handler.py

from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch
import librosa
import io
import base64
from typing import Dict, Any

class EndpointHandler:
    def __init__(self, path=""):
        print("Loading Whisper model...")
        try:
            try:
                self.model = WhisperForConditionalGeneration.from_pretrained(
                    path,
                    torch_dtype=torch.bfloat16,
                    device_map={"": 0},
                    attn_implementation="flash_attention_2"
                )
                print("✅ Flash Attention 2 activated!")
            except ImportError:
                print("⚠️ Flash Attention not available, fallback to eager")
                self.model = WhisperForConditionalGeneration.from_pretrained(
                    path,
                    torch_dtype=torch.float16,
                    device_map="auto"
                )

            self.processor = WhisperProcessor.from_pretrained(path)
            self.model.eval()

            if hasattr(torch, 'compile'):
                try:
                    self.model = torch.compile(self.model, mode="max-autotune")
                    print("Model compiled with max-autotune!")
                except Exception as e:
                    print(f"Max-autotune compilation failed: {e}")
                    try:
                        self.model = torch.compile(self.model, mode="reduce-overhead")
                        print("Model compiled with reduce-overhead!")
                    except Exception as e2:
                        print(f"Compilation failed: {e2}")

            # forced_decoder_ids pour français (comme fastapi)
            self.french_decoder_ids = self.processor.get_decoder_prompt_ids(
                language="french", task="transcribe"
            )

            print("Model loaded and optimized successfully!")
        except Exception as e:
            print(f"Error loading model: {e}")
            raise e

    def __call__(self, data: Dict[str, Any]) -> Dict[str, str]:
        try:
            inputs = data.get("inputs", "")
            parameters = data.get("parameters", {})

            # decode audio (base64 string or bytes)
            if isinstance(inputs, str):
                try:
                    audio_bytes = base64.b64decode(inputs)
                except Exception:
                    return {"error": "Invalid base64 encoded audio"}
            elif isinstance(inputs, bytes):
                audio_bytes = inputs
            else:
                return {"error": "Invalid input format. Expected base64 string or bytes"}

            # check size
            if len(audio_bytes) > 25 * 1024 * 1024:
                return {"error": "File too large (max 25MB)"}

            # load audio
            audio_array, _ = librosa.load(
                io.BytesIO(audio_bytes),
                sr=16000,
                mono=True,
                duration=30
            )

            if len(audio_array) == 0:
                return {"error": "Invalid or empty audio file"}

            model_inputs = self.processor(
                audio_array,
                sampling_rate=16000,
                return_tensors="pt"
            )
            model_inputs = {
                k: v.to(self.model.device).half() if v.dtype == torch.float32 else v.to(self.model.device)
                for k, v in model_inputs.items()
            }

            # params
            max_length = parameters.get("max_length", 256)
            num_beams = parameters.get("num_beams", 6)
            temperature = parameters.get("temperature", 0.0)

            # generate
            with torch.no_grad(), torch.inference_mode(), torch.autocast(device_type="cuda", dtype=torch.float16):
                predicted_ids = self.model.generate(
                    **model_inputs,
                    max_length=max_length,
                    num_beams=num_beams,
                    temperature=temperature,
                    do_sample=False,
                    early_stopping=True,
                    no_repeat_ngram_size=3,
                    repetition_penalty=1.1,
                    length_penalty=1.0,
                    use_cache=True,
                    pad_token_id=self.processor.tokenizer.eos_token_id,
                    forced_decoder_ids=self.french_decoder_ids,  # ✅ identique à fastapi
                    suppress_tokens=[],
                    begin_suppress_tokens=[]
                )

            transcription = self.processor.batch_decode(predicted_ids, skip_special_tokens=True)
            return {"transcription": transcription[0]}
        except Exception as e:
            return {"error": f"Transcription error: {str(e)}"}