dsh-plugin-vllm-ascend-profiler

Pinpointing whether slow vLLM inference on Ascend NPUs is caused by host scheduling, NPU compute, cross-card communication, or H2D/D2H copies is hard when the raw CANN and torch_npu profiling artifacts (trace_view.json, kernel_details.csv, op_statistic.csv, and more) carry bare JSON arrays, CSV headers that shift across CANN versions, and virtually no Prefill-or-Decode labels. This plugin absorbs all of that parsing complexity: drag in a zip, point it at a server directory, or upload individual files, and you get a full analysis without maintaining a hand-rolled pipeline.

The six-step workflow goes: import, overview and bottleneck identification, Host and Device operator swim-lane timeline, 100-percent category bar plus operator latency treemap, a five-node reasoning chain (bottleneck scoring, quantitative evidence vs. thresholds, root-cause inference, prioritized actions, expected gains), and a before-and-after comparison. All three visualization modules share one linked filter state, so clicking an operator in the swim lane highlights the same operator across the treemap and category bar in sync. Every conclusion traces back to a specific metric, threshold, and formula, and Prefill and Decode phases are reported separately. The output is a self-contained Markdown report or a print-to-PDF document with vector chart snapshots, ready for a design review or a ticket.

Built for ML engineers and DevOps working on Ascend NPU inference optimization who need to isolate the dominant bottleneck fast, verify that a configuration change actually helped, and produce a defensible report without maintaining a front-end visualization stack. The plugin has zero runtime dependencies, is a pure ESM package, and is ready to use the moment the profile restarts.