VRAMScout v0.7.0

VRAMScout v0.7.0

VRAMScout is now a composable, validation-backed deployment-memory preflight tool for modern open-weight LLMs.

Highlights

• Refactored cache/state accounting into reusable memory components instead of one monolithic formula per model family.
• Covers current families including Qwen3.8 / Flash-Next, DeepSeek-V4, GLM-5.2/5.3, Kimi-K2.x/K3, Hy4, MiniMax-M2.7/M3, Gemma 4, Nemotron 3.5 Lightning, Mistral Medium 3.5 / Small 4, LongCat-2.0 and MiMo-V2.5/Pro.
• Adds vLLM memory-budget semantics, TP/DCP-aware placement and explicit engine-layout calibration.
• Adds a public Predicted vs Actual corpus from vLLM/SGLang deployment receipts.
• Adds vramscout --parse-log for normalizing vLLM/SGLang startup telemetry.
• Adds Chinese documentation in README_CN.md.

Validation snapshot

Across six new hard cache receipts added in v0.7:

• architecture-only MAPE: 2.850%
• audited public-profile fit after explicit calibration: 0.237%

These are cache/layout validation figures, not a guarantee that total process peak or max-context OOM boundaries have the same error.

See validation/PREDICTED_VS_ACTUAL.md for the full auditable table and sources.

Install from source

git clone https://github.com/S3vryn/vramscout.git
cd vramscout
pip install -e .

PyPI publishing is prepared separately; until the trusted publisher is authorized on PyPI, source installation remains the supported path.