Skip to content

Changelog

Auto-generated from git tags + commit history. Regenerate with python docs/changelog/generate.py. No CI writes this file.

Unreleased

unreleased

Features

  • feat(onnx): add standalone numpy-only ONNX import/export toolkit
  • feat(integrations): add bi-directional HuggingFace integration module
  • feat(edge): add lightweight C edge runtime skeleton, build system, and model converter
  • feat(compiler): add static graph optimizer with constant folding and IR serialization
  • feat(api): add PyTorch-like high-level Python API layer and 10 examples
  • feat(hub): add model hub CLI, server, registry, storage, and C client

Bug Fixes

  • fix(security): Dilithium S0 sign rejection from NTT domain + zeta typos
  • fix(arc,security): argon2 timing-safe password clobber + ARC inputguard outputdim projection
  • fix(security): S7 version enforcement + S8 symex paths + S9 redteam return
  • fix(security): VM dispatch scrambling + HALT register wipe
  • fix(autodiff): nograd scope zeroes requiresgrad on outputs
  • fix(tensor): matmul of two 1-D vectors returns NULL
  • fix(attention): correct in-place RoPE contamination + NULL sin bail
  • fix(kernel): SNEPPXarchconfigdefault input/outputdim 16 -> 512
  • fix(compress): add SNEPPXCOMPRESSNONE pass-through codec
  • fix(ser): correct malloc/alignedfree mismatch in route+forward
  • fix(hub): correct download card path, total_size, upload overrides, and CLI errors
  • fix(vizmon): fix access violation, DLL export macros, and JSON format in histogram snapshot
  • fix(crypto): correct Dilithium NTT twiddle table and Kyber/DRBG crypto primitives
  • fix: tensor.c header insertion broke BINARYOPF32 macro
  • fix: argon2.c header insertion broke macro definition
  • fix: replace missing tensor.h include and wrong API names in attention_module.c

Build / Tooling

  • build(ninja): auto-resolve x64 MASM to ml64 so 'cmake -B build -G Ninja' works
  • build: add Ninja generator to profiles README
  • build: set Ninja generator in devcontainer config
  • build: add Ninja generator to remaining cmake .. commands
  • build: switch docker-compose to Ninja generator
  • build: switch install.sh and Dockerfile to Ninja generator
  • build: switch install.sh to Ninja generator
  • build: add Ninja generator to shell scripts
  • build: add Ninja generator to Makefile
  • build: add Ninja generator to run_sanitizers.ps1
  • build: add Ninja generator to install.ps1
  • build: switch build.ps1 to Ninja generator
  • build: wire sneppx-format --docs into pre-commit and dev-tools.ps1

Other

  • test: fix RoPE self-attention + causal masking failures
  • test: use INT32 index tensor in embedding grad check
  • test: migrate C/C++ test suite to GoogleTest (135 .c + 6 .cpp)
  • docs(onnx): document ONNX import/export toolkit
  • docs: update READMEs with edge runtime information
  • docs(readme): update README with Model Hub, serving guides, and v1.1.1 highlights
  • docs: update main documentation index for v1.1.1
  • docs: update roadmap with v1.1 completion and model hub status
  • docs(guide): add model hub and serving system guides
  • docs(guide): add model hub and serving system guides
  • Serving system: C control plane + Python engine/client + FastAPI endpoints + example config
  • Fix save/loadhfmodel for modules without named_parameters
  • Add paged attention, MX formats, JIT tracing, MLflow-like observability, hw backends
  • Fix stale summary-matrix rows for layer API and experiment tracking
  • Document hyperparameter-search closure and complete gap analysis
  • Add hyperparameter-search orchestrator with random/grid/halving/bayesian samplers
  • Document edge runtime closure in gap analysis
  • Add mobile/edge runtime with device detection and INT8 inference
  • Document graph compiler closure in gap analysis
  • Add graph compiler with element-wise fusion, tiling and C codegen
  • Document Keras-style layer API and autograd fix in gap analysis
  • Add Keras-style layer API with compile/fit/evaluate/predict/summary
  • Fix broadcast-aware backward for Add/Sub/Mul/Div autograd ops
  • Document experiment/run tracking closure in gap analysis
  • Add experiment/run tracking with metadata and metrics persistence
  • Document ONNX shape inference and QAT closure in gap analysis
  • Add ONNX shape inference and op-schema checks (onnx_check)
  • Document Python canonical binary ONNX bridge in gap analysis
  • Add canonical binary ONNX export/import to Python bindings
  • Mark QAT fake-quant as implemented in gap analysis
  • Fix allocator mismatch in reducegradto_shape error path
  • Add QAT fake_quant op with straight-through estimator backward
  • Fix double-free in autodiff tape backward tests
  • Add ONNX structural validator (SneppXonnxvalidate) with rejection tests
  • Reflect general-graph ONNX export implementation in gap analysis
  • Add general-graph ONNX exporter with multi-node graph test
  • Add ML framework feature gap analysis and roadmap
  • Add real binary ONNX exporter for linear models with round-trip test
  • docs: fix stale Ninja build output paths (build_test/Release subdirs)
  • docs(changelog): record Ninja build switch + Dilithium/Kyber/DRBG fixes
  • docs(quickstart): use release preset + correct ninja build output paths
  • docs(testing): align test dirs/commands to Ninja preset layout
  • docs(agents): note stable Ninja + auto-ml64 for Windows builds
  • docs(build): align build/test instructions to Ninja presets + VS+MASM notes
  • docs: add Ninja generator to remaining README build commands
  • docs: add Ninja generator to remaining doc build commands
  • docs: add Ninja generator to Docs.md build commands
  • docs: add Ninja generator to README build commands
  • docs: add Ninja generator to build.md
  • docs: add Ninja generator to build commands
  • docs: index the five pipeline code walkthroughs
  • docs: add FM code walkthrough
  • docs: add NPE code walkthrough
  • docs: add ARC code walkthrough
  • docs: add SER code walkthrough
  • docs: add HSS code walkthrough
  • docs: add pipeline code-walkthrough overview
  • chore: remove obsolete docs/process_all.py batch annotation script
  • docs: add Layer-1 headers and Doxygen blocks to lib, bindings and standalone tests
  • docs: add Layer-1 headers across examples, benchmarks, samples and tests
  • docs: add Doxygen blocks and Layer-1 headers across drivers and tools
  • docs: add Doxygen blocks and Layer-1 headers across memory and filesystem modules
  • docs: add Doxygen blocks and Layer-1 headers across network modules
  • docs: add Doxygen blocks and Layer-1 headers across all kernel modules
  • docs: add Doxygen blocks and Layer-1 headers across all algorithm modules
  • docs: add Doxygen blocks and Layer-1 headers across all security modules
  • docs: add Doxygen blocks and Layer-1 headers to crypto C sources
  • docs: add Layer-1 header to coverage.h
  • docs: add Doxygen blocks and Layer-1 headers to drivers, model_zoo, network headers
  • docs: add Doxygen blocks and Layer-1 headers to all kernel headers
  • docs: add Doxygen blocks to all architecture headers
  • docs: add Doxygen blocks and Layer-1 headers to all security headers
  • docs: add Doxygen blocks to kyber and aes_gcm headers
  • docs: add commenting standard reference to security architecture
  • docs: add commenting standard reference to C API reference
  • docs: add commenting standard to docs index
  • docs: add commenting standard to contributing guide
  • docs: add commenting standard reference to README
  • docs: add commenting standard reference to doxygen guide
  • docs: add commenting standards reference to architecture guide
  • docs: add commenting standard to development guide
  • docs: add commenting standards to code review checklist
  • docs: add commenting standard reference to style guide
  • docs: add commenting standards reference to index
  • docs: add Layer-1 file headers, API docs, and commenting conventions
  • Bump version to 1.1.1 and add v1.1.1 changelog
  • Add C HTTP REST API, packaging improvements, quickstart, and dev-tools chain
  • Fix unsafe C functions: replace gets/scanf/strcpy/sprintf/alloca with safe alternatives
  • Fix spec: correct tensor header and source file paths
  • Fix fuzz harness: register targets via static init, add NULL target check
  • Phase 4-5: C HTTP server + middleware + API reference docs + CMake integration
  • chore: reorganize documentation system
  • chore: remove all CI/CD — manual-only workflow
  • fix CI: install numpy for CMake, add build step for CodeQL analysis, fix pre-commit deps
  • fix CI workflows: correct action versions, add CodeQL init, fix cross-platform shell compatibility
  • update PR template to tier-based system
  • chore: cleanup project structure

:material-git-compare: View diff vs previous

v1.2.0

2026-07-30

No tagged changes; showing commits up to this tag.

