🐛[bugfix] Model : Added the ability to automatically select calculation types from GPU devices

src-python/models/transcription/transcription_whisper.py

@@ -4,6 +4,7 @@ from typing import Callable
 import huggingface_hub
 from faster_whisper import WhisperModel
 import logging
+from utils import getBestComputeType
 
 logger = logging.getLogger('faster_whisper')
 logger.setLevel(logging.CRITICAL)
@@ -73,7 +74,7 @@ def downloadWhisperWeight(root, weight_type, callback=None, end_callback=None):
 
 def getWhisperModel(root, weight_type, device="cpu", device_index=0):
     path = os_path.join(root, "weights", "whisper", weight_type)
-    compute_type = "int8" if device == "cpu" else "float16"
+    compute_type = getBestComputeType(device, device_index)
     return WhisperModel(
         path,
         device=device,

src-python/models/translation/translation_translator.py

@@ -6,7 +6,7 @@ from .translation_utils import ctranslate2_weights
 
 import ctranslate2
 import transformers
-from utils import errorLogging
+from utils import errorLogging, getBestComputeType
 
 import warnings
 warnings.filterwarnings("ignore")
@@ -37,7 +37,7 @@ class Translator():
         weight_path = os_path.join(path, "weights", "ctranslate2", directory_name)
         tokenizer_path = os_path.join(path, "weights", "ctranslate2", directory_name, "tokenizer")
 
-        compute_type = "int8" if device == "cpu" else "float16"
+        compute_type = getBestComputeType(device, device_index)
         self.ctranslate2_translator = ctranslate2.Translator(
             weight_path,
             device=device,

src-python/utils.py

@@ -1,68 +1,19 @@
 import base64
 from typing import Any
 import json
-import random
-from typing import Union
-from os import path as os_path, rename as os_rename
 import traceback
 import logging
-from PIL.Image import open as Image_open
 
-def getImageFile(file_name):
+from ctranslate2 import get_supported_compute_types
-    img = Image_open(os_path.join(os_path.dirname(__file__), "img", file_name))
-    return img
 
-def callFunctionIfCallable(function, *args):
+def getBestComputeType(device, device_index) -> str:
-    if callable(function) is True:
+    compute_types = get_supported_compute_types(device, device_index)
-        function(*args)
+    compute_types = set(compute_types)
+    preferred_types = ["int8_bfloat16", "int8_float16", "int8", "bfloat16", "float16", "int8_float32", "float32"]
 
-def isEven(number):
+    for preferred_type in preferred_types:
-    return number % 2 == 0
+        if preferred_type in compute_types:
-
+            return preferred_type
-def makeEven(number, minus:bool=False):
-    if minus is True:
-        return number if isEven(number) else number - 1
-    return number if isEven(number) else number + 1
-
-def intToPctStr(value:int):
-    return f"{value}%"
-
-def floatToPctStr(value:float):
-    return f"{int(value*100)}%"
-
-def strPctToInt(value:str):
-    return int(value.replace("%", ""))
-
-def isUniqueStrings(unique_strings:Union[str, list], input_string:str, require=False):
-    import re
-    if isinstance(unique_strings, str):
-        unique_strings = [unique_strings]
-    patterns = [re.escape(s) for s in unique_strings]
-
-    counts = [len(re.findall(pattern, input_string)) for pattern in patterns]
-
-    if require is True:
-        # If require is True, unique_strings must appear once
-        return all(count == 1 for count in counts) and counts.count(1) == 2
-    else:
-        # If require is False, check if unique strings are used exactly once
-        return all(count == 1 for count in counts)
-
-# path先のweightフォルダがある場合にはそのフォルダ名をweightsに変更する
-def renameWeightFolder(path):
-    weight_path = os_path.join(path, "weight")
-    if os_path.exists(weight_path):
-        os_rename(weight_path, os_path.join(path, "weights"))
-
-def splitList(lst:list, split_count:int, to_shuffle:bool=False):
-    if to_shuffle is True:
-        random.shuffle(lst)
-
-    split_lists = []
-    for i in range(0, len(lst), split_count):
-        sub_list = lst[i:i+split_count]
-        split_lists.append(sub_list)
-    return split_lists
 
 def encodeBase64(data:str) -> dict:
     return json.loads(base64.b64decode(data).decode('utf-8'))