Spaces:

---

karlexmarin

/

taf-agent

like 0

Running

⚡ TL;DR — two things in v0.12, both born from what the community actually asked for:

💾 Fit Check — "will it fit on my GPU?"

The #1 recurring forum question, answered before you download anything — including the part every VRAM calculator forgets: the KV cache.

Llama-3.1-8B · fp16 · RTX 4090 (24 GB) · context 131072
  ⚖️  weights   15.0 GB
  🧊  KV cache  16.0 GB   ← 128 KB per token, 50% of the total!
  📦  total     31.9 GB   → 🚨 DOES NOT FIT — and it's the KV cache, not the model
  ✂️  max context that DOES fit: ~66,000 tokens
  💡  cheapest rescues: q8_0 KV cache → Q6_K weights → partial offload

Model id + precision (fp16/bf16/int8/nf4 or any GGUF quant) + GPU + target context → the full budget, which side is the problem (weights-bound vs KV-bound), the max context that fits, and the cheapest fix.

🎬 The demo runs itself: https://karlesmarin.github.io/tafagent/?demo=fitcheck

🌍 Your language, end to end

Until v0.11 the menus were translated but demos and recipe results came out in English. v0.12 fixes it at the root: guided demos follow your browser language automatically, and the recipe engine now emits message codes the UI localizes into EN / ES / FR / ZH (English always kept as fallback). Guarded by a real-browser regression test + a CI sweep of every mode in every language on each push.

⚖️ Honesty corner

• 📐 Everything from config.json is a prediction, not a measurement — the Diagnose CLI + Prediction-vs-Reality modes exist to check.
• 🚧 Verdict pill labels ("GO", "MEMORY-LIMITED"…) are still English-only; next batch.
• 🔒 Nothing you type leaves your browser. No server, no telemetry, no signup.

📦 Source: https://github.com/karlesmarin/tafagent · Feedback and refutations welcome — there's a registry with issue templates for both.