Releases: NagusameCS/HyperTensor
Releases · NagusameCS/HyperTensor
HyperRetro v0.3.6 -- Auto-Certificate + GPU Distill Proven
v0.3.6
New: Auto-Certificate Generation
Every compress run now auto-generates a quality certificate with trust tier, spectral efficiency, jury-proof PPL bounds. Disable with --no-cert.
Gap Closures
- CI now tests hypercore suite (21 tests) on all 9 platforms
- AxiomGauge wired into compress pipeline (--gauge flag)
- version added to hypercore (1.0.0) and hypersort (0.1.0)
- GPU distill pipeline proven end-to-end on SmolLM2-135M
- Certificate system wired into CompressedModel.certificate property
Install
\\�ash
pip install hyperretro==0.3.6
hyperretro setup
\\
HyperRetro v0.3.5 — Comprehensive Documentation
HyperRetro v0.3.5
Complete documentation overhaul
- Full README rewrite — covers all 32 API exports, 13 CLI commands, 4 model backends, 6 kernel backends
- CLI help updated — compress backend flag now lists all 4 backends
- Model registry docstrings — updated to list GGUF + vLLM
- Certificate system documented — with trust tiers and PPL bound formulas
- Benchmark table — current numbers, GPU speedup figure
- Install guide — quickstart, extras, auto-detection rules
- Shield badges — PyPI version, Python versions, license
Install
\\�ash
pip install hyperretro==0.3.5
hyperretro setup
\\
HyperRetro v0.3.4 — GGUF + vLLM Model Backends
HyperRetro v0.3.4
New: 4 Model Backends
- HuggingFace (always available) — AutoModelForCausalLM
- GGUF (new!) — load llama.cpp / Ollama .gguf files directly: \hyperretro compress model.gguf\
- OpenMythos — OpenMythos models
- vLLM — vLLM LLM instances for speculative decode
GGUF Backend Features
- Auto-detects .gguf files by extension
- Extracts config, architecture, hidden size, layers from GGUF metadata
- Indexes all 272+ tensors with shapes (no weight loading until needed)
- Maps common GGUF architectures (llama, mistral, qwen2, phi3, gemma2, 40+)
- \pip install hyperretro[gguf]\ to add GGUF support
Install
\
pip install hyperretro==0.3.4
hyperretro setup
\\
HyperRetro v0.3.3 — GPU Backend, 9.5x CUDA Speedup
HyperRetro v0.3.3
GPU Backend Optimisation
- 9.5x GPU speedup on dual-Q8_0 GEMV vs CPU numpy (RTX 4070: 21.5ms vs 204.9ms at 4096x4096)
- Clean CUDA path — pure-PyTorch GPU tensor ops, no compiler needed
- CUDA kernel source shipped (
csrc/cuda/gemv_dual_q8_0.cu) — compiles with NVCC + host compiler when available - C++ JIT extension (
csrc/gemv_dual_q8_0.cpp) — auto-builds with ninja + C++ compiler - 6-tier backend resolution: cuda_cext → cext → cpu_opt → gpu → torch → numpy
- Suppressed MSVC warnings on Windows without Visual Studio (GPU path works fine)
- ninja from venv auto-detected for JIT builds
Kernel Benchmarks (RTX 4070 Laptop, 4096x4096)
| Backend | Latency | GB/s | Speedup |
|---|---|---|---|
| GPU (torch CUDA) | 21.5ms | 1.8 | — |
| CPU (numpy) | 204.9ms | 0.2 | 9.5x |
Install
pip install hyperretro==0.3.3
hyperretro setup
HyperRetro v0.3.2 — Fix author email
HyperRetro v0.3.2
What's New in v0.3.2
- Fixed author email to nagusamecs@gmail.com (was @proton.me)
- All prior v0.3.1 features included:
- Interactive setup wizard (hyperretro setup)
- 32 public API exports, 13 CLI subcommands
- Certificate system with trust tiers + jury-proof PPL bounds
- Q8_0 quantize/dequantize kernels
- vLLM speculative decode adapter
- AxiomGauge, NativeLinear, RiemannianAdamW, KExpansionScheduler
- Red team attacks (GCG, AutoPrompt, PAIR)
- GRC light distillation
- GGUF, safetensors, HuggingFace export
Install
pip install hyperretro
hyperretro setup
Links
HyperRetro v0.3.1 — Interactive Setup Wizard
HyperRetro v0.3.1
New: Interactive Setup Wizard
One command to get the right installation:
pip install hyperretro
hyperretro setup
The wizard detects what's installed and offers guided options:
- [1] Base only - numpy + safetensors
- [2] + HuggingFace - compress/export HF models, distill, certify
- [3] + vLLM - speculative decode adapter
- [4] Full stack - HF + vLLM + benchmarks + hypercore
- [5] + hypercore - geometric tools (AxiomGauge, RiemannianAdamW)
- [6] Everything - all of the above
Install
pip install hyperretro
pip install "hyperretro[hf]"
pip install "hyperretro[vllm]"
pip install "hyperretro[hf,vllm,bench]"
Links
HyperRetro v0.3.0 — Standalone Geometric LLM Compression Toolkit
HyperRetro v0.3.0
Geometric LLM compression with verifiable quality certificates.
Quick Install
pip install git+https://github.com/NagusameCS/HyperTensor.git#subdirectory=hyperretro
What's Inside
- 32 public API exports
- 12 CLI subcommands
- Certificate system (trust tiers + jury-proof PPL bounds)
- Q8_0 quantize/dequantize kernels
- vLLM speculative decode adapter
- AxiomGauge diagonal gauge optimization
- Native k-space training (NativeLinear, RiemannianAdamW)
- Red team attacks (GCG, AutoPrompt, PAIR)
- GRC light distillation
- GGUF, safetensors, HuggingFace export
- 77/77 tests passing
Full details: https://github.com/NagusameCS/HyperTensor/tree/main/hyperretro
HyperSort v0.1.0
O(1) Instant Sort via Riemannian Comparison Manifold. Based on HyperTensor Geometric Jury framework (Papers I-XVIII). Cross-language: Python, Java, JavaScript.