v1.0.4: Modularized Rust & Python Engine with Free-Threaded No-GIL Support

Release v1.0.4 - Modular Architecture Refactoring, Free-Threaded No-GIL Support & Parity Hardening

LuminaMoon v1.0.4 introduces a comprehensive architectural modularization across the entire workspace (Rust engine crates, Python optimizer package, and canonical reference implementations) while introducing verified support for Free-Threaded No-GIL CPython (PEP 703).

All public APIs, numerical invariants, stochastic rounding distributions, and hardware dispatch parity remain strictly preserved.

---

Key Highlights

1. Free-Threaded CPython (No-GIL / PEP 703) Compatibility

• Added native support and verification for free-threaded Python runtimes (3.14t, 3.15t).
• Automated CI pipeline verification ensuring CPython extensions compile cleanly and run with sys._is_gil_enabled() == False.

2. Rust Workspace Modularization (optimizer/crates/ & optimizer/)

• luminav-core:
  • Decomposed monolithic reference_cpu.rs into specialized submodules: reference_cpu/dual.rs, reference_cpu/tdclm.rs, and reference_cpu/mod.rs.
  • Added dedicated integration test suite tests/core_tests.rs covering PRNG seed determinism, Padé approximation bounds, scratch buffer EMA smoothing, and config error typo suggestions.
• luminav-simd:
  • Restructured avx2.rs, avx512.rs, and parallel.rs into uniform submodule directories (dual.rs, tdclm.rs, mod.rs).
• luminav-vulkan:
  • Extracted UMA dual-buffer and single-buffer TD-CLM compute dispatches from lib.rs into dedicated dual.rs and tdclm.rs execution files.
• luminav-triton-aot:
  • Integrated master_weights mode mapping and buffer mode default ema_decay within c_api.rs.
• python_ffi:
  • Decomposed monolithic optimizer/src/python_ffi.rs into modular components: config_parser.rs, extract.rs, worker.rs, and mod.rs.

3. Python Package Modularization (python/luminamoon/optimizer/)

• luminav.py:
  • Extracted pure standalone optimizer implementation into standalone.py.
  • Extracted tensor and configuration validation utilities into validation.py.
  • Extracted CPU NumPy fallback step execution into numpy_backend.py.
• gpu_backend.py:
  • Extracted distributed FSDP buffer synchronization routines into fsdp_sync.py.
  • Extracted PyTorch tensor math and stochastic rounding operations into torch_ops.py.
  • Extracted Triton GPU JIT kernels into triton_kernels.py.
• version.py:
  • Implemented standard cross-platform cache directory resolution via _resolve_default_cache_root() (LOCALAPPDATA on Windows, XDG_CACHE_HOME on POSIX).
  • Updated remote modular file fetch list.

4. Canonical Reference Implementation Overhaul (optimizer/reference/)

• Replaced the single-file luminav_v153.py (3,996 lines) with an organized reference package:
  • core/: Base optimizer state and parameter validation (optimizer.py, state_manager.py, validation.py).
  • kernels/: Mathematical primitives and update rules (common.py, dual_kernels.py, tdclm_kernels.py).
  • single_backend/: Single-GPU PyTorch implementations (dual_step.py, tdclm_step.py, dual_chunked.py, tdclm_chunked.py, fsdp_pass1.py).
  • triton_backend/: Triton fused GPU kernels (base.py, dual.py, tdclm.py).
  • utils/: Stochastic rounding and distributed communication helpers (fsdp_sync.py, math_sr.py).
  • Added reference documentation in README.md.

---

Verification & Parity Guarantees

• Rust Workspace: cargo test --workspace passed 100% (unit tests, SIMD/AVX differential parity, Vulkan UMA parity, and dispatch routing).
• Python Test Suite: pytest -q passed across all supported Python versions (3.10 through 3.15, including free-threaded variants).
• Format & Linters: Zero clippy warnings with -D warnings, zero pyright type-check errors.

