Pull requests / #22
#22 tools: add real-time web dashboard for hardware and engine monitoring
closed · @code-martin · 0 评论 · 在 GitHub 查看
BenchmarksServer & APIMulti-GPUAMD / HIPNVIDIA / CUDAModels & quantsWindowsLinux
描述
## What Adds a self-contained real-time monitoring web dashboard (`tools/dashboard_server.py`) for Strata: - Zero external dependencies: implemented purely with Python standard library (`http.server`, `urllib`, `json`, `subprocess`, `ctypes`). - Single-page responsive dark UI with live SVG telemetry gauges, mini trend charts, and real-time generation metrics. - Runs alongside the Strata engine server: `python tools/dashboard_server.py --port 8081 --server http://127.0.0.1:8080`. ## Preview  ## Features - **Strata Engine Tracking**: - Live state, prompt/generation progress bar, current token throughput (tok/s), context window utilization, and active KV cache mode. - Completed request history log with prefill/gen timing, token counts, and effective speeds. - **Hardware Telemetry (Windows & Linux)**: - **GPU**: Util %, VRAM usage/capacity, temperature, board power draw, active PCIe generation and lane width (e.g. `PCIe 4.0 x16` via NVML), and expert cache slots in VRAM. - **CPU & RAM**: Core/thread count, CPU utilization, system RAM used/total. - **NVMe Activity**: Per-disk read/write MB/s and load percentage (queries PDH on Windows, `/proc/diskstats` on Linux) to monitor DirectStorage/SSD streaming during MoE expert offloading. - **Resilient**: Gracefully handles engine reconnects, idle periods, and background streaming. ## Verification - Tested on Windows 11 with NVIDIA RTX 3080 + AMD Ryzen 9 5950X + Dual NVMe SSDs while running Strata with Qwen3.8-Flash-Next (Swift 1.5). - Verified live polling updates, responsive gauge animations, and request history tracking.
站内延伸阅读
链到安装、模型与版本说明,便于 SEO/GEO,非官方 issue 正文。