:material-git-compare: View diff vs previous

v1.1.1

2026-08-01

Other

  • Bump version to 1.1.1 and add v1.1.1 changelog
  • Add C HTTP REST API, packaging improvements, quickstart, and dev-tools chain
  • Fix unsafe C functions: replace gets/scanf/strcpy/sprintf/alloca with safe alternatives
  • Fix spec: correct tensor header and source file paths
  • Fix fuzz harness: register targets via static init, add NULL target check
  • Phase 4-5: C HTTP server + middleware + API reference docs + CMake integration
  • chore: reorganize documentation system

:material-git-compare: View diff vs previous

v1.1.0

2026-07-30

Features

  • feat: add serving benchmark; update pyproject.toml with new CLIs; fix continuous_batching decode
  • feat: integrate continuous batching + quantized serving into inference server; fix rlhf_cli
  • feat: add sneppx-eval, sneppx-quantize, sneppx-rlhf CLI entry points
  • feat: add synthetic data loaders for MMLU/GSM8K/HumanEval eval tasks
  • feat: add FP8/INT4 quantized model serving module
  • feat: add continuous batching scheduler for inference serving
  • feat: complete v1.1.0 — LoRA/QLoRA, Eval Harness, FSDP
  • feat: add serving profile to Docker Compose

Bug Fixes

  • fix: replace missing tensor.h include and wrong API names in attention_module.c
  • fix: use .item() for numpy scalar conversion in ultra_trainer test
  • fix: add model_implementations test file
  • fix: correct API calls in LLM/vision test files (model names, create_model args)
  • fix: align quantizedserve with QuantizedLinear API, fix quantmode refs
  • fix: implement DPO/GRPO trainer logprob methods (replaced stubs)

Other

  • chore: remove all CI/CD — manual-only workflow
  • fix CI: install numpy for CMake, add build step for CodeQL analysis, fix pre-commit deps
  • fix CI workflows: correct action versions, add CodeQL init, fix cross-platform shell compatibility
  • update PR template to tier-based system
  • chore: cleanup project structure
  • docs: add contribution framework with tiers, branching strategy, CI workflows
  • Fix test bugs found during regression suite: - testcryptoasm.py: add missing import hashlib - testevalharness.py: update assertion for synthetic data fallback - testkeyvault.py: adapt to auditlogger.py API (FileAuditBackend, AuditAction enums) - tests4network.py: adapt to ddosmitigation.py API (DDoSConfig, checkrequest) - testtrainerv2.py: add missing import numpy as np - auditlogger.py: string-to-enum conversion in log(), hash computation fix in FileAuditBackend.write(), verifychain sort order fix, add clear() - benchmarking.py: fix RNN/transformer weight shape mismatches - testalgowrappers.py: add _test__ = False to prevent pytest collection
  • docs: update AGENTS.md with v1.2-1.3 features and new safe test list
  • test: add tests for continuous batching, quantized serve, and CLI entry points
  • docs: add .nojekyll for GitHub Pages serving
  • test: add batch 3 test files (security, CLI, benchmark, trainer_v3, audit, firewall, DDoS)
  • test: add batch 2 test files (dataloader, advancedops, augmentation, pruning, distillation)
  • test: add 5 missing test files (AMP, gradcheckpoint, tokenizer, autogradops, schedulers) — 60 tests
  • docs: add documentation system with Doxygen config, feature guides, and API references
  • chore: update Arix-Site submodule to b62b5d2 (add Models, Benchmarks, Playground, API Reference)
  • chore: bump version to 1.1.0 and prepare pre-release

:material-git-compare: View diff vs previous

v1.0.0

2026-07-26

Features

  • feat: add 128K context extension and MHA forward pass

Bug Fixes

  • fix: pyproject.toml dependencies format for PEP 621 compatibility

Other

  • chore: update Arix-Site submodule to latest (v1.0.0 site rebuild)
  • chore: update Arix-Site submodule to v1.0.0 site content
  • repo SEO & infrastructure: badges, CITATION.cff, keywords, classifiers, docs refresh
  • docs: comprehensive v1.0.0 documentation refresh
  • v1.0.0: Stable Release
  • 1.8: C + Python integration tests; weights.c buffer overrun fix
  • 1.7: C++ Model Factory with RAII wrappers; modelcardto_json hang fix; 1.6 Python ModelHub
  • 1.5: Python ModelConfig dataclass with JSON serialization, validation, C config conversion, registry integration; added to LlamaConfig, MistralConfig, Qwen2Config, DeepSeekV2Config
  • 1.4: Model Cards - Metadata with JSON serialization, validation, file I/O; unit tests
  • 1.3: Pretrained Weights - WeightCollection API, dtype conversion (f32/f16), INT8 quantization, safetensors/gguf/npz loaders (stubs); unit tests
  • 1.2: Model Registry - Registration, discovery, search, save/load, deprecation; C API with tests
  • 1.1: Model Config Schema - C config API with JSON serialize/parse, presets for LLaMA2/3, Mistral, Qwen2, BERT, ViT, SDXL; unit tests
  • docs: Complete documentation overhaul for v0.5.0
  • Update ROADMAP.md to reflect v0.5.0 completed features
  • Add Python API wrappers for ARC/NPE/FM/Trainer features
  • Add CUDA-accelerated training loop with optimizer bridge
  • Add NPE JIT pipeline with fusion passes, auto-JIT in VM
  • Add FM NCCL distributed sync bridge with callback pattern, tests
  • Wire ARC adversarial training: FGSM/PGD/CW attack injection in training graph, config epsilon, test
  • Add SER learned MLP gater: config flag, 2-layer MLP forward, autodiff subgraph, bindings, tests
  • Enable HSS Blelloch parallel scan by default; update ARCHITECTURE.md
  • Remove remaining CI/CD mentions (cliff.toml, ROADMAP)
  • Remove CI/CD: drop GitHub workflows/dependabot/CODEOWNERS, strip CI mentions from docs and configs
  • Docs: bump v0.9.7.890e, refresh stale content, add release/migration/HSS-training docs

:material-git-compare: View diff vs previous

v0.9.7.890e

2026-07-19

Bug Fixes

  • Fix: HSS backward corruption — layer_norm gamma/beta gradient pointer deref + missing math.h (sqrt)
  • Fix: build lib/internal/*.c + vulkan typo + CI matrix + tape guard

Other

  • Phase E: E2E integration tests + gradient flow tests + CI l0-validation job
  • Phase D: Layer 2 Memory cognitive architecture
  • Phase C: realize 6 stub drivers with reference-compute + opt-in flags + tests

:material-git-compare: View diff vs previous

v0.9.7.890b

2026-07-18

Other

  • Phase B: populate 4 kernel directories (activations, position, algorithms, drivers)

:material-git-compare: View diff vs previous

v0.9.7.890a

2026-07-18

Other

  • Phase A: eliminate last C stubs (vmem evict + FM params/train_graph)
  • Update VERSION with correct LOC/file/test counts
  • Docs: document v0.9.7.890 — real implementations and opt-in Vulkan/TPU/HTTP/ZK backends

:material-git-compare: View diff vs previous

v0.9.7.890

2026-07-18

Features

  • feat: add S8/S9 security layer stubs (verification, hardware, c/asm)

Bug Fixes

  • fix: implement real C networking layer (replace skeletons with working code)
  • fix: critical bugs, security hardening, and code quality improvements
  • fix: replace unsafe strcpy/sprintf/sscanf with bounds-checked snprintf across 5 security modules
  • fix: remove duplicate version param in FastAPI init

