File size: 10,423 Bytes
54fb37d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
import base64
import datetime
import io
import logging
import os
import numpy as np
import soundfile as sf
from faster_whisper import WhisperModel
from typing import Dict, Any, Union

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

class EndpointHandler:
    def __init__(self, path=""):
        """
        Initialize the endpoint handler.

        Args:
            path: Path to the model directory. In Hugging Face Inference Endpoints,
                 this will be the directory containing the model files.
        """

        logger.info("Initializing EndpointHandler")

        if os.environ.get("LOG_DIAGNOSTICS") == "true":
            self.__log_diagnostics__()

        cudaIsAvailable = os.environ.get("CUDA_VISIBLE_DEVICES")
        if cudaIsAvailable:
            logger.info("CUDA is available")
        else:
            logger.info("CUDA is not available, using CPU")

        # Load the model with fallback to CPU if CUDA fails
        device = "cuda" if cudaIsAvailable else "cpu"
        try:
            # First attempt with CUDA if available
            if device == "cuda":
                logger.info("Attempting to load model with CUDA support")
                self.model = WhisperModel(
                    path or ".",
                    compute_type="float16",
                    device="cuda",
                )
                logger.info("Model loaded successfully with CUDA")
            else:
                # CPU fallback
                raise ValueError("CUDA not available, using CPU")
        except Exception as e:
            # Log the error and fall back to CPU
            logger.warning(f"Error loading model with CUDA: {e}")
            logger.info("Falling back to CPU model")
            try:
                self.model = WhisperModel(
                    path or ".",
                    compute_type="int8",
                    device="cpu",
                )
                logger.info("Model loaded successfully with CPU")
            except Exception as cpu_err:
                logger.error(f"Error loading CPU model: {cpu_err}")
                raise

        # Set default parameters
        self.sampling_rate = 16000

        logger.info("EndpointHandler initialized")

    def __call__(self, data: Dict[str, Any]) -> Dict[str, Any]:
        """
        Process a request.

        Args:
            data: Request data containing audio input and optional parameters.
                 Expected format:
                 - For batch processing: {"inputs": audio_data, ...parameters}
                 - For streaming: {"inputs": audio_chunk, "stream": true, "session_id": "unique_id", ...parameters}

        Returns:
            Transcription result or error message.
        """
        logger.info("Processing request")

        try:
            # Extract inputs and parameters
            if "inputs" not in data:
                return {"error": "No inputs provided"}

            inputs = data.pop("inputs")
            parameters = data.pop("parameters", {})

            # Process audio input
            audio, sampling_rate = self._process_audio_input(inputs)

            # Get transcription parameters
            language = data.pop("language", "da")  # Default to Danish
            beam_size = data.pop("beam_size", 5)

            return self.transcribe(
                audio,
                sampling_rate,
                language,
                beam_size,
                **parameters,
            )

        except Exception as e:
            logger.error(f"Error processing request: {e}")
            return {"error": str(e)}

    def _process_audio_input(self, inputs: Union[str, Dict[str, Any]]) -> tuple:
        """
        Process audio input in various formats.

        Args:
            inputs: Audio input as base64 string, URL, or numpy array.

        Returns:
            Tuple of (audio_array, sampling_rate).
        """
        # Handle different input formats
        if isinstance(inputs, str):
            logger.info("Received audio input as base64 encoded string")
            # Base64 encoded audio
            if inputs.startswith(("data:", "data%3A")):

                # Handle data URI
                if "base64," in inputs:
                    audio_b64 = inputs.split("base64,")[1]
                else:
                    audio_b64 = inputs

                # Decode base64
                audio_bytes = base64.b64decode(audio_b64)

                # Read audio data
                with io.BytesIO(audio_bytes) as audio_io:
                    audio, sampling_rate = sf.read(audio_io)

                return audio, sampling_rate

            # URL or file path (not implemented for security reasons)
            else:
                raise ValueError("URL or file path inputs are not supported")

        # Dictionary with audio data
        elif isinstance(inputs, dict) and "audio" in inputs:
            logger.info("Received audio input as dictionary")
            if isinstance(inputs["audio"], list):
                # Convert list to numpy array
                audio = np.array(inputs["audio"], dtype=np.float32)
            else:
                # Assume it's already a numpy array
                audio = inputs["audio"]

