aadebuger-ComfyUI-HyMT2/cli

ComfyUI-HyMT2 — CLI tools

Two helpers for running Hy-MT2 through a remote ComfyUI server without opening the UI.

Script	Style	Purpose
hymt2_cli.py	v2.0 套路 A/B	One-shot translation.
hymt2_batch.py	v2.0 套路 C	JSON batch with resumable runs.

Both default to 127.0.0.1:8188; point them at your GPU server with --ip 10.8.0.12 (or --server 10.8.0.12:8188).

For best startup time we recommend calling the venv's python directly (per v2.0 §2.4):

PY=/mnt/modelall/ComfyUI/.venv/bin/python

1. hymt2_cli.py — single shot

Inline text

$PY hymt2_cli.py "今天天气真好,我们去散步吧。" \
    --target English --prompt-lang en --ip 10.8.0.12

Read from a file

$PY hymt2_cli.py /path/to/article.zh.md \
    --target English --prompt-lang en --ip 10.8.0.12 \
    --out /path/to/article.en.md

Read from stdin

cat article.zh.md | $PY hymt2_cli.py - \
    --target Japanese --prompt-lang en --ip 10.8.0.12

With a glossary

cat > /tmp/glossary.txt <<'EOF'
# tech glossary
endpoint => endpoint
webhook => webhook
token    => token
EOF

$PY hymt2_cli.py "调用 API 时,请在 endpoint 后附 token" \
    --target English --prompt-lang en --ip 10.8.0.12 \
    --terminology /tmp/glossary.txt

With background context

$PY hymt2_cli.py "本判决系终审判决。" \
    --target English --prompt-lang en --ip 10.8.0.12 \
    --background "Closing paragraph of a Chinese administrative-litigation judgment."

With a style descriptor

$PY hymt2_cli.py "Hi guys, this is a great product!" \
    --target Chinese --prompt-lang zh --ip 10.8.0.12 \
    --style "正式书面语"

Useful flags

Flag	Default	Notes
--target	English	Full language name. EN names with --prompt-lang en, ZH names with --prompt-lang zh.
--prompt-lang	en	Selects the EN or ZH prompt template.
--model-path	tencent/Hy-MT2-1.8B	Use tencent/Hy-MT2-7B for the 7B model.
--gpu	0	Maps to cuda:N.
--temperature / --top-p / --top-k / --repetition-penalty / --max-new-tokens	model-card defaults	Override the sampling params.
--seed	-1	Set a value for reproducible runs.
--dump-workflow PATH	—	Write the API workflow JSON for UI loading.
--ws-recv-timeout	600 s	Cold-load tolerance for the WebSocket idle timeout (v2.0 §10).
--force-reload	off	Bypass the loader cache.

2. hymt2_batch.py — batch translation

Input formats accepted (autodetected):

// (a) list of strings
[
  "今天天气真好",
  "明天去散步吧"
]

// (b) list of {id, text}
[
  { "id": "weather_1", "text": "今天天气真好" },
  { "id": "weather_2", "text": "明天去散步吧" }
]

// (c) dict of id → text
{
  "weather_1": "今天天气真好",
  "weather_2": "明天去散步吧"
}

Each item produces one output file <out-dir>/<sanitized-id>.json containing:

{
  "id": "weather_1",
  "source": "今天天气真好",
  "translation": "The weather is really nice today.",
  "target_lang": "English",
  "prompt_lang": "en",
  "model_path": "tencent/Hy-MT2-1.8B"
}

Resumable batch run

$PY hymt2_batch.py /mnt/modelall/in/ai_news_vocab.json \
    --out-dir /mnt/modelall/out/hymt2_ai_news \
    --target Japanese --prompt-lang en \
    --ip 10.8.0.12 \
    --limit 100 \
    --dump-workflow /mnt/modelall/tmp/hymt2_first.json

Re-run the same command — already-finished files are skipped. Pass --force to retranslate everything.

Apply a glossary to every item

$PY hymt2_batch.py words.json \
    --out-dir /mnt/modelall/out/hymt2_words \
    --target English --prompt-lang en --ip 10.8.0.12 \
    --terminology /mnt/modelall/in/glossary.txt

Debugging tips

1. --dump-workflow /tmp/wf.json and load that into the ComfyUI UI — easiest way to confirm the prompt was built the way you expected.
2. On the server: tail -f /mnt/modelall/tmp/comfy.log while the CLI runs to see [hymt2/translate], [hymt2/gen], and VRAM diagnostic lines.
3. If the WebSocket disconnects on a cold first-load, bump --ws-recv-timeout to 1200 or higher.