fix type hints for Python 3.8 and add verbose logging

whisper_trt/model.py

@@ -34,6 +34,8 @@
 
 import torch.nn as nn
 import torch2trt
+import tensorrt
+
 from dataclasses import asdict
 from .cache import get_cache_dir, make_cache_dir
 from .__version__ import __version__
@@ -132,7 +134,7 @@ def __init__(self,
             dims: ModelDimensions,
             encoder: AudioEncoderTRT,
             decoder: TextDecoderTRT,
-            tokenizer: Tokenizer | None = None
+            tokenizer: Tokenizer = None
         ):
         super().__init__()
         self.dims = dims
@@ -150,7 +152,7 @@ def forward(self, mel: Tensor, tokens: Tensor):
         return self.decoder(tokens, self.encoder(mel))
 
     @torch.no_grad()
-    def transcribe(self, audio: str | np.ndarray):
+    def transcribe(self, audio): #: str | np.ndarray):
 
         if isinstance(audio, str):
             audio = whisper.audio.load_audio(audio)
@@ -184,6 +186,7 @@ class WhisperTRTBuilder:
     model: str
     fp16_mode: bool = True
     max_workspace_size: int = 1 << 30
+    verbose: bool = False
 
     @classmethod
     @torch.no_grad()
@@ -222,7 +225,8 @@ def build_text_decoder_engine(cls) -> torch2trt.TRTModule:
             input_names=["x", "xa", "mask"],
             output_names=["output"],
             max_workspace_size=cls.max_workspace_size,
-            fp16_mode=cls.fp16_mode
+            fp16_mode=cls.fp16_mode,
+            log_level=tensorrt.Logger.VERBOSE if cls.verbose else tensorrt.Logger.ERROR,
         )
 
         return engine
@@ -266,7 +270,8 @@ def build_audio_encoder_engine(cls) -> torch2trt.TRTModule:
             input_names=["x", "positional_embedding"],
             output_names=["output"],
             max_workspace_size=cls.max_workspace_size,
-            fp16_mode=cls.fp16_mode
+            fp16_mode=cls.fp16_mode,
+            log_level=tensorrt.Logger.VERBOSE if cls.verbose else tensorrt.Logger.ERROR,
         )
 
         return engine
@@ -298,8 +303,9 @@ def get_audio_encoder_extra_state(cls):
 
     @classmethod
     @torch.no_grad()
-    def build(cls, output_path: str):
-
+    def build(cls, output_path: str, verbose: bool = False):
+        cls.verbose = verbose
+
         checkpoint = {
             "whisper_trt_version": __version__,
             "dims": asdict(load_model(cls.model).dims),
@@ -398,7 +404,7 @@ class SmallEnBuilder(EnBuilder):
     "small.en": SmallEnBuilder
 }
 
-def load_trt_model(name: str, path: str | None = None, build: bool = True):
+def load_trt_model(name: str, path: str = None, build: bool = True, verbose: bool = False):
 
     if name not in MODEL_BUILDERS:
         raise RuntimeError(f"Model '{name}' is not supported by WhisperTRT.")
@@ -413,6 +419,6 @@ def load_trt_model(name: str, path: str | None = None, build: bool = True):
         if not build:
             raise RuntimeError(f"No model found at {path}.  Please call load_trt_model with build=True.")
         else:
-            builder.build(path)
+            builder.build(path, verbose=verbose)
 
     return builder.load(path)