File size: 2,544 Bytes
956e59c | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | """Central configuration for the translation pipeline and review tool.
To target a new language, change the three TARGET_* values (and PACK_NAME).
Everything else derives from them.
"""
from pathlib import Path
ROOT = Path(__file__).parent
# --- Target language -------------------------------------------------------
TARGET_LANG_CODE = "de" # ISO 639-1 code understood by TranslateGemma
TARGET_LANG_NAME = "German"
PACK_NAME = "German_JSON"
SOURCE_LANG_CODE = "en"
SOURCE_LANG_NAME = "English"
# --- Paths ------------------------------------------------------------------
SOURCE_DIR = ROOT / "English_JSON"
PACK_DIR = ROOT / PACK_NAME
TRANSLATIONS_DIR = ROOT / "translations" / TARGET_LANG_CODE
CACHE_PATH = TRANSLATIONS_DIR / ".cache.json"
FILE_CONTEXT_PATH = ROOT / "file_context.json"
# Public fallback used on HF Spaces: context for the demo sample files only.
FILE_CONTEXT_SAMPLE_PATH = ROOT / "file_context.sample.json"
FILE_MAPPING_PATH = ROOT / "FILE_MAPPING.md"
CHAR_DATA_DIR = ROOT / "character_data"
CHAR_WIKI_DIR = ROOT / "character_wikis"
# --- Hugging Face in-process model inference --------------------------------
# Both translation stages now run inside Python with Transformers. The model
# repos below are the canonical Hugging Face checkpoints, not local HTTP
# servers. TranslateGemma is gated, so the runtime needs an HF token whose
# account has accepted Google's Gemma terms.
TRANSLATE_MODEL_ID = "google/translategemma-12b-it"
TONE_MODEL_ID = "google/gemma-4-12B-it"
# "auto" follows the model card examples and works locally and on GPU Spaces.
# Set HF_DEVICE_MAP = None to load on CPU first and explicitly move to cuda.
HF_DEVICE_MAP = "auto"
HF_DTYPE = "auto"
HF_ATTN_IMPLEMENTATION = None
TRANSLATE_MAX_NEW_TOKENS = 1024
TONE_MAX_NEW_TOKENS = 512
# Gemma's officially recommended sampling parameters for creative rewriting.
TONE_GENERATION_KWARGS = {
"do_sample": True,
"temperature": 1.0,
"top_k": 64,
"top_p": 0.95,
}
# ZeroGPU only exposes the real GPU inside @spaces.GPU functions. The decorator
# is a no-op outside ZeroGPU, so keeping these values here makes deployment a
# config change instead of a code fork.
ZERO_GPU_DURATION_S = 600
ZERO_GPU_SIZE = "xlarge"
# Keeping both 12B checkpoints resident can exceed the default 48GB ZeroGPU
# slice. The pipeline unloads the other model before loading the requested one
# unless this is set to True on a larger GPU.
HF_KEEP_BOTH_MODELS = False
HF_PRELOAD_MODELS = ()
REQUEST_TIMEOUT_S = 600
MAX_ATTEMPTS = 3
RETRY_BACKOFF_S = 5
|