translator.py

import os
import pathlib
import torch
from fairseq2.assets import InProcAssetMetadataProvider, asset_store
from huggingface_hub import snapshot_download
from seamless_communication.inference import Translator

CHECKPOINTS_PATH = pathlib.Path(os.getenv("CHECKPOINTS_PATH", "/home/user/app/models"))
if not CHECKPOINTS_PATH.exists():
    snapshot_download(repo_id="facebook/seamless-m4t-v2-large", repo_type="model", local_dir=CHECKPOINTS_PATH)
asset_store.env_resolvers.clear()
asset_store.env_resolvers.append(lambda: "demo")
demo_metadata = [
    {
        "name": "seamlessM4T_v2_large@demo",
        "checkpoint": f"file://{CHECKPOINTS_PATH}/seamlessM4T_v2_large.pt",
        "char_tokenizer": f"file://{CHECKPOINTS_PATH}/spm_char_lang38_tc.model",
    },
    {
        "name": "vocoder_v2@demo",
        "checkpoint": f"file://{CHECKPOINTS_PATH}/vocoder_v2.pt",
    },
]
asset_store.metadata_providers.append(InProcAssetMetadataProvider(demo_metadata))

if torch.cuda.is_available():
    device = torch.device("cuda:0")
    dtype = torch.float16
else:
    device = torch.device("cpu")
    dtype = torch.float32

translator = Translator(
    model_name_or_card="seamlessM4T_v2_large",
    vocoder_name_or_card="vocoder_v2",
    device=device,
    dtype=dtype,
    apply_mintox=True,
)