[v0.2.5+] JIT Compiler Stabilization (NVRTC Robustness)

[m96-chan]

Overview

Improve NVRTC JIT compiler robustness with error handling, retry logic, and fallback paths.

Tasks

NVRTC Error Handling

• Comprehensive error code mapping
• Retry logic for transient failures
• Detailed error messages with context
• Log compilation errors for debugging

JIT Warmup Cache

• Pre-compile common kernels on first import
• Background compilation thread
• Warmup progress indicator (optional)

PTX ISA Mismatch Fallback

• Detect PTX version vs driver capability mismatch
• Fallback to compatible PTX version
• Warning when using fallback path
• Document minimum driver versions

Implementation Location

• native/jit/ - C++ NVRTC wrapper
• rust/pygpukit-core/ - Rust coordination

Related Issues

• Closed PyGPUkit v0.2.3 — Reliability Phase + TF32 TensorCore #42 (v0.2.3 reliability)
• [v0.2.4] Make NVRTC Optional (No Bundling) #50, [v0.3+] test: Verify Driver-Only Runtime on Clean Systems #51, [v0.2.4] Graceful Error Handling for Missing GPU Components #52 (v0.2.4 driver-only mode)

Notes

This is separate from v0.2.4's goal of shipping NVRTC DLLs. This issue focuses on making the JIT compiler more robust regardless of how NVRTC is loaded.

[m96-chan]

Completed in v0.2.5

JIT Compiler Stabilization

• Improved NVRTC error handling and reporting
• Architecture fallback chain (sm_90 → sm_89 → sm_86 → sm_80)
• Better driver compatibility checks
• Warmup mechanism for reliable initialization

Robustness Improvements

• check_driver_compatibility() for version validation
• get_driver_requirements() returns minimum versions
• Graceful fallback when NVRTC unavailable
• Clear error messages for common issues

Integration with Kernel Cache

• Combined with LRU cache ([v0.2.5+] Kernel Cache LRU Persistence #54) for efficient kernel reuse
• Cached PTX avoids recompilation overhead

PR: #61