ContextQMD
Back to Mnn

dumpapp Command Guide

apps/Android/MnnLlmChat/tools/dumpapp.md

3.5.08.9 KB
Original Source

dumpapp Command Guide

tools/dumpapp is a thin wrapper around Stetho dumpapp. It connects from your computer to the running Android app process and forwards the remaining arguments to the app-side dumper plugins.

This document covers the custom dumpapp commands registered by MnnLlmChat in debug builds. Stetho's built-in default commands are not listed here.

Basic Usage

bash
./tools/dumpapp <command> [args]

Examples:

bash
./tools/dumpapp models list
./tools/dumpapp llm status
./tools/dumpapp openai start --model ModelScope/MNN/Qwen3___5-0___8B-MNN

Connection Notes

  • -p or --process: override the target Android process name. Default is com.alibaba.mnnllm.android.
  • ANDROID_SERIAL: target a specific device when multiple devices are connected.
  • STETHO_PROCESS: set the default process name through environment variables.
  • ANDROID_ADB_SERVER_PORT or ADB_SERVER_SOCKET: override adb server connection settings if needed.

Available Commands

models

Inspect and maintain the app's model list and filesystem entries.

  • dumpapp models list [-v|--verbose]: list models as they appear to the UI.
  • dumpapp models dump: dump the raw current model-list state for debugging.
  • dumpapp models refresh: force a model-list refresh.
  • dumpapp models tags [modelId|--all]: print model tags and extra tags.
  • dumpapp models find <keyword>: search models by id or display name.
  • dumpapp models files: list .mnnmodels/ entries and symlink targets.
  • dumpapp models unlink <modelId|name>: remove only the outer symlink entry while preserving underlying data.

logs

Control runtime logging verbosity and file logging.

  • dumpapp logs get: show the current log level.
  • dumpapp logs set <LEVEL>: change log level. Supported values are VERBOSE, DEBUG, INFO, WARN, ERROR, ASSERT.
  • dumpapp logs file enable: enable file logging.
  • dumpapp logs file disable: disable file logging.
  • dumpapp logs file status: show whether file logging is enabled.

market

Inspect and control the model market network source used by the app.

  • dumpapp market status: show current market loading state.
  • dumpapp market allow on|off: allow or block network-based market fetches.
  • dumpapp market env dev|prod: switch the market endpoint environment.
  • dumpapp market refresh: force-refresh market data from the network.

download

Operate model downloads without touching the UI.

  • dumpapp download source: show the current preferred download source.
  • dumpapp download source set <SOURCE>: switch source. Aliases are hf, ms, ml.
  • dumpapp download test <modelId>: start a test download for the given model.
  • dumpapp download check-update <modelId>: check whether the remote model has updates.
  • dumpapp download status <modelId>: print download state for one model.
  • dumpapp download delete <modelId>: delete a downloaded model.
  • dumpapp download pause <modelId>: pause an active download.

sana

Run Sana image-generation models from the command line.

  • dumpapp sana list: list installed Sana models.
  • dumpapp sana run <model_path> <prompt> [options]: generate an image with a Sana model.

sana run options:

  • --output <path>: output image path. Default is /sdcard/sana_output.jpg.
  • --input <path>: input image path for img2img mode.
  • --steps <n>: inference step count.
  • --seed <n>: random seed.
  • --cfg <true|false>: enable or disable CFG mode.
  • --cfg-scale <f>: CFG scale.
  • --width <n>: output width.
  • --height <n>: output height.

diffusion

Run non-Sana diffusion image models from the command line.

  • dumpapp diffusion list: list available diffusion-capable models.
  • dumpapp diffusion run <modelId|modelPath> <prompt> [options]: generate an image with a diffusion model.

diffusion run options:

  • --output <path>: output image path. Default is /sdcard/diffusion_output.jpg.
  • --steps <n>: inference step count.
  • --seed <n>: random seed.

benchmark

Run on-device LLM benchmark cases from adb.

  • dumpapp benchmark list: list benchmarkable models.
  • dumpapp benchmark run <modelId> [options]: run a benchmark for one model.
  • dumpapp benchmark status: show the current benchmark state.
  • dumpapp benchmark stop: stop a running benchmark.

benchmark run options:

  • --backend <cpu|opencl>: select backend.
  • --prompt <n>: prompt token count.
  • --gen <n>: generated token count.
  • --repeat <n>: repeat count.
  • --threads <n>: thread count.

storage

Inspect and clean app internal storage, especially mmap cache leftovers.