            # Get sampling rate
            sampling_rate = inputs.get("sampling_rate", self.sampling_rate)

            return audio, sampling_rate

        elif isinstance(inputs, bytes):
            logger.info("Received raw bytes input")
            # Read audio data from bytes
            with io.BytesIO(inputs) as audio_io:
                audio, sampling_rate = sf.read(audio_io)
            return audio, sampling_rate

        # Unsupported input format
        else:
            raise ValueError(f"Unsupported input format: {type(inputs)}")

    def transcribe(
        self,
        audio: np.ndarray,
        sampling_rate: int,
        language: str,
        beam_size: int,
        **kwargs,
    ) -> Dict[str, Any]:
        """
        Perform the transcription

        Args:
            audio: Audio data as numpy array.
            sampling_rate: Sampling rate of the audio.
            language: Language code.
            beam_size: Beam size for transcription.
            **kwargs: Additional parameters.

        Returns:
            Transcription result.
        """
        logger.info(f"Batch transcription: {len(audio)} samples, {sampling_rate} Hz")

        # Resample if needed
        if sampling_rate != self.sampling_rate:
            logger.warning(
                f"Sampling rate mismatch: {sampling_rate} Hz vs {self.sampling_rate} Hz"
            )

        logger.info(f"Parameters: {kwargs}")

        # Transcribe
        # see parameters here: <https://github.com/SYSTRAN/faster-whisper/blob/master/faster_whisper/transcribe.py>
        now = datetime.datetime.now()
        segments, info = self.model.transcribe(
            audio,
            language=language,
            beam_size=beam_size,
            **kwargs,
        )

        logger.info(f"Transcription info: {info}")

        # Format results
        result = {
            "text": "",  # Full text will be populated below
            "segments": [],
            "language": info.language,
            "language_probability": info.language_probability,
        }

        # Process segments
        # Segments is a generator and the actual transcription is done in the loop below,
        all_text = []
        for segment in segments:
            segment_info = {
                "id": segment.id,
                "text": segment.text,
                "start": segment.start,
                "end": segment.end,
                "temperature": segment.temperature,
                "avg_logprob": segment.avg_logprob,
                "compression_ratio": segment.compression_ratio,
                "no_speech_prob": segment.no_speech_prob,
            }
            if kwargs.get("word_timestamps", False):
                # Add word timestamps if requested
                segment_info["words"] = [
                    {
                        "word": word.word,
                        "start": word.start,
                        "end": word.end,
                    }
                    for word in segment.words
                ]

            all_text.append(segment.text)
            result["segments"].append(segment_info)

        elapsed_time = datetime.datetime.now() - now
        logger.info(f"Transcription time: {elapsed_time}")
        logger.info(f"Segments: {len(result['segments'])}")
        # Combine text from all segments
        result["text"] = " ".join(all_text)

        return result

    def __log_diagnostics__(self):
        """
        Log diagnostics information for debugging.
        This includes CUDA availability, library paths, installed packages,
        and environment variables.

        Very useful as the HF endpoint runtime is rather secretive about its environment.
        """
        logger.info("Logging environment diagnostics")

        logger.info("LD_LIBRARY_PATH:")
        if "LD_LIBRARY_PATH" in os.environ:
            logger.info(os.environ["LD_LIBRARY_PATH"])
        else:
            logger.info("LD_LIBRARY_PATH not set")

        # and the files
        logger.info("LD_LIBRARY_PATH files:")
        if "LD_LIBRARY_PATH" in os.environ:
            for ld_path in os.environ["LD_LIBRARY_PATH"].split(":"):
                if os.path.exists(ld_path):
                    logger.info(f"  {ld_path}:")
                    for file in os.listdir(ld_path):
                        logger.info(f"    {file}")
                else:
                    logger.info(f"  {ld_path} does not exist")

        logger.info(f"Installed Python packages:")
        import pkg_resources

        for package in pkg_resources.working_set:
            logger.info(f"  {package.key}=={package.version}, {package.location}")

        # dump environment variables
        logger.info("Environment variables:")
        for key, value in os.environ.items():
            logger.info(f"  {key}: {value}")

        logger.info("NVIDIA environment:")
        import subprocess

        try:
            result = subprocess.run(["nvidia-smi"], capture_output=True, text=True)
            logger.info(result.stdout)
        except Exception as e:
            logger.warning(f"Could not run nvidia-smi: {e}")