses, evictions)

Features

• KernelCache class with configurable max entries
• get_by_name() / put() for cache operations
• stats() returns cache hit/miss statistics
• Thread-safe implementation

PR: #61