  • dumpapp storage list: list internal storage directories.
  • dumpapp storage analysis: print a summarized storage analysis.
  • dumpapp storage analysis detail: print per-entry drill-down details.
  • dumpapp storage analysis <entryModelId>: inspect a single storage entry.
  • dumpapp storage mmap: list mmap cache directories.
  • dumpapp storage orphans: find orphaned mmap caches.
  • dumpapp storage clean: clean all orphan mmap caches.
  • dumpapp storage clean <path>: clean one relative path under filesDir.
  • dumpapp storage verify: output a machine-readable integrity check.

config

Validate model config loading (guards issue #4259).

  • dumpapp config validate <modelId>: verify config path is a file (not directory) and loadMergedConfig succeeds. Fails if settings would receive wrong path (e.g. model directory instead of config.json).
  • dumpapp config dump <modelId>: dump merged config (config_path, llm_model, llm_weight, system_prompt). RESULT=FAIL if llm_model/llm_weight empty (corrupted).

llm

Manage the OpenAI-service-side LLM runtime session.

  • dumpapp llm status: show whether a runtime session exists and which model is active.
  • dumpapp llm ensure <modelId> [--force-reload]: create or refresh the runtime session for a model.
  • dumpapp llm run <modelId> <prompt> [--force-reload] [--use-app-config]: ensure the runtime session for a model, then execute one prompt. By default uses LlmSession.submitFullHistory(...) with base config only (no custom_config merge). With --use-app-config: aligns with ChatActivity—uses ModelUtils.getConfigPathForModel, merges custom_config.json, useNewConfig=false, and calls LlmSession.generate(...) (submitNative); useful for reproducing issues that occur in the app chat but not in dumpapp.
  • dumpapp llm thinking get: read the current thinking-mode flag from the active runtime session.
  • dumpapp llm thinking set <on|off>: enable or disable runtime thinking mode.
  • dumpapp llm mock-latex <on|off> [path]: enable mocked streaming markdown output, optionally loading content from a file. The command name is historical; current smoke flows reuse it for LaTeX and markdown-table rendering.
  • dumpapp llm release: release the active runtime session.

Note: llm ensure remains the lifecycle primitive for loading or switching the active model session. llm run is the direct execution entrypoint layered on top of it.

openai

Control the app's OpenAI-compatible local API service.

  • dumpapp openai status: show service state and current config.
  • dumpapp openai diag: print service config plus runtime diagnostics.
  • dumpapp openai start [--model <modelId>]: start OpenAIService, optionally with a specific model preselected.
  • dumpapp openai stop: stop the service.
  • dumpapp openai reset-config: reset API settings and regenerate a default API key.
  • dumpapp openai test-chat [msg]: send a test message through the service to validate chat flow.

history

Inspect stored chat-session history.

  • dumpapp history list: list all sessions and chat-item counts.
  • dumpapp history show <sessionId>: print all saved chat data for one session.
  • dumpapp history check <sessionId>: run integrity checks for one session.
  • dumpapp history diag: run diagnostics across all history data.

voice

Test and validate Voice Chat capabilities (TTS/ASR).

  • dumpapp voice status: check voice models status (TTS and ASR paths and readiness).
  • dumpapp voice tts init: initialize the TTS service with the default model.
  • dumpapp voice tts test [text]: run TTS on the given text and report audio generation metrics.
  • dumpapp voice tts destroy: release the TTS service.
  • dumpapp voice tts status: show TTS service internal state.
  • dumpapp voice asr status: show ASR model path and configuration.

Note: Full ASR testing requires audio input (microphone). For end-to-end ASR validation, use the Voice Chat UI smoke test (16_regress_voice_ui.sh).

Examples:

bash
dumpapp voice status
dumpapp voice tts init
dumpapp voice tts test "Hello, this is a TTS smoke test."
dumpapp voice tts destroy

Source of Truth

The commands above are registered in app/src/debug/java/com/alibaba/mnnllm/android/StethoInitializer.kt, and their behavior is implemented in the dumper plugin classes under app/src/debug/java/com/alibaba/mnnllm/android/debug/ (including VoiceDumperPlugin.kt) and app/src/debug/java/com/alibaba/mnnllm/api/openai/debug/.