hwLedger

Appearance

GUI Journey: Planner Export to vLLM Flags ​

This page documents the export-gui-vllm journey, which exercises the Planner export pipeline — fixture loading, Export menu, vLLM flag modal, and clipboard copy.

Overview ​

Journey ID: export-gui-vllmStatus: Implemented (placeholder artefacts — real recording pending on user Mac) Last Updated: 2026-04-19

What you'll see ​

  • Planner opens at defaults (Llama-3.1-8B, 4096 tokens).
  • User clicks Load fixture...; dropdown lists 4 fixtures; the DeepSeek-V3 @ 32k / 8 users entry is selected.
  • Fixture loads: seq-len slider jumps to 32768, users slider to 8, stacked bar recomputes showing 71.4 GB VRAM total.
  • User clicks Export — menu slides down with vLLM flags, llama.cpp args, MLX JSON, TorchServe.
  • Selecting vLLM flags opens a modal sheet containing a monospaced flag string starting --model deepseek-v3 --max-model-len 32768 --max-num-seqs 8 --gpu-memory-utilization 0.92 --dtype bf16 ....
  • Copy pulses green-checked and a bottom-center toast reads Copied vLLM flags (148 chars).

Journey not yet recorded.

Run the journey recorder to capture interactions:

./apps/macos/HwLedgerUITests/scripts/run-journeys.sh

What to watch for ​

  • Flag-string parity — the modal string must match what hwledger plan --export vllm emits on the CLI for the same fixture. This is the visual half of FR-PLAN-007.
  • Recompute before export — the stacked bar recomputes after fixture load before the Export menu opens; exporting stale totals is a bug.
  • Toast timing — the toast fires only after the clipboard write succeeds; the modal stays open so the user can re-copy if they missed it.

Reproduce ​

bash
cd apps/macos/HwLedgerUITests
./scripts/bundle-app.sh --no-codesign debug

swift test --filter ExportVLLMJourneyTests/testExportGUIVLLM

cd ../../..
bash docs-site/scripts/sync-journey-artefacts.sh

Source ​

Released under the Apache 2.0 License.

Copyright © 2024-2026 hwLedger Contributors