shengbinxu/dsh-llm-latency
A passive LLM latency dashboard for DeepSeek Harness: see how long models take to respond, compare vendors and inspect requests without extra model calls.
dsh-llm-latency is a passive latency dashboard for DeepSeek Harness: it shows how long models take to respond, compares vendors and lets you inspect individual requests — all without starting extra model calls. The v0.2.0 README ships a native dashboard (overview, time-window comparison, session comparison and a request log) plus a composer quick view for the last 15 minutes, hour or three hours, reporting latest response delay, median response delay, median first-text delay and successful/total samples. It deliberately separates total output speed from strict visible-answer rates and surfaces metric explanations on hover or keyboard focus.
Install
dsh plugin --profile web add 'github:shengbinxu/dsh-llm-latency#v0.2.0'Copied verbatim from the README's install section (the README also gives a local tarball form, dsh plugin --profile web add ./dsh-llm-latency-0.2.0.tgz). The README stresses it is passive: 'without starting extra model calls'.
Compatibility
The README says it reuses already-recorded data — the legacy /llm-latency/ JSON endpoints, the latency_report tool, recorded statistics and CSV export all remain available — and it never probes an unused model. Idle models honestly show 'No recent samples' rather than substituting a historical average for a current measurement.
Details
- Repo: shengbinxu/dsh-llm-latency
- Category: Models & Providers
- Stars: 0
- Version: GitHub repo shengbinxu/dsh-llm-latency; installed from the tagged github: spec rather than a published npm package. The README's 'What's in v0.2.0' section identifies this as the v0.2.0 release.
FAQ
- Does it cost extra model calls?
- No — the README says it is passive and inspects already-recorded data without starting extra model calls.
- What does an idle model show?
- The README says idle models show 'No recent samples'; the panel never substitutes a historical average for a current measurement or probes an unused model.
Alternatives
BruceLanLan/dsh-tier-router · NOirBRight/dsh-llm-ollama · wenzetan/dsh-llm-newapi