v2.0.0 - config-grounded memory ledger

v2.0.0 release notes draft

Config-grounded memory ledger for two supported dense model families: Qwen2.5-7B-Instruct and Qwen3-8B.

What changed

• Offline catalogue built from official Hugging Face configs and file-size metadata at pinned revisions. No model weights downloaded.
• KV bytes use layers, KV heads and head dimension from the actual config, not a guessed parameter-count curve.
• Distinguishes weight-file bytes, allocated KV cache, arithmetic subtotal and an optional explicit manual overhead bound. Unknown runtime overhead is not reported as zero. The total is not a measured GPU reading or a fit guarantee.
• CLI accepts a supported model id and exports JSON. Old parameter-only invocations have explicit migration/refusal; opt-in legacy estimates remain labelled heuristic.
• Browser supports validated URL state, copied reports and opt-in Telegram sharing. Channel subscription is never required. No tracking or identifier collection added.
• Unsupported MoE, MLA, hybrid/sliding attention, quantized KV and partial offload receive explicit limitations/refusals.
• The old streaming benchmark is disabled with benchmark_disabled and no network call. No throughput result is claimed.

Known limits

Only the two reviewed catalogue entries and fp16/bf16 KV are supported. File bytes are not identical to runtime-resident allocation; CUDA context, workspaces, alignment, allocator fragmentation, graph buffers and server effects must be measured separately. fp16 file-size equivalence is not a separate downloaded official fp16 artifact. No GPU/runtime measurements were performed.

Verification

155 application tests passed in isolated Docker, covering backend/CLI and Chromium plus Firefox UI. 20 Node state tests passed. A clean wheel built and installed; pinned catalogue data and the disabled-benchmark refusal were checked. Independent same-model source review completed with parent empirical adjudication of findings. GitHub CI on Python 3.9/3.11/3.13, package build and both browsers passed; main merge and Pages deployment succeeded. Parent read back the deployed HTML byte-for-byte and exercised its actual HTTPS UI in both browsers. No GPU measurement is claimed.

Try the owned project

https://dangerousanen.github.io/llm-vram-calculator/
https://github.com/DangerousANEN/llm-vram-calculator

AI assistance: implementation and routine review used deepseek-v4.1-flash-1017. A focused GPT advisor reviewed release security/measurement gates. Human account ownership does not imply a human authored every line; this is an AI-assisted project.

v1.1.0 — live calculator + pip-installable

• Live web calculator: https://ai.zxc1x1.ru/tools/vram
• Install: pip install https://github.com/DangerousANEN/llm-vram-calculator/releases/download/v1.1.0/llm_vram_calculator-1.0.0-py3-none-any.whl
• CLI: llm-vram-calc
• Zero dependencies, Python 3.8+
• Telegram: @llm_hubs

v1.0.0 - LLM VRAM & KV-Cache Calculator

First official release of the zero-dependency LLM VRAM Calculator, KV-Cache estimator, and streaming benchmark suite.

• Live WebApp: https://dangerousanen.github.io/llm-vram-calculator/
• Official Channel: https://t.me/llm_hubs
• Includes support for GQA, MHA, and DeepSeek MLA architectures.