v0.1.4

What's New

• Server benchmarks (RTX 4000 SFF Ada 20GB)
• Multi-language report generation (EN, DE, FR, ES, FA, AR)
• Prioritized server benchmark charts in README
• Fixed npm/GitHub Packages publish workflow

Server Benchmark Results

Model	Throughput	TTFT
GPT-2	407.1 tok/s	4.0ms
Qwen2.5-0.5B	140.7 tok/s	10.9ms
TinyLlama-1.1B	93.0 tok/s	30.6ms
Phi-1.5	78.8 tok/s	37.2ms

v0.1.3

New Features

• Multi-language Support: Added README translations for French, Spanish, Farsi (Persian), and Arabic
• Improved Report Branding: TensorFlow-style header with Keyvan.ai, GitHub, and LinkedIn links
• Language Navigation: Quick language switcher in HTML reports

Fixes

• Fixed GitHub Packages workflow (replaced grep -oP with portable sed)

Publishing

This release will be automatically published to:

• PyPI - pip install kvat[full]
• npm - npm install kvat
• GitHub Packages - @keyvanhardani/kvat

v0.1.2

Fixes

• Context Length Bug Fix: Models with limited max_position_embeddings (like GPT-2's 1024) now properly limit context/output lengths before benchmarking
• Uses AutoConfig to fetch model config without loading full model weights

Improvements

• German README now uses proper umlauts (ü, ö, ä)
• Generated reports include Keyvan.ai marketing footer
• Added GitHub Packages publishing to release workflow

Publishing

This release will be automatically published to:

• PyPI
• npm
• GitHub Packages

v0.1.1 - Auto Context Limiting

What's New

Bug Fixes

• Auto context length limiting: Automatically limits context length to model's max_position_embeddings to prevent CUDA errors
• Fixed GPT-2 with chat-agent profile crash

Improvements

• Cleaner README with better explanation of why kvat exists
• Details sections open by default
• npm badge added

CI/CD

• Auto-publish to npm when releasing
• Version sync between pyproject.toml and package.json

Installation

pip install kvat[full]==0.1.1

Or npm:

npm install kvat@0.1.1

v0.1.0 - Initial Release

KVCache Auto-Tuner v0.1.0

Automatic KV-Cache Optimization for HuggingFace Transformers

Features

• Automatic optimization of cache strategy, attention backend, and dtype
• Built-in profiles: chat-agent, rag, longform, ci-micro
• CLI interface: kvat tune, kvat apply, kvat compare
• Production-ready code snippets and JSON configs
• Markdown & HTML benchmark reports
• CUDA/GPU memory tracking
• Windows & Linux support

Installation

pip install kvat[full]

Quick Start

kvat tune gpt2 --profile ci-micro
kvat tune meta-llama/Llama-3.2-1B --profile chat-agent

Benchmarks

• GPT-2: 124.6 tok/s (RTX 4060)
• Server: 365 tok/s (RTX 4000 Ada)

---

Made from Germany with dedication for the HuggingFace community