Other

  • Build v0.9.7.890: real implementations across formats, kernel, architectures, security, and backends
  • Update files and add new kernel algorithms
  • Rewrite README with accurate stats, correct links, updated architecture tree
  • Create comprehensive skeleton framework for future planned features
  • Update tests for real implementations + fix vmem.c field name
  • Fill in remaining 4 stub files with real implementations
  • Fill in remaining stubs with real implementations (+1,227 lines)
  • Fix autodiff backward regressions: matmul, pow, sum, and add reducegradto_shape
  • Wire CUDA optimizer dispatch for all optimizer types, add LARS enum
  • v0.9.5.937: Fix CUDA autodiff bugs, implement NCCL reduce/reduce_scatter, add RAdam and Schedule-Free AdamW C optimizers
  • Implement all 26 missing autodiff forward+backward ops, add header declarations
  • Fix package name in pyproject.toml, add Python auditlogger/containersecurity/ddos_mitigation modules, fix SyslogBackend default arg and Tuple import
  • Fix autodiff ops.c deduplication, add missing helpers, implement all 25 forward+backward ops
  • Re-enable 13 previously excluded C tests by implementing missing APIs and fixing test mismatches
  • Fix pre-existing C compilation errors (CUDA guards, missing includes, duplicate symbols)
  • Add scripts/README.md documenting sanitizer build/CI scripts
  • Add include/README.md documenting C/C++ header structure and include conventions
  • Add examples/README.md documenting demo programs
  • Add drivers/README.md for CUDA, ROCm, and TPU hardware abstraction layers
  • Add net/README.md documenting topology, socket, RDMA, gRPC, and NCCL layers
  • Add docs/security_layers.txt with deep-dive into all 10 security layers (S0-S9), post-quantum crypto, and compliance mappings
  • Add docs/build.txt (build instructions, options, troubleshooting) and docs/api_quickref.txt (C + Python API reference)
  • Add docs/architecture.txt with system layers, data flow, memory model, distributed arch, and security stack
  • Add tests/README.md documenting 31+ test suites, Python/C/CUDA tests, and excluded tests
  • Add bindings/README.md documenting Python bindings architecture, 8 phases, build, and pure-Python fallback
  • Add security/README.md documenting S0-S9 layers, post-quantum crypto, 4-ring firewall, and MASM routines
  • Add algorithms/README.md covering HSS, SER, ARC, NPE, FM with configs and CUDA kernel details
  • Add kernel/README.md documenting the entire computational substrate (tensor, autodiff, optimizer, CUDA, distributed, quantization)
  • Expand root README.md with full project structure, C networking layer, post-quantum crypto, and build instructions
  • Split bindings.cpp into modular include files
  • docs: comprehensive README for v0.9.5.748 release

:material-git-compare: View diff vs previous

v0.9.5.748

2026-07-15

Features

  • feat: add test_watermark (tests/python)
  • feat: add testsecuritymiddleware (tests/python)
  • feat: add tests5safety (tests/python)
  • feat: add test_onnx (tests/python)
  • feat: add test_nccl (tests/python)
  • feat: add testinferenceserver (tests/python)
  • feat: add test_hparams (tests/python)
  • feat: add testgraphcompiler (tests/python)
  • feat: add test_generation (tests/python)
  • feat: add testexperimenttracker (tests/python)
  • feat: add testdistributedwrapper (tests/python)
  • feat: add testdifferentialprivacy (tests/python)
  • feat: add testcudakernels (tests/python)
  • feat: add test_cuda (tests/python)
  • feat: add testcheckpointmanager (tests/python)
  • feat: add test_autograd (tests/python)
  • feat: add test_adversarial (tests/python)
  • feat: add example (config/training)
  • feat: add watermark (bindings/python/SneppXALG/interfacebindings)
  • feat: add trainerv3 (bindings/python/SneppXALG/interface_bindings)
  • feat: add traincli (bindings/python/SneppXALG/interface_bindings)
  • feat: add tokenizer (bindings/python/SneppXALG/interfacebindings)
  • feat: add servecli (bindings/python/SneppXALG/interface_bindings)
  • feat: add securitymiddleware (bindings/python/SneppXALG/interface_bindings)
  • feat: add s5safety (bindings/python/SneppXALG/interface_bindings)
  • feat: add nccl (bindings/python/SneppXALG/interfacebindings)
  • feat: add inferenceserver (bindings/python/SneppXALG/interface_bindings)
  • feat: add hparams (bindings/python/SneppXALG/interfacebindings)
  • feat: add graphcompiler (bindings/python/SneppXALG/interface_bindings)
  • feat: add generation (bindings/python/SneppXALG/interfacebindings)
  • feat: add experimenttracker (bindings/python/SneppXALG/interface_bindings)
  • feat: add experimentcli (bindings/python/SneppXALG/interface_bindings)
  • feat: add distributedwrapper (bindings/python/SneppXALG/interface_bindings)
  • feat: add differentialprivacy (bindings/python/SneppXALG/interface_bindings)
  • feat: add cudakernels (bindings/python/SneppXALG/interface_bindings)
  • feat: add cudadevice (bindings/python/SneppXALG/interface_bindings)
  • feat: add checkpointmanager (bindings/python/SneppXALG/interface_bindings)
  • feat: add autogradops (bindings/python/SneppXALG/interface_bindings)
  • feat: add autograd (bindings/python/SneppXALG/interfacebindings)
  • feat: add adversarial (bindings/python/SneppXALG/interfacebindings)

Bug Fixes

  • fix(N4): fix autocast duplicate import syntax error
  • fix(N3): Optimizer saves lr, step() works without C backend, autocast is proper context manager
  • fix(N2): add Trainer.train() method and fix LSTM benchmark shape
  • fix(N1): MSELoss returns scalar tensor
  • fix: resolve test import errors in testtrainerv2 and testultratrainer
  • fix(C3): add fallback stub when C extension not available
  • fix(C2): add missing Tensor import and fix matmul in benchmarking.py
  • fix(C1): prevent SimpleTokenizer shadowing from .tokenizer over .data version
  • fix(build): compile CUDA extension sources into existing exported targets via target_sources
  • fix(build): link CUDA extension libs privately to avoid export-set errors
  • fix(build): add CMakeLists.txt for algorithms/*/cuda CUDA extension libraries
  • fix(ci): pin CUDA builds to ubuntu-22.04 for CUDA 12.4 repo compatibility
  • fix(ci): install CUDA toolkit on runner (avoid container checkout break), mark macOS/Windows non-blocking
  • fix(ci): build Linux via CUDA container, add net/distributed CMakeLists, mark macOS/Windows non-blocking
  • fix: guard CUDA-only source files for CPU builds, fix CI workflow
  • fix(ci): install numpy for CMake, fix markdownlint, cuda keyring
  • fix(ci): remove remaining empty with block in security-scan
  • fix(ci): remove submodules recursive (breaks on Arix-Site dir)
  • fix(ci): remove duplicate runs-on, embedded dependabot, and other invalid YAML

Refactor / Internal

  • refactor: update advanced_ops, quantization, C tests
  • refactor: update C source files (checkpoint_reader, dilithium)
  • refactor: update test files and CMakeLists
  • refactor: update init.py exports
  • refactor: update benchmarks/export/distillation
  • refactor: update training/infra modules
  • refactor: update model/hf/vision modules
  • refactor: update core tensor/nn/optim/data modules
  • refactor(ci): restructure jobs, fix deps, make CUDA/security/docs non-blocking

