Releases: codepawl/turboquant-torch
Releases · codepawl/turboquant-torch
v0.4.0
v0.3.0
v0.2.2
v0.2.1
Full Changelog: v0.2.0...v0.2.1
v0.2.0
What's New
Features
- HuggingFace Integration Demo (
examples/huggingface_demo.py): End-to-end KV cache compression on real SmolLM2-135M model with multi-bitwidth benchmarking, DynamicCache support, attention MSE comparison, and generation quality comparison - Benchmark Suite (
benchmarks/run_benchmarks.py): Real model + synthetic configs (Llama-7B at 2K/8K/32K ctx, Llama-70B at 2K ctx), outputs JSON + formatted markdown tables - Visual Benchmark Cards: Card D (real model benchmark table) and Card E (KV cache memory at scale chart) — 1200x675 PNGs for X posts
Fixes
- Fix README image paths to absolute GitHub raw URLs for proper PyPI rendering
- Fix CI badge and repo links to correct GitHub org
- Add TestPyPI badge to README
- Add staging branch to CI triggers
Chores
- Refine
.gitignorewith IDE, OS, and benchmark data patterns - Remove
RELEASING.md(info consolidated inCLAUDE.md) - Add
benchmarks/to ruff check/format paths in CI
Full Changelog: v0.1.1...v0.2.0
v0.1.1
v0.1.1
- Replace assert statements with ValueError for production safety
- Suppress CVE-2026-4539 (pygments ReDoS, no upstream patch)
- Migrate CI from pip to uv for faster builds
- Add Python 3.13 and 3.14 to test matrix
- Automatic versioning via setuptools-scm (no manual version edits)
- Fix shallow clone issue for setuptools-scm in CI
- Drop twine version pin to fix metadata validation
- 94% test coverage across Python 3.10–3.14