Tools · Featured · 2 of 37

VRAM allocator

Weights = P × bytes. KV = 2·L·n_kv·d_h·S·B·bytes. GQA labeled.

Need

What this desk is for.

A vendor screenshot treated as a capacity plan.

How

What runs in this tab.

Weights = P × bytes. KV = 2·L·n_kv·d_h·S·B·bytes. GQA labeled.

Role analog: MLOps Engineer · 2026 US year-1 replacement $255,300 (market, not payroll). Kernel runs here. This tab stays local. Not zg, not SignalR, not Copilot, not tiktoken. Lite profile hides this canvas from default nav. Deep-link is calculator-only, not a 404. Device none.

Desks that sit this analog: MLOps Engineer. Unique pain, not a shared slogan.

Calculator-only. Device none. Lite profile hides this canvas from default nav. Deep-link is not a 404. This host does not start CUDA.

16.06 GBWeights / TP
1.07 GBKV cache
2.41 GBWorkspace est.
19.54 GBTotal

Fits the 24 GB class. GQA 32:8. KV is not sharded here. Implementation-dependent. Identity: Llama 3.1 8B fp16 KV/token = 2×32×8×128×2 = 131072 B.

Register a GQA card

Custom cards persist at engos.tools.vram.registry. Number() only. No eval.

Talk Route in the workspaceSave to a workspaceAll 37 tools

Featured cluster

Same cluster. Nothing deleted.

EngOS