ESPnet3 Publication Stages

Masao SomekiLess than 1 minute

ESPnet3 Publication Stages

Current publication flow is centered on:

pack_model
upload_model

Implementation:

espnet3.utils.publish_utils.pack_model
espnet3.utils.publish_utils.upload_model

Quick usage

python run.py \
  --stages pack_model upload_model \
  --training_config conf/training.yaml \
  --publication_config conf/publication.yaml

`pack_model`

pack_model creates a publishable bundle directory, usually:

<exp_dir>/model_pack

Strategy

pack_model.strategy supports:

auto
espnet2
espnet3

Current behavior:

if training_config.task is set, packing defaults to the espnet2 path
otherwise it uses the espnet3 path

What gets bundled

Current ESPnet3 packing copies recipe assets such as:

conf/
src/
run.py
pixi.toml
pixi.lock
.python-version

and usually includes the experiment directory and, when enabled, the recipe data_dir.

In practice the packed tree often looks like:

model_pack/
  conf/
  src/
  exp/
  data/
  run.py
  meta.yaml
  README.md
  scores.json

data/ is included when include_data_dir: true and the configured data_dir exists.

Additional file controls

publication.yaml can also define:

include
extra
exclude
files
yaml_files

These allow adding or explicitly registering artifacts in meta.yaml.

`meta.yaml`

pack_model writes meta.yaml into the bundle root. That metadata is later used by tools such as InferenceSession.

`InferenceSession`

Direct packaged-model inference is provided by:

espnet3.publication.InferenceSession

Typical use:

from espnet3.publication import InferenceSession

session = InferenceSession.from_pretrained(
    "espnet/your-model-tag",
    trust_user_code=True,
)
result = session(audio_array)

Important behavior:

it can load conf/inference.yaml from the packed bundle
it can use bundle metadata from meta.yaml
it can enable bundled recipe code such as src/ when trust_user_code=True

This is the main reason current espnet3 packing includes recipe configs and recipe-local user code.

`upload_model`

upload_model uploads the packed directory to Hugging Face.

Required field:

publication_config.upload_model.hf_repo

The packed directory must already exist before upload runs.