---

Commits Included

• Merge pull request #13 from cloverx-id/development: refactor: modularize Rust engine kernels, Python optimizer backends, and reference implementation
• b6c6a23 ci: fix interpreter target for free-threaded wheel verification
• d72724b chore: ignore scratch files and clean up sync_version script
• 383174b refactor(python): decompose luminav optimizer and gpu backend modules
• 0cf017a refactor(reference): decompose monolithic luminav_v153 implementation
• 1640909 refactor(rust): modularize engine kernels, ffi bindings, and add core tests

v1.0.3

What's Changed

• release(v1.0.3): SIMD finite BF16 clamps, modular engine version decoupling, and manifest sanitization by @cloverxion in #11
• update ci.yaml by @cloverxion in #12

Full Changelog: v1.0.2...v1.0.3

v1.0.2.post2

Overview

PEP 440 post-release patch implementing automated Git tag version synchronization and PyPI file immutability compliance.

Key Changes

• Dynamic Tag Versioning: Integrated tools/sync_version.py into CI/CD build matrix to dynamically synchronize pyproject.toml and wheel metadata from Git tags.
• PyPI Immutability Compliance: Bumped distribution package version to 1.0.2.post2 to comply with PyPI non-reusable filename policy.
• Multi-Platform Support: Built and verified across Ubuntu, Windows, and macOS for Python 3.10–3.14 (including Free-Threaded No-GIL 3.13t/3.14t).

Installation

pip install --upgrade luminamoon

Full Changelog: v1.0.2...v1.0.2.post2

Full Changelog: v1.0.2...v1.0.2.post2

v1.0.2.post1

Overview

PEP 440 post-release patch hardening distribution and automated publishing pipelines.

Changes

• Configured PyPI publish workflow with skip-existing: true to prevent pipeline failures on pre-existing wheel releases.
• Validated multi-platform wheels across Ubuntu, Windows, and macOS on Python 3.10–3.14 (including Free-Threaded No-GIL 3.13t/3.14t).

Installation

pip install --upgrade luminamoon

Full Changelog: v1.0.2...v1.0.2.post1

Full Changelog: v1.0.2...v1.0.2.post1

v1.0.2

Overview

Release v1.0.2 introduces the modular data preprocessing engine (luminamoon.data) and establishes LuminaV v1.5.3 as the default canonical optimizer baseline with dynamic versioning.

Key Changes

Data Preprocessing Engine (luminamoon.data)

• SIMD text normalization and whitespace trimming (AVX-512, AVX2, NEON).
• Dual-layout StringTensor:
  • CPU StringView: 16-byte structure with inlined short strings (<= 12 bytes) and 4-byte prefix comparison.
  • GPU Contiguous Buffers: Flat uint32 offsets and uint8 payload buffer for coalesced VRAM access.
• Zero-copy export to Apache Arrow IPC, Polars Series (to_polars()), NumPy arrays, and PyTorch GPU tensors.
• Fault-tolerant checkpointing with 128-byte dual-slot Ping-Pong metadata and CRC32 integrity validation.
• Deduplication via parallel XXH3-64, MinHash, Shannon entropy filter, and SIMD PII masking.

LuminaV Optimizer

• Standardized default optimizer algorithm version to canonical v1.5.3 reference baseline.
• Dynamic per-instance and class-level versioning via LuminaV.setVersion() and getVersion().
• On-demand version cache resolver (~/.cache/luminamoon/versions/).

Installation

pip install luminamoon==1.0.2

Full Changelog: v1.0.1...v1.0.2

Full Changelog: v1.0.1...v1.0.2

v1.0.1

Overview

Initial multi-platform release of LuminaMoon (luminamoon), providing streaming multimodal data ingestion, low-precision numerical optimization, and formal verification harnesses across Linux, macOS, and Windows.

Key Modules

DataLoader (luminamoon.dataloader)

• Streaming ingestion for Apache Arrow IPC / Feather v2, Apache Parquet, WebDataset TAR, and RFC 4180 CSV/TSV.
• Kernel-level memory mapping with asynchronous HTTP 206 range requests and Hugging Face Hub integration.
• Direct decompression streams via flate2 (gzip), zstd, and lz4_flex.

LuminaV Optimizer (luminamoon.optimizer)

• Fused single-pass parameter update with AVX-512 vectorization, Vulkan Wave32 compute shaders, and Triton AOT compiled kernels.
• Low-precision numerical stability via PRNG stochastic rounding, cautious gradient masking, and trust-region bounding.
• Pure-Python and NumPy fallback runtime (py3-none-any).

Installation

pip install luminamoon==1.0.1

Full Changelog: https://github.com/cloverx-id/luminamoon/commits/v1.0.1

Full Changelog: https://github.com/cloverx-id/luminamoon/commits/v1.0.1