selfhostyourtech/reflector
2
0
Fork 0
mirror of https://github.com/Monadical-SAS/reflector.git synced 2025-12-21 04:39:06 +00:00
Code Issues Packages Projects Releases Wiki Activity

gpu self hosted setup guide (no-mistakes)

Browse Source
This commit is contained in:
Igor Loskutov
2025-12-09 11:25:09 -05:00
parent 5779478d3c
commit 2b3f28993f
14 changed files with 799 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
gpu/modal_deployments/reflector_diarizer.py
View File

@@ -24,6 +24,12 @@ app = modal.App(name="reflector-diarizer")
upload_volume = modal.Volume.from_name("diarizer-uploads", create_if_missing=True)
# IMPORTANT: This function is duplicated in multiple files for deployment isolation.
# If you modify the audio format detection logic, you MUST update all copies:
# - gpu/self_hosted/app/utils.py
# - gpu/modal_deployments/reflector_transcriber.py (2 copies)
# - gpu/modal_deployments/reflector_transcriber_parakeet.py
# - gpu/modal_deployments/reflector_diarizer.py (this file)
def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
parsed_url = urlparse(url)
url_path = parsed_url.path
@@ -39,6 +45,8 @@ def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtens
return AudioFileExtension("wav")
if "audio/mp4" in content_type:
return AudioFileExtension("mp4")
if "audio/webm" in content_type or "video/webm" in content_type:
return AudioFileExtension("webm")
raise ValueError(
f"Unsupported audio format for URL: {url}. "

25
gpu/modal_deployments/reflector_transcriber.py
View File

@@ -99,6 +99,12 @@ image = (
)
# IMPORTANT: This function is duplicated in multiple files for deployment isolation.
# If you modify the audio format detection logic, you MUST update all copies:
# - gpu/self_hosted/app/utils.py
# - gpu/modal_deployments/reflector_transcriber.py (this file - 2 copies!)
# - gpu/modal_deployments/reflector_transcriber_parakeet.py
# - gpu/modal_deployments/reflector_diarizer.py
def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
parsed_url = urlparse(url)
url_path = parsed_url.path
@@ -114,6 +120,8 @@ def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtens
return AudioFileExtension("wav")
if "audio/mp4" in content_type:
return AudioFileExtension("mp4")
if "audio/webm" in content_type or "video/webm" in content_type:
return AudioFileExtension("webm")
raise ValueError(
f"Unsupported audio format for URL: {url}. "
@@ -316,6 +324,11 @@ class TranscriberWhisperFile:
import numpy as np
from silero_vad import VADIterator
# IMPORTANT: This VAD segment logic is duplicated in multiple files for deployment isolation.
# If you modify this function, you MUST update all copies:
# - gpu/modal_deployments/reflector_transcriber.py (this file)
# - gpu/modal_deployments/reflector_transcriber_parakeet.py
# - gpu/self_hosted/app/services/transcriber.py
def vad_segments(
audio_array,
sample_rate: int = SAMPLERATE,
@@ -323,6 +336,7 @@ class TranscriberWhisperFile:
) -> Generator[TimeSegment, None, None]:
"""Generate speech segments as TimeSegment using Silero VAD."""
iterator = VADIterator(self.vad_model, sampling_rate=sample_rate)
audio_duration = len(audio_array) / float(SAMPLERATE)
start = None
for i in range(0, len(audio_array), window_size):
chunk = audio_array[i : i + window_size]
@@ -342,6 +356,9 @@ class TranscriberWhisperFile:
start / float(SAMPLERATE), end / float(SAMPLERATE)
)
start = None
# Handle case where audio ends while speech is still active
if start is not None:
yield TimeSegment(start / float(SAMPLERATE), audio_duration)
iterator.reset_states()
upload_volume.reload()
@@ -407,6 +424,12 @@ class TranscriberWhisperFile:
return {"text": " ".join(all_text), "words": all_words}
# IMPORTANT: This function is duplicated in multiple files for deployment isolation.
# If you modify the audio format detection logic, you MUST update all copies:
# - gpu/self_hosted/app/utils.py
# - gpu/modal_deployments/reflector_transcriber.py (this file - 2 copies!)
# - gpu/modal_deployments/reflector_transcriber_parakeet.py
# - gpu/modal_deployments/reflector_diarizer.py
def detect_audio_format(url: str, headers: dict) -> str:
from urllib.parse import urlparse
@@ -424,6 +447,8 @@ def detect_audio_format(url: str, headers: dict) -> str:
return "wav"
if "audio/mp4" in content_type:
return "mp4"
if "audio/webm" in content_type or "video/webm" in content_type:
return "webm"
raise HTTPException(
status_code=400,

13
gpu/modal_deployments/reflector_transcriber_parakeet.py
View File

@@ -90,6 +90,12 @@ image = (
)
# IMPORTANT: This function is duplicated in multiple files for deployment isolation.
# If you modify the audio format detection logic, you MUST update all copies:
# - gpu/self_hosted/app/utils.py
# - gpu/modal_deployments/reflector_transcriber.py (2 copies)
# - gpu/modal_deployments/reflector_transcriber_parakeet.py (this file)
# - gpu/modal_deployments/reflector_diarizer.py
def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtension:
parsed_url = urlparse(url)
url_path = parsed_url.path
@@ -105,6 +111,8 @@ def detect_audio_format(url: str, headers: Mapping[str, str]) -> AudioFileExtens
return AudioFileExtension("wav")
if "audio/mp4" in content_type:
return AudioFileExtension("mp4")
if "audio/webm" in content_type or "video/webm" in content_type:
return AudioFileExtension("webm")
raise ValueError(
f"Unsupported audio format for URL: {url}. "
@@ -301,6 +309,11 @@ class TranscriberParakeetFile:
audio_array, sample_rate = librosa.load(file_path, sr=SAMPLERATE, mono=True)
return audio_array
# IMPORTANT: This VAD segment logic is duplicated in multiple files for deployment isolation.
# If you modify this function, you MUST update all copies:
# - gpu/modal_deployments/reflector_transcriber.py
# - gpu/modal_deployments/reflector_transcriber_parakeet.py (this file)
# - gpu/self_hosted/app/services/transcriber.py
def vad_segment_generator(
audio_array,
) -> Generator[TimeSegment, None, None]:

137
gpu/self_hosted/DEV_SETUP.md