Other

  • Release v0.9.5.748 — Unified Python Bindings (Phases 1-7), CLI commands, version bump
  • Fix CLI entry points (consolescripts), experimentcli.py import time bug, add missing deps, sync pyproject.toml
  • bump: v0.9.4.467
  • firewall: test files (24 tests), fix path traversal detection, lazy SSL context build
  • firewall: CLI flags for serve_cli.py, kwargs safety in firewall constructors
  • firewall: wire into inferenceserver.py middleware, setsecurity accepts firewall config
  • firewall: integrate into SecurityMiddleware with checkfirewall, releaseconcurrent
  • firewall: orchestrator with YAML config, env/CLI overrides, 3-ring dispatch
  • firewall: transport ring with TLS/mTLS, cert pinning, ALPN
  • firewall: application ring with injection filter, path normalization, concurrent limiter
  • firewall: network ring with IP CIDR filtering, rate limiting, port knock
  • firewall: assembly routines for IP match, rate counter, conn track, port knock
  • chore: bump version to 0.9.2.094 across all files
  • ci: mark C/CUDA build jobs non-blocking (pre-existing source compile errors); keep Lint/Security/Docs as real gates
  • chore(ci): add workflow_dispatch for manual runs

:material-git-compare: View diff vs previous

v0.9.0-fix4

2026-07-11

Bug Fixes

  • fix: remove CUDA job (needs Docker Hub secrets not set)

:material-git-compare: View diff vs previous

v0.9.0-fix3

2026-07-11

Bug Fixes

  • fix: wrap secrets check in template expr for job if

:material-git-compare: View diff vs previous

v0.9.0-fix2

2026-07-11

Bug Fixes

  • fix: skip CUDA Docker if Docker Hub secrets not set

:material-git-compare: View diff vs previous

v0.9.0-fix1

2026-07-11

No tagged changes; showing commits up to this tag.

:material-git-compare: View diff vs previous

v0.9.0

2026-07-11

Features

  • feat(packaging): add comprehensive packaging for PyPI, Docker, Conda, CI/CD
  • feat(models): add complete LLaMA, Mistral, Qwen2, DeepSeek V2 architectures
  • feat(vision): add Vision Transformer (ViT), DeiT, Swin, MAE architectures
  • feat: add benchmarking suite + all 11 tests pass
  • feat: massive expansion - distillation, pruning, advanced ops, augmentation, ONNX, model zoo
  • feat: distillation, pruning, advanced ops, augmentation, ONNX export
  • feat(augmentation): comprehensive data augmentation pipeline
  • feat(train): add AMP, gradient checkpointing, advanced DataLoader, UltraTrainer
  • feat(optim): add 7 extra optimizers + 10 schedulers + 6 advanced optimizers
  • feat(tensor): add advanced ops - conv, pooling, RNN, attention, norms, tensor manip
  • feat(core): add SIMD GEMM + tensor expression IR with operator fusion
  • feat(crypto): Dilithium DRBG-based sampling, 10-bit packing, SHA512 challenge
  • feat(zoo): model configs, weight converters, from_pretrained for LLaMA2/3, Mistral, Qwen2, DeepSeek V2 + 49 tests
  • feat(profile): profiler, logger, NVTX markers, compute-sanitizer CI scripts + 13 tests
  • feat(ft): async checkpointing, heartbeat, elastic training, fault tolerance (C + Python) + 23 tests
  • feat(quant): INT8/FP8/AWQ/GPTQ quantization kernels (C, CUDA, Python) + 17 tests
  • feat(py): self-contained Tensor with optional C backend, update exports and pip package config
  • feat(py): implement neural network, optimizers, data, distributed, hf, model, train modules
  • feat(arch): Mamba-2 selective SSM with HiPPO initialization, conv1d, discretized scan
  • feat(arch): FlexAttention block-sparse with mask modulation, multi-modal cross-attention, MoD, gated activations, YaRN NTK-RoPE, ALiBi
  • feat(arch): Differential Attention and Multi-head Latent Attention (DeepSeek MLA style)
  • feat(arch): advanced architectures master header (DifferentialAttn, MLA, FlexAttn, Mamba2, MoD, YaRN, ALiBi)
  • feat(dist): distributed sampler with epoch-based shuffling and gradient accumulation manager
  • feat(dist): distributed checkpoint coordinator with async save and fault tolerance
  • feat(dist): hierarchical all-reduce (NVLink+RDMA) and Top-K gradient compression with error feedback
  • feat(dist): DDP with bucket-based gradient all-reduce and compute overlap
  • feat(dist): expert parallelism all-to-all dispatch and FM distributed communication
  • feat(dist): tensor parallelism row/column split linear with all-reduce
  • feat(dist): pipeline parallelism 1F1B schedule with microbatches
  • feat(dist): ZeRO-1/2/3 optimizer state partitioning and fused AdamW step
  • feat(dist): distributed training config header (ZeRO, pipeline, tensor, expert parallel)
  • feat(dist): NCCL dynamic loading, all-reduce, process group part 2/2
  • feat(dist): NCCL dynamic loading, all-reduce, process group part 1/2
  • feat(dist): NCCL communication primitives header
  • feat(hss): extended SSM step, conv, selective scan kernels
  • feat(arc): PGD/FGSM attacks, gradient obfuscation, smoothing
  • feat(arc): adversarial robustness CUDA header
  • feat(npe): differentiable program execution GPU kernel part 2/2
  • feat(npe): differentiable program execution GPU kernel part 1/2
  • feat(npe): neural VM instruction dispatch CUDA header
  • feat(fm): ring/butterfly all-reduce, gradient quantization part 2/2
  • feat(fm): ring/butterfly all-reduce, gradient quantization part 1/2
  • feat(fm): all-reduce, quantization, federated avg CUDA header
  • feat(ser): fused MoE forward, load balancing loss
  • feat(ser): top-k gating, dispatch, combine kernels
  • feat(ser): top-k gating and fused MoE CUDA header
  • feat(hss): Mamba/S6 selective scan, SSM conv, HiPPO matrix CUDA kernels
  • feat(hss): selective scan, S4, HiPPO CUDA header
  • feat(cuda): RNG API header (Philox, distributions, init schemes)
  • feat(cuda): random number generation kernels part 3/3
  • feat(cuda): random number generation kernels part 2/3
  • feat(cuda): random number generation kernels part 1/3
  • feat(cuda): memory pool/stream/event API header
  • feat(cuda): memory management (pool, streams, events) part 3/3
  • feat(cuda): memory management (pool, streams, events) part 2/3
  • feat(cuda): memory management (pool, streams, events) part 1/3
  • feat(cuda): optimizer header with all step types
  • feat(cuda): fused optimizer kernels part 3/3
  • feat(cuda): fused optimizer kernels part 2/3
  • feat(cuda): fused optimizer kernels part 1/3
  • feat(cuda): autodiff backward header declarations
  • feat(cuda): autodiff backward kernels part 4/4
  • feat(cuda): autodiff backward kernels part 3/4
  • feat(cuda): autodiff backward kernels part 2/4
  • feat(cuda): autodiff backward kernels part 1/4
  • feat(cuda): attention kernels part 5/5
  • feat(cuda): attention kernels part 4/5
  • feat(cuda): attention kernels part 3/5
  • feat(cuda): attention kernels part 2/5
  • feat(cuda): attention kernels part 1/5
  • feat: expand assembly code to 3258 lines with constant-time, speculation-safe, cache-resistant security improvements across 15 files
  • feat: security infrastructure expansion - compliance, PQ crypto, threat intel, IR, automation, ZT, chaos, AI guardrails, supply chain, fuzzing + docs
  • feat(tokenizer): BPE tokenizer with train/encode/decode/save/load
  • feat(autodiff): gradient checkpointing + view-aware storage
  • feat(autodiff): min/max with selective gradient routing
  • feat(autodiff): ref-counted backward lifecycle with ctx cleanup

