funasr_wss_server.py ignores --asr_model and --asr_model_revision: the offline model is always paraformer-zh v2.0.4 #3886

runtime/python/websocket/funasr_wss_server.py parses --asr_model and --asr_model_revision (funasr_wss_server.py:45-51), and the README lists --asr_model in the server's API reference (runtime/python/websocket/README.md:29). But the server builds its offline model from fixed values (funasr_wss_server.py:237-245):

model_asr = AutoModel(
    model="paraformer-zh",
    model_revision="v2.0.4",
    ngpu=args.ngpu,
    ...

So both options have no effect. A server started with --asr_model paraformer-en, or with the directory of a fine-tuned model, still runs paraformer-zh for its offline and 2pass-offline results. The default of --asr_model names yet another model, the Contextual Paraformer (iic/speech_paraformer-large-contextual_asr_nat-zh-cn-16k-common-vocab8404), which the server does not load either. The streaming, VAD and punctuation models are built from their options (--asr_model_online, --vad_model, --punc_model and their revisions).

The fixed values came with #2791. Before it, the server called AutoModel(model=args.asr_model, model_revision=args.asr_model_revision, ...).

Reproduction on main (6f52414), no model files. The server builds its models when it is imported, so this runs the real server module with AutoModel replaced by a recorder and prints the offline model for three command lines. It needs only numpy. Run it from the root of a FunASR checkout:

import contextlib
import importlib.util
import io
import sys
import types

# The server builds its models with AutoModel when it is imported. Replace AutoModel with a
# recorder (each call returns its keyword arguments), and the modules that the server only
# uses to serve requests (websockets, scipy, torch) with empty stand-ins.
funasr = types.ModuleType("funasr")
funasr.AutoModel = types.SimpleNamespace
sys.modules["funasr"] = funasr
for name in ("websockets", "torch", "scipy", "scipy.spatial", "scipy.spatial.distance"):
    sys.modules[name] = types.ModuleType(name)
sys.modules["scipy.spatial.distance"].cosine = None

SERVER = "runtime/python/websocket/funasr_wss_server.py"
for options in [
    [],
    ["--asr_model", "paraformer-en"],
    ["--asr_model", "/models/my-paraformer", "--asr_model_revision", "master"],
]:
    sys.argv = [SERVER, *options]
    spec = importlib.util.spec_from_file_location("funasr_wss_server", SERVER)
    server = importlib.util.module_from_spec(spec)
    with contextlib.redirect_stdout(io.StringIO()):  # "model loading" ...
        spec.loader.exec_module(server)
    server.EXECUTOR.shutdown()
    print(f"options {options}")
    print(f"  offline model: {server.model_asr.model} {server.model_asr.model_revision}")

Output:

options []
  offline model: paraformer-zh v2.0.4
options ['--asr_model', 'paraformer-en']
  offline model: paraformer-zh v2.0.4
options ['--asr_model', '/models/my-paraformer', '--asr_model_revision', 'master']
  offline model: paraformer-zh v2.0.4

Expected: the model that --asr_model and --asr_model_revision name (paraformer-en v2.0.4, /models/my-paraformer master), and paraformer-zh v2.0.4, as now, when they are not given.

Environment: Windows 11, Python 3.11.17, numpy 1.26.4, sources from main.