Continous increase in Memory usage

I have the same issue with Wav2vec2 model which is deployed on my local GPU system the continuous growth of RAM causes the system to crash, and model is loaded using huggingface pipeline.below is the code that I used

@router.post("/transcribe")

async def quran(audio_file: UploadFile = File(...)):

    process = psutil.Process()

    start_ram = process.memory_info().rss / (1024**2)

# Track paths for strict cleanup

    temp_wav_path = None




try:

        audio_bytes = await audio_file.read()

# 2. Load with Torchaudio

# waveform shape: [channels, time]

        waveform, sample_rate = torchaudio.load(io.BytesIO(audio_bytes))




# 3. Pre-processing: Convert to Mono if Stereo

if waveform.shape[0] > 1:

            waveform = torch.mean(waveform, dim=0, keepdim=True)




# 4. Resample to 16kHz (Standard for most AI speech models)

        target_sample_rate = 16000

if sample_rate != target_sample_rate:

            resampler = torchaudio.transforms.Resample(orig_freq=sample_rate, new_freq=target_sample_rate)

            waveform = resampler(waveform)




# 5. Squeeze to 1D if your model expects [samples] instead of [1, samples]

# Most transformers models prefer a flat 1D array/tensor

        input_tensor = waveform.squeeze()




# 6. Inference

with torch.no_grad(): # Reduce memory usage during inference

            transcript_raw = ai_models["wav2vec"](input_tensor)

            transcript = transcript_raw.get("text", "")




return JSONResponse(content={"transcript": transcript}, status_code=200)




except Exception as e:

        logger.error(f"Transcription Error: {e}")

raise HTTPException(status_code=500, detail="Internal processing error")




finally:

# --- AGGRESSIVE CLEANUP ---

# 1. Delete file immediately

if temp_wav_path and os.path.exists(temp_wav_path):

try:

                os.remove(temp_wav_path)

except: pass




# 2. Clear local tensors

if 'waveform' in locals():

del waveform

del audio_bytes

# 3. GPU and Python GC

if torch.cuda.is_available():

            torch.cuda.empty_cache()

        gc.collect()




# 4. FORCE OS RELEASE (The "Malloc Trim")

# This tells Linux to actually take the memory back

if libc:

            libc.malloc_trim(0)




        final_ram = process.memory_info().rss / (1024**2)

        logger.info(f"RAM Status: {start_ram:.1f}MB -> {final_ram:.1f}MB")

Anyone can help.