Bug Fixes

  • fix(docker): make CPU Dockerfile self-contained, fix packages workflow
  • fix(version): bump setup.py to 0.9.0 to match pyproject.toml and tag
  • fix(crypto): Kyber uses DRBG for deterministic noise sampling instead of rand()
  • fix: full build cleanup — fix 8 test files for API changes, stub sha512, fix auditlogger stdarg, fix securewipe ABI, exclude unimplemented tests
  • fix(asm): rewrite poly1305sse.asm with correct 5-limb scalar algorithm, fix 8 ABI signature mismatches in asmexports.h
  • fix(asm): fix 6 ABI signature mismatches in asm_exports.h header
  • fix(asm): fix 5 assembly correctness bugs
  • fix(crypto): fix 7 critical bugs - auth bypasses, memory overflows, X25519, Kyber
  • fix: update metadata to align with sneppx-alg identity
  • fix: update build.ps1 CMake flags to SNEPPX_ prefix
  • fix: rename arixalgo package to SneppXALG, fix Python relative imports
  • fix: sha3 finalize->finish, fm_forward 4 args, m.lib MSVC guard
  • fix: add csignal, time.h, link m for math in kernel lib
  • fix: move sys/prctl.h to linux guard, ifdef s9_extensions _finddata for WIN32
  • fix: s9_extensions io.h guard, blake3 finalize->finish across 4 files
  • fix: use asm/unistd.h for _NR* in seccomp filter
  • fix: memory_hardening sys/syscall.h, hmac.c sha512 func name
  • fix: memory_hardening linux headers, drbg sha512 func name, asm only on MSVC
  • fix: ed25519 umul128 compat + pointisoncurve ordering; fix cache.c aarch64 prfm syntax
  • fix: sign scheme S=r+h*a mod L with proper sc_reduce64
  • fix: fetobytes final p-subtraction via limb comparison
  • fix: fe_mul rewrite with _umul128 128-bit arithmetic
  • fix: point_add formula (X3=E·F, Y3=G·H, T3=E·H, Z3=F·G)
  • fix: point_double formula (Y3=E·G, T3=F·G, Z3=D·H) + identity guard
  • fix: fe_sub per-limb bias constants
  • fix: pointscalarmult cswap mask uint8t->uint64t
  • fix: add two-round carry chain to fetobytes
  • fix: add shebangs to build.sh, test.sh, clean.sh

Other

  • ci(packages): add GitHub Packages workflow for Docker and Python publishing
  • chore: update Arix-Site submodule reference after cleanup
  • test(crypto): add Kyber debug output for pk/sk/shared-secret bytes
  • docs: update AGENTS.md with Phases 5-8, deprecate old lib/python bindings
  • test(py): add 7 Python API test suites (82 tests)
  • test(cuda): GPU kernel test suite part 3/3
  • test(cuda): GPU kernel test suite part 2/3
  • test(cuda): GPU kernel test suite part 1/3
  • docs: update AGENTS.md with Phase 1 CUDA backend overview
  • chore: bump version to 0.8.6 for 12 critical bug fixes across C and asm
  • Fix ASMMASM compiler (ml64.exe) and add asmexports.h header
  • Update README for algo0.8.2: assembly stats, version bump, S0 detail
  • Fix MASM syntax errors: all 15 asm files assemble cleanly
  • Fix assembly bugs across 8 files
  • Remove CI/CD workflows, rebrand releases to SneppX-ALG
  • Full rebrand: ARIXAlgo -> SneppXALG
  • Fix ed25519 intrin.h, entropy_pool stdio.h, clang-tidy target order
  • Fix ASM language, macOS func ptr types, clz64 order, skip broken MSVC tests
  • Fix CI/CD: ASAN Debug build, macOS march, clang-tidy targets, Codecov v4, release changelog
  • algo0.8.0: Infrastructure overhaul — build fixes, CI/CD, dev tooling, project organization
  • Add Python package scaffolding, update Makefile/CI/gitignore/pre-commit [algo0.7.8 infra]
  • Update README for algo0.7.8: Python bindings, PQ benchmarks, stats
  • Bump version to algo0.7.8
  • Add PQ crypto benchmark suite (Kyber, Dilithium, SPHINCS+) [algo0.7.8 #7]
  • Enable Python bindings (pybind11) with linker + naming fixes [algo0.7.8 #6]
  • Fix test projects infrastructure [algo0.7.8 #5]
  • Add docs for 11 new security modules [algo0.7.8 #4]
  • Fix C++ obfuscation layer errors (neuralsecuritycpp) [algo0.7.8 #3]
  • Add CI/CD pipeline with GitHub Actions
  • Add AGENTS.md for AI-assisted development workflow
  • algo0.7.5: S4-S9 security additions - PQ crypto, ASM optimizations, DP, DDoS, container security, fuzzer, leak detector, breakout detection, RLHF safety
  • Fix website link: aixsite -> arixsite.vercel.app
  • Update README for algo0.7: S0-S9 complete, 64,589 total lines, full architecture details
  • algo0.7: Complete S0-S9 security system (21,809 lines) + 64,589 total codebase
  • bump version to algo0.5.4
  • Fix README cd path (arix-algo -> arixalgo), update pyproject.toml version to 0.5.0
  • Update VERSION file to reflect algo0.5 release

  • 81 v0.5.0: ARC/NPE/FM training graphs, trainer fix, NPE training test

  • 80 Fix loss computation in trainer, multi-module train graphs, and tests

  • 79 Full attention training graph, reshape op, trainer fix

  • Add 24 infrastructural files across 7 directories
  • Remove CI/CD concepts
  • README, VERSION, changelog, and 20 infra files
  • Attention, inference, data pipeline, arch improvements
  • Project-wide nomenclature restructuring: renamed all module identifiers to extended descriptive nomenclature for enhanced clarity and semantic precision
  • Multi-head attention + RoPE + KV-cache + batched matmul
  • test(gradient): conv2d finite-difference gradient verification
  • L0.1 audit fixes: include paths, orphaned tests, pre-commit hook, ed25519 bit unpack, docs, gitignore
  • docs: update README with skeleton infrastructure stats, fix links and back-to-top
  • chore: add .gitkeep for papers/figures
  • skeleton: samples directory with basic demo
  • skeleton: generic library — rbtree, hashtable, pqueue, strutil, Rust, Python bindings
  • skeleton: test suites — fuzz, unit, algorithm tests, HSS CUDA
  • skeleton: tools — benchmark runner, CLI, fuzz harness, scripts
  • skeleton: security language bindings — C, C++, C#, Go, Rust, secure allocator, integrity monitor
  • skeleton: kernel internal implementations — autodiff, memory, optimizer, tensor, thread
  • skeleton: checkpoint format, slab alloc, vmem, compression
  • skeleton: network subsystem — socket, RDMA, gRPC, topology
  • skeleton: ROCm and TPU drivers
  • skeleton: CUDA driver interface
  • chore: gitignore target/ and Cargo.lock
  • chore: add target/ to .gitignore (Rust build artifacts)
  • v0.2.0: rewrite all stubs into comprehensive implementations
  • docs: update PROGRESS & README for restructure
  • restructure: remove old src/ tree
  • restructure: move security source to security/
  • restructure: move algorithm sources to algorithms/
  • restructure: move kernel source to kernel/
  • restructure: move public headers to include/arix/
  • T2-T5: implement stubs (inverse, det, conv1d/2d, pool1d/2d, save/load) + 54/56 tests pass, add specs
  • Update README: 50/52 tests, 13 dtypes, 80+ ops
  • T1: creation & shape tests + fixes
  • T0: tensor audit & foundation
  • Fix pre-existing benchmark errors: add seed arg to hssmodelcreate, fix npe instruction initializers
  • Add backward gradient passes for sub, div, neg, pow and extend test coverage
  • Update all GitHub URLs to ammar49-cyber/arixalgo
  • Normalize all contact emails to algoarix@gmail.com
  • Project infrastructure: git config, GitHub templates (no CI/CD), editorconfig, clang-format/tidy, Docker, pre-commit hooks, release scripts, security.txt, GOVERNANCE, DESIGN, STYLEGUIDE, AUTHORS, NEWS, INSTALL, COPYING; fix stale URLs
  • Remove GitHub CI/CD workflow
  • K0 foundation: extend NPE to 32 opcodes, expand autodiff tape/variable, optimizer factories, Python bindings (ARC/NPE/FM/SER), tensor/model/train Python API, checkpoint v2, API docs, benchmarks
  • Remove black background from logo via CSS mix-blend-mode
  • Resize logo to 30%
  • Add ARIX logo to README
  • Add back-to-top link at bottom of README
  • ✨ Supercharged README with emojis, TOC, badges, and visual design
  • Markdown documentation rewrite
  • Phase 8: Training graphs (HSS multi-timestep + SER soft MoE), test suite hardening (edge cases, shared header, build config), benchmark infra (tensor + autodiff), CI workflow, CMakePresets, LTO support, Windows build scripts
  • L3: HSS paper draft (LaTeX)
  • Release system: scripts, VERSION, CHANGELOG, installers

:material-git-compare: View diff vs previous

v0.1.0

2026-06-24

No tagged changes; showing commits up to this tag.

