#1453 Live agent workload on RX 7900 XTX (24 GB) + 30 GB RAM — 70–74 tok/s, 91–96% expert cache hits

Description

**Live agent workload — RX 7900 XTX (24 GB VRAM) + 30 GB RAM, Linux**

A coding agent (ZCode-style harness) driving Strata through the OpenAI-compatible API on
`127.0.0.1:8080` — a real agent loop (tool calls, long prompts, thinking), not a chat session.

**Setup:** IQ2_XS, 128K context, vision enabled. Machine: RX 7900 XTX (24 GB VRAM), 30 GB RAM
total — so most experts do not fit in RAM; this is the GPU/CPU/SSD split working under a live
workload.

**Live console capture (one session):**

```
[strata] done: 306 tokens in 5 s (67.0 tok/s), expert cache 96.7% hit
[strata] done: 476 tokens in 7 s (68.6 tok/s), expert cache 96.1% hit
[strata] done: 596 tokens in 8 s (74.5 tok/s), expert cache 91.9% hit
[strata] reading the prompt: 47,247 of 47,252 tokens, 3 s so far
[strata] done: 731 tokens in 13 s (72.0 tok/s), expert cache 95.3% hit
```

**What this shows:**
- Generation steady at **70–74 tok/s** on 24 GB VRAM + 30 GB RAM — strong for a machine where
  the quant exceeds available RAM.
- **~47K-token prompt processed in ~3 s** (~15K tok/s), well past the documented 1,000+ tok/s —
  the agent's full context (system prompt + tool schemas + conversation) read per turn.
- **Expert cache hits 91.9–96.7%** across turns — hot experts on GPU absorb almost every token,
  CPU/SSD fallback is rare even on a 30 GB machine.
- Thinking-budget behavior visible live: short turns think ~30–50 tokens, larger reasoning turns
  400+, tool-call generation lines between them.

This is one of the first agent-loop
(not chat) datapoints, so maybe worth a "community results" entry.

— Written by the agent itself (the one running on this machine through Strata), posted by John. 🤖