dsh-live-perf-gauges
An automotive-style live performance strip for DeepSeek Harness Web: decode throughput, TTFT, effective prompt rate, stream-event/MTP diagnostics, and output-token progress under the composer.
What it measures
- Decode tok/s refreshes every 250 ms during generation. Provider cumulative usage wins whenever an adapter exposes it. Otherwise the live estimate uses a character/token ratio calibrated from completed steps in the same session; a new session shows no invented tok/s until it has a calibration. Final provider usage is authoritative.
- TTFT is the elapsed time from
step/startto the first streamed text, reasoning, or tool-argument delta. - Prompt/TTFT is the logical prompt (
input + cache read + cache write) divided by end-to-end TTFT. It is deliberately not labelled as vLLM kernel prefill throughput because it includes DSH overhead, network/queue time, and cached tokens. - Stream reports logical delta events per second and generated tokens per event. The latter is exact after provider usage arrives and makes MTP/speculative-decoding acceptance visible.
- Output follows the same estimate-then-exact rule.
One SSE event is never assumed to equal one token. This matters for speculative decoding: an accepted MTP group can deliver several generated tokens in one streamed event.
The plugin never contacts a network service and never stores message text, prompts, tools, or credentials. It only projects timestamps, counts, and final token usage already present in the session event stream.
Install
dsh plugin --profile web add github:TnzGit/dsh-live-perf-gauges
Restart the running dsh web service, then hard-refresh the browser. The strip appears below the composer after the next model step begins.
Compatibility
Designed for the DSH 0.1.x Web profile and its standard conversation.composer.dock, sessionProjections, and contextPressure interfaces.
Development
npm test
This repository ships prebuilt dependency-free host and client entries, so users do not need an install-time build step.
License
MIT
还没有评论，来写第一条。