:material-git-compare: View diff vs previous

algo0.9.4.467

2026-07-14

Bug Fixes

  • fix(N4): fix autocast duplicate import syntax error
  • fix(N3): Optimizer saves lr, step() works without C backend, autocast is proper context manager
  • fix(N2): add Trainer.train() method and fix LSTM benchmark shape
  • fix(N1): MSELoss returns scalar tensor
  • fix: resolve test import errors in testtrainerv2 and testultratrainer
  • fix(C3): add fallback stub when C extension not available
  • fix(C2): add missing Tensor import and fix matmul in benchmarking.py
  • fix(C1): prevent SimpleTokenizer shadowing from .tokenizer over .data version

Other

  • bump: v0.9.4.467
  • firewall: test files (24 tests), fix path traversal detection, lazy SSL context build
  • firewall: CLI flags for serve_cli.py, kwargs safety in firewall constructors
  • firewall: wire into inferenceserver.py middleware, setsecurity accepts firewall config
  • firewall: integrate into SecurityMiddleware with checkfirewall, releaseconcurrent
  • firewall: orchestrator with YAML config, env/CLI overrides, 3-ring dispatch
  • firewall: transport ring with TLS/mTLS, cert pinning, ALPN
  • firewall: application ring with injection filter, path normalization, concurrent limiter
  • firewall: network ring with IP CIDR filtering, rate limiting, port knock
  • firewall: assembly routines for IP match, rate counter, conn track, port knock
  • chore: bump version to 0.9.2.094 across all files

:material-git-compare: View diff vs previous

algo0.9.2.094

2026-07-13

Features

  • feat: add test_watermark (tests/python)
  • feat: add testsecuritymiddleware (tests/python)
  • feat: add tests5safety (tests/python)
  • feat: add test_onnx (tests/python)
  • feat: add test_nccl (tests/python)
  • feat: add testinferenceserver (tests/python)
  • feat: add test_hparams (tests/python)
  • feat: add testgraphcompiler (tests/python)
  • feat: add test_generation (tests/python)
  • feat: add testexperimenttracker (tests/python)
  • feat: add testdistributedwrapper (tests/python)
  • feat: add testdifferentialprivacy (tests/python)
  • feat: add testcudakernels (tests/python)
  • feat: add test_cuda (tests/python)
  • feat: add testcheckpointmanager (tests/python)
  • feat: add test_autograd (tests/python)
  • feat: add test_adversarial (tests/python)
  • feat: add example (config/training)
  • feat: add watermark (bindings/python/SneppXALG/interfacebindings)
  • feat: add trainerv3 (bindings/python/SneppXALG/interface_bindings)
  • feat: add traincli (bindings/python/SneppXALG/interface_bindings)
  • feat: add tokenizer (bindings/python/SneppXALG/interfacebindings)
  • feat: add servecli (bindings/python/SneppXALG/interface_bindings)
  • feat: add securitymiddleware (bindings/python/SneppXALG/interface_bindings)
  • feat: add s5safety (bindings/python/SneppXALG/interface_bindings)
  • feat: add nccl (bindings/python/SneppXALG/interfacebindings)
  • feat: add inferenceserver (bindings/python/SneppXALG/interface_bindings)
  • feat: add hparams (bindings/python/SneppXALG/interfacebindings)
  • feat: add graphcompiler (bindings/python/SneppXALG/interface_bindings)
  • feat: add generation (bindings/python/SneppXALG/interfacebindings)
  • feat: add experimenttracker (bindings/python/SneppXALG/interface_bindings)
  • feat: add experimentcli (bindings/python/SneppXALG/interface_bindings)
  • feat: add distributedwrapper (bindings/python/SneppXALG/interface_bindings)
  • feat: add differentialprivacy (bindings/python/SneppXALG/interface_bindings)
  • feat: add cudakernels (bindings/python/SneppXALG/interface_bindings)
  • feat: add cudadevice (bindings/python/SneppXALG/interface_bindings)
  • feat: add checkpointmanager (bindings/python/SneppXALG/interface_bindings)
  • feat: add autogradops (bindings/python/SneppXALG/interface_bindings)
  • feat: add autograd (bindings/python/SneppXALG/interfacebindings)
  • feat: add adversarial (bindings/python/SneppXALG/interfacebindings)

Bug Fixes

  • fix(build): compile CUDA extension sources into existing exported targets via target_sources
  • fix(build): link CUDA extension libs privately to avoid export-set errors
  • fix(build): add CMakeLists.txt for algorithms/*/cuda CUDA extension libraries
  • fix(ci): pin CUDA builds to ubuntu-22.04 for CUDA 12.4 repo compatibility
  • fix(ci): install CUDA toolkit on runner (avoid container checkout break), mark macOS/Windows non-blocking
  • fix(ci): build Linux via CUDA container, add net/distributed CMakeLists, mark macOS/Windows non-blocking
  • fix: guard CUDA-only source files for CPU builds, fix CI workflow
  • fix(ci): install numpy for CMake, fix markdownlint, cuda keyring
  • fix(ci): remove remaining empty with block in security-scan
  • fix(ci): remove submodules recursive (breaks on Arix-Site dir)
  • fix(ci): remove duplicate runs-on, embedded dependabot, and other invalid YAML
  • fix: remove CUDA job (needs Docker Hub secrets not set)
  • fix: wrap secrets check in template expr for job if
  • fix: skip CUDA Docker if Docker Hub secrets not set

Refactor / Internal

  • refactor: update advanced_ops, quantization, C tests
  • refactor: update C source files (checkpoint_reader, dilithium)
  • refactor: update test files and CMakeLists
  • refactor: update init.py exports
  • refactor: update benchmarks/export/distillation
  • refactor: update training/infra modules
  • refactor: update model/hf/vision modules
  • refactor: update core tensor/nn/optim/data modules
  • refactor(ci): restructure jobs, fix deps, make CUDA/security/docs non-blocking

Other

  • ci: mark C/CUDA build jobs non-blocking (pre-existing source compile errors); keep Lint/Security/Docs as real gates
  • chore(ci): add workflow_dispatch for manual runs

:material-git-compare: View diff vs previous

algo0.9.0

2026-07-11

Features

  • feat(packaging): add comprehensive packaging for PyPI, Docker, Conda, CI/CD
  • feat(models): add complete LLaMA, Mistral, Qwen2, DeepSeek V2 architectures
  • feat(vision): add Vision Transformer (ViT), DeiT, Swin, MAE architectures
  • feat: add benchmarking suite + all 11 tests pass
  • feat: massive expansion - distillation, pruning, advanced ops, augmentation, ONNX, model zoo
  • feat: distillation, pruning, advanced ops, augmentation, ONNX export
  • feat(augmentation): comprehensive data augmentation pipeline
  • feat(train): add AMP, gradient checkpointing, advanced DataLoader, UltraTrainer
  • feat(optim): add 7 extra optimizers + 10 schedulers + 6 advanced optimizers
  • feat(tensor): add advanced ops - conv, pooling, RNN, attention, norms, tensor manip
  • feat(core): add SIMD GEMM + tensor expression IR with operator fusion
  • feat(crypto): Dilithium DRBG-based sampling, 10-bit packing, SHA512 challenge
  • feat(zoo): model configs, weight converters, from_pretrained for LLaMA2/3, Mistral, Qwen2, DeepSeek V2 + 49 tests
  • feat(profile): profiler, logger, NVTX markers, compute-sanitizer CI scripts + 13 tests
  • feat(ft): async checkpointing, heartbeat, elastic training, fault tolerance (C + Python) + 23 tests
  • feat(quant): INT8/FP8/AWQ/GPTQ quantization kernels (C, CUDA, Python) + 17 tests
  • feat(py): self-contained Tensor with optional C backend, update exports and pip package config
  • feat(py): implement neural network, optimizers, data, distributed, hf, model, train modules
  • feat(arch): Mamba-2 selective SSM with HiPPO initialization, conv1d, discretized scan
  • feat(arch): FlexAttention block-sparse with mask modulation, multi-modal cross-attention, MoD, gated activations, YaRN NTK-RoPE, ALiBi
  • feat(arch): Differential Attention and Multi-head Latent Attention (DeepSeek MLA style)
  • feat(arch): advanced architectures master header (DifferentialAttn, MLA, FlexAttn, Mamba2, MoD, YaRN, ALiBi)
  • feat(dist): distributed sampler with epoch-based shuffling and gradient accumulation manager
  • feat(dist): distributed checkpoint coordinator with async save and fault tolerance
  • feat(dist): hierarchical all-reduce (NVLink+RDMA) and Top-K gradient compression with error feedback
  • feat(dist): DDP with bucket-based gradient all-reduce and compute overlap
  • feat(dist): expert parallelism all-to-all dispatch and FM distributed communication
  • feat(dist): tensor parallelism row/column split linear with all-reduce
  • feat(dist): pipeline parallelism 1F1B schedule with microbatches
  • feat(dist): ZeRO-1/2/3 optimizer state partitioning and fused AdamW step
  • feat(dist): distributed training config header (ZeRO, pipeline, tensor, expert parallel)
  • feat(dist): NCCL dynamic loading, all-reduce, process group part 2/2
  • feat(dist): NCCL dynamic loading, all-reduce, process group part 1/2
  • feat(dist): NCCL communication primitives header
  • feat(hss): extended SSM step, conv, selective scan kernels
  • feat(arc): PGD/FGSM attacks, gradient obfuscation, smoothing
  • feat(arc): adversarial robustness CUDA header
  • feat(npe): differentiable program execution GPU kernel part 2/2
  • feat(npe): differentiable program execution GPU kernel part 1/2
  • feat(npe): neural VM instruction dispatch CUDA header
  • feat(fm): ring/butterfly all-reduce, gradient quantization part 2/2
  • feat(fm): ring/butterfly all-reduce, gradient quantization part 1/2
  • feat(fm): all-reduce, quantization, federated avg CUDA header
  • feat(ser): fused MoE forward, load balancing loss
  • feat(ser): top-k gating, dispatch, combine kernels
  • feat(ser): top-k gating and fused MoE CUDA header
  • feat(hss): Mamba/S6 selective scan, SSM conv, HiPPO matrix CUDA kernels
  • feat(hss): selective scan, S4, HiPPO CUDA header
  • feat(cuda): RNG API header (Philox, distributions, init schemes)
  • feat(cuda): random number generation kernels part 3/3
  • feat(cuda): random number generation kernels part 2/3
  • feat(cuda): random number generation kernels part 1/3
  • feat(cuda): memory pool/stream/event API header
  • feat(cuda): memory management (pool, streams, events) part 3/3
  • feat(cuda): memory management (pool, streams, events) part 2/3
  • feat(cuda): memory management (pool, streams, events) part 1/3
  • feat(cuda): optimizer header with all step types
  • feat(cuda): fused optimizer kernels part 3/3
  • feat(cuda): fused optimizer kernels part 2/3
  • feat(cuda): fused optimizer kernels part 1/3
  • feat(cuda): autodiff backward header declarations
  • feat(cuda): autodiff backward kernels part 4/4
  • feat(cuda): autodiff backward kernels part 3/4
  • feat(cuda): autodiff backward kernels part 2/4
  • feat(cuda): autodiff backward kernels part 1/4
  • feat(cuda): attention kernels part 5/5
  • feat(cuda): attention kernels part 4/5
  • feat(cuda): attention kernels part 3/5
  • feat(cuda): attention kernels part 2/5
  • feat(cuda): attention kernels part 1/5
  • feat: expand assembly code to 3258 lines with constant-time, speculation-safe, cache-resistant security improvements across 15 files
  • feat: security infrastructure expansion - compliance, PQ crypto, threat intel, IR, automation, ZT, chaos, AI guardrails, supply chain, fuzzing + docs

Bug Fixes

  • fix(docker): make CPU Dockerfile self-contained, fix packages workflow
  • fix(version): bump setup.py to 0.9.0 to match pyproject.toml and tag
  • fix(crypto): Kyber uses DRBG for deterministic noise sampling instead of rand()
  • fix: full build cleanup — fix 8 test files for API changes, stub sha512, fix auditlogger stdarg, fix securewipe ABI, exclude unimplemented tests
  • fix(asm): rewrite poly1305sse.asm with correct 5-limb scalar algorithm, fix 8 ABI signature mismatches in asmexports.h
  • fix(asm): fix 6 ABI signature mismatches in asm_exports.h header
  • fix(asm): fix 5 assembly correctness bugs
  • fix(crypto): fix 7 critical bugs - auth bypasses, memory overflows, X25519, Kyber
  • fix: update metadata to align with sneppx-alg identity
  • fix: update build.ps1 CMake flags to SNEPPX_ prefix
  • fix: rename arixalgo package to SneppXALG, fix Python relative imports
  • fix: sha3 finalize->finish, fm_forward 4 args, m.lib MSVC guard
  • fix: add csignal, time.h, link m for math in kernel lib
  • fix: move sys/prctl.h to linux guard, ifdef s9_extensions _finddata for WIN32
  • fix: s9_extensions io.h guard, blake3 finalize->finish across 4 files
  • fix: use asm/unistd.h for _NR* in seccomp filter
  • fix: memory_hardening sys/syscall.h, hmac.c sha512 func name
  • fix: memory_hardening linux headers, drbg sha512 func name, asm only on MSVC
  • fix: ed25519 umul128 compat + pointisoncurve ordering; fix cache.c aarch64 prfm syntax

Other

  • ci(packages): add GitHub Packages workflow for Docker and Python publishing
  • chore: update Arix-Site submodule reference after cleanup
  • test(crypto): add Kyber debug output for pk/sk/shared-secret bytes
  • docs: update AGENTS.md with Phases 5-8, deprecate old lib/python bindings
  • test(py): add 7 Python API test suites (82 tests)
  • test(cuda): GPU kernel test suite part 3/3
  • test(cuda): GPU kernel test suite part 2/3
  • test(cuda): GPU kernel test suite part 1/3
  • docs: update AGENTS.md with Phase 1 CUDA backend overview
  • chore: bump version to 0.8.6 for 12 critical bug fixes across C and asm
  • Fix ASMMASM compiler (ml64.exe) and add asmexports.h header
  • Update README for algo0.8.2: assembly stats, version bump, S0 detail
  • Fix MASM syntax errors: all 15 asm files assemble cleanly
  • Fix assembly bugs across 8 files
  • Remove CI/CD workflows, rebrand releases to SneppX-ALG
  • Full rebrand: ARIXAlgo -> SneppXALG
  • Fix ed25519 intrin.h, entropy_pool stdio.h, clang-tidy target order
  • Fix ASM language, macOS func ptr types, clz64 order, skip broken MSVC tests
  • Fix CI/CD: ASAN Debug build, macOS march, clang-tidy targets, Codecov v4, release changelog

:material-git-compare: View diff vs previous

algo0.8.0

2026-07-08

Other

  • algo0.8.0: Infrastructure overhaul — build fixes, CI/CD, dev tooling, project organization
  • Add Python package scaffolding, update Makefile/CI/gitignore/pre-commit [algo0.7.8 infra]
  • Update README for algo0.7.8: Python bindings, PQ benchmarks, stats

