SneppX-ALG v0.9.0 is now released. This is the largest single update to the project, spanning 8 completed phases with 171K+ lines of code across 500+ files. The release adds production-grade packaging — Docker images on GitHub Container Registry, an npm package for UI components, and a PyPI wheel.
What's New
Quantization & Compression (Phase 5)
INT8 symmetric/asymmetric, INT4 packed, FP8 E4M3/E5M2
AWQ optimal scale search
GPTQ with Cholesky-based inverse Hessian
CUDA INT8/FP8 kernels with shared memory reduction and Hopper native FP8
Python `QuantizedLinear` with all modes
Async Checkpointing & Fault Tolerance (Phase 6)
Real binary checkpoint format with file I/O
Double-buffered async D2D→D2H checkpoint saves
UDP heartbeat with SUSPECT→DEAD state machine
Elastic training with join/leave/failure/reconfigure lifecycle
Python `CheckpointWriter/Reader`, `Coordinator`, `HeartbeatMonitor`, `ElasticTrainer`
Profiling & Debugging (Phase 7)
C profiler with named aggregation and JSON export
6-severity structured logger with color stdout + JSON file
Real NVTX range push/pop with software fallback
Python `Profiler`, `@timeit`, `MemoryTracker`, `TrainProfiler`
Model Zoo (Phase 8)
LLaMA 2 (7B/13B/70B), LLaMA 3 (8B/70B)
Mistral (7B with sliding window), Qwen2 (7B/72B)
DeepSeek V2 (lite/full with MLA)
Vision Transformers (ViT, DeiT, Swin, MAE)
`from_pretrained()` with automatic HF safetensors → SneppX conversion
Additional Modules
SIMD GEMM (AVX2/AVX-512/SSE4.2)
Tensor expression IR with fusion planner
14 advanced optimizers (Lion, LAMB, Sophia, Adan, RAdam, etc.)
14 schedulers (OneCycle, CosineWarmRestarts, TriStage, etc.)
AMP/GradScaler, gradient checkpointing, advanced DataLoader
Mixup/CutMix/AugMix, SpecAugment, RandomErasing
Pruning (magnitude/Taylor/movement/lottery ticket)
Distillation (KD/AT/FM/CRD, multi-teacher)
ONNX export/import, benchmarking suite
Packaging & Distribution
**Docker**: Multi-stage images (CPU, CUDA, dev, py-runtime) on ghcr.io
**npm**: `@ammar49-cyber/arix-ui` — reusable UI components
**PyPI**: `sneppx-alg` wheel (714 KB, pure Python + NumPy)
**GitHub Actions**: Full CI/CD with lint, build (Linux/macOS/Windows), test, release, Docker
**Dependabot**: Auto-updates for pip, cargo, cmake, docker, github-actions
By the Numbers
171,000+ lines of code
500+ source files
198+ commits
11 test suites, 172 tests (all passing)
11/11 test suites: tensor, nn, optim, data, distributed, hf_integration, train, quantization, checkpoint, profiler, model_zoo
Quick Start
# pip
pip install sneppx-alg
# Docker (CPU)
docker pull ghcr.io/ammar49-cyber/sneppx-alg:cpu-latest
# Docker (CUDA)
docker pull ghcr.io/ammar49-cyber/sneppx-alg:cuda-latest
# npm
npm install @ammar49-cyber/arix-ui
# Source
git clone https://github.com/ammar49-cyber/sneppx-alg.git
See the Packages page for full installation details.