Quantica v0.2.1

[gurukasi-2006]

Quantica JIT Compiler - Patch Notes v0.2.1

"The Stability Update"

Core Compiler & JIT Backend (src/codegen/mod.rs)

• Fixed len() Intrinsic Dispatch
  • Bug: The compiler previously treated Strings as Arrays, attempting to read length headers from raw string pointers, which resulted in reading garbage memory (negative lengths).
  • Fix: Updated mod.rs to check the argument type. It now correctly dispatches to quantica_rt_string_len for Strings and quantica_rt_array_len for Arrays.
• Implemented != (Not Equal) Operator
  • Feature: Added backend support for the != comparison operator, enabling logic such as if val != 0.0.

AI Library (stdlib/ai.qc)

• Fixed Model Saving Crash
  • Bug: The original save() method used massive string concatenation (str = str + val) to build CSV rows. This created thousands of temporary string objects, leading to memory exhaustion and crashes.
  • Fix: Rewrote save() to use an incremental flushing strategy via file_append(). Values are now written to disk one by one, keeping memory footprint minimal.
• Fixed NaN / Division by Zero
  • Bug: Training on empty or malformed datasets caused division by zero in accuracy calculations.
  • Fix: Added guard clauses in evaluate_network to check if dataset.size == 0 before computing accuracy.
• Stabilized Gradient Descent
  • Optimization: Verified and stabilized matrix_update and backpropagation logic.

Runtime Environment (src/runtime.rs)

• Implemented file_append
  • Feature: Added quantica_rt_file_append to support the new incremental model saving strategy.
• Robust File I/O
  • Fix: Updated quantica_rt_file_read to print clear error messages to stderr instead of failing silently when a file is not found.
• Stable Memory Model
  • Adjustment: Standardized all array allocations to use a hidden 8-byte header layout. This ensures compatibility between compiled JIT code and Runtime helper functions.

Known Issues & Design Decisions

• Memory Leaks (By Design):
  • To prevent STATUS_HEAP_CORRUPTION and maintain high stability during this development phase, manual memory deallocation (free) has been disabled.
  • Impact: Long-running training jobs (e.g., >10,000 epochs) will consume increasing amounts of RAM. This is a known trade-off for current stability.

Current Status

The engine is Stable. It successfully performs the entire machine learning pipeline:

---

This discussion was created from the release Quantica v0.2.1.