:material-git-compare: View diff vs previous

algo0.7.8

2026-07-08

Other

  • Bump version to algo0.7.8
  • Add PQ crypto benchmark suite (Kyber, Dilithium, SPHINCS+) [algo0.7.8 #7]
  • Enable Python bindings (pybind11) with linker + naming fixes [algo0.7.8 #6]
  • Fix test projects infrastructure [algo0.7.8 #5]
  • Add docs for 11 new security modules [algo0.7.8 #4]
  • Fix C++ obfuscation layer errors (neuralsecuritycpp) [algo0.7.8 #3]
  • Add CI/CD pipeline with GitHub Actions
  • Add AGENTS.md for AI-assisted development workflow

:material-git-compare: View diff vs previous

algo0.7.5

2026-07-07

Other

  • algo0.7.5: S4-S9 security additions - PQ crypto, ASM optimizations, DP, DDoS, container security, fuzzer, leak detector, breakout detection, RLHF safety
  • Fix website link: aixsite -> arixsite.vercel.app
  • Update README for algo0.7: S0-S9 complete, 64,589 total lines, full architecture details
  • algo0.7: Complete S0-S9 security system (21,809 lines) + 64,589 total codebase
  • bump version to algo0.5.4
  • Fix README cd path (arix-algo -> arixalgo), update pyproject.toml version to 0.5.0
  • Update VERSION file to reflect algo0.5 release

:material-git-compare: View diff vs previous

algo0.5

2026-07-01

Features

  • feat(tokenizer): BPE tokenizer with train/encode/decode/save/load
  • feat(autodiff): gradient checkpointing + view-aware storage
  • feat(autodiff): min/max with selective gradient routing
  • feat(autodiff): ref-counted backward lifecycle with ctx cleanup

Bug Fixes

  • fix: sign scheme S=r+h*a mod L with proper sc_reduce64
  • fix: fetobytes final p-subtraction via limb comparison
  • fix: fe_mul rewrite with _umul128 128-bit arithmetic
  • fix: point_add formula (X3=E·F, Y3=G·H, T3=E·H, Z3=F·G)
  • fix: point_double formula (Y3=E·G, T3=F·G, Z3=D·H) + identity guard
  • fix: fe_sub per-limb bias constants
  • fix: pointscalarmult cswap mask uint8t->uint64t
  • fix: add two-round carry chain to fetobytes
  • fix: add shebangs to build.sh, test.sh, clean.sh

Other

  • 81 v0.5.0: ARC/NPE/FM training graphs, trainer fix, NPE training test

  • 80 Fix loss computation in trainer, multi-module train graphs, and tests

  • 79 Full attention training graph, reshape op, trainer fix

  • Add 24 infrastructural files across 7 directories
  • Remove CI/CD concepts
  • README, VERSION, changelog, and 20 infra files
  • Attention, inference, data pipeline, arch improvements
  • Project-wide nomenclature restructuring: renamed all module identifiers to extended descriptive nomenclature for enhanced clarity and semantic precision
  • Multi-head attention + RoPE + KV-cache + batched matmul
  • test(gradient): conv2d finite-difference gradient verification
  • L0.1 audit fixes: include paths, orphaned tests, pre-commit hook, ed25519 bit unpack, docs, gitignore
  • docs: update README with skeleton infrastructure stats, fix links and back-to-top
  • chore: add .gitkeep for papers/figures
  • skeleton: samples directory with basic demo
  • skeleton: generic library — rbtree, hashtable, pqueue, strutil, Rust, Python bindings
  • skeleton: test suites — fuzz, unit, algorithm tests, HSS CUDA
  • skeleton: tools — benchmark runner, CLI, fuzz harness, scripts
  • skeleton: security language bindings — C, C++, C#, Go, Rust, secure allocator, integrity monitor
  • skeleton: kernel internal implementations — autodiff, memory, optimizer, tensor, thread
  • skeleton: checkpoint format, slab alloc, vmem, compression
  • skeleton: network subsystem — socket, RDMA, gRPC, topology
  • skeleton: ROCm and TPU drivers
  • skeleton: CUDA driver interface
  • chore: gitignore target/ and Cargo.lock
  • chore: add target/ to .gitignore (Rust build artifacts)
  • v0.2.0: rewrite all stubs into comprehensive implementations
  • docs: update PROGRESS & README for restructure
  • restructure: remove old src/ tree
  • restructure: move security source to security/
  • restructure: move algorithm sources to algorithms/
  • restructure: move kernel source to kernel/
  • restructure: move public headers to include/arix/
  • T2-T5: implement stubs (inverse, det, conv1d/2d, pool1d/2d, save/load) + 54/56 tests pass, add specs
  • Update README: 50/52 tests, 13 dtypes, 80+ ops
  • T1: creation & shape tests + fixes
  • T0: tensor audit & foundation
  • Fix pre-existing benchmark errors: add seed arg to hssmodelcreate, fix npe instruction initializers
  • Add backward gradient passes for sub, div, neg, pow and extend test coverage
  • Update all GitHub URLs to ammar49-cyber/arixalgo
  • Normalize all contact emails to algoarix@gmail.com
  • Project infrastructure: git config, GitHub templates (no CI/CD), editorconfig, clang-format/tidy, Docker, pre-commit hooks, release scripts, security.txt, GOVERNANCE, DESIGN, STYLEGUIDE, AUTHORS, NEWS, INSTALL, COPYING; fix stale URLs
  • Remove GitHub CI/CD workflow
  • K0 foundation: extend NPE to 32 opcodes, expand autodiff tape/variable, optimizer factories, Python bindings (ARC/NPE/FM/SER), tensor/model/train Python API, checkpoint v2, API docs, benchmarks
  • Remove black background from logo via CSS mix-blend-mode
  • Resize logo to 30%
  • Add ARIX logo to README
  • Add back-to-top link at bottom of README
  • ✨ Supercharged README with emojis, TOC, badges, and visual design
  • Markdown documentation rewrite
  • Phase 8: Training graphs (HSS multi-timestep + SER soft MoE), test suite hardening (edge cases, shared header, build config), benchmark infra (tensor + autodiff), CI workflow, CMakePresets, LTO support, Windows build scripts
  • L3: HSS paper draft (LaTeX)
  • Release system: scripts, VERSION, CHANGELOG, installers
  • Add release artifacts to gitignore
  • Update docs with new vision: verifiable inference, on-device, federated contribution, safety guarantees
  • S2: obfuscation engine (C++) — CFG flattening, string encryption, instruction substitution, opaque predicates, code VM, anti-debug, pipeline, tests, demo
  • Add x86_64 assembly: ed25519 scalarmult, constant-time cmp, SC cmov ops
  • Phase 7: S1 - Secure Memory & Side-Channel Resistance
  • Phase 7: Security - AES-free crypto suite (ChaCha20-Poly1305 AEAD, Ed25519, Argon2, SHA-3, BLAKE3)
  • Phase 6: Training Loop & Python Bindings
  • Phase 5: FM - Federated Memory
  • Phase 4: NPE - Neural Program Executor
  • Phase 3: ARC - Adversarial Robustness Core - input guard with anomaly detection, gradient obfuscator with noise+clamp, output verifier with consistency history, attack simulation (FGSM/PGD/CW), multi-layer model, tests, integration with HSS+SER, demo
  • Phase 2: SER - Sparse Expert Routing - expert create/destroy, routing with softmax+top-k, expert forward (ReLU/GELU/Swish), layer forward with gather/scatter, load balance loss, multi-layer model, tests, integration test with HSS, demo
  • Phase 1: HSS forward pass - layer create/destroy, model create/destroy, discretization, single step, seq scan, hierarchical scan stub, forward pass with layer norm and input projection, test suite, demo
  • Add .gitignore, remove build artifacts from tracking
  • Fix CMakeLists.txt for MSVC compatibility, update README
  • Initial foundation: tensor ops, memory allocator, thread pool stub, CMake build system, C test suite, Python/Rust/C security stubs, MIT license
  • Initial commit

:material-git-compare: View diff vs previous

Each release header links to a GitHub compare view ({prev}...{tag}) which renders a side-by-side visual diff of the source tree between releases. Tag the repo (git tag v1.2.0) and re-run python docs/changelog/generate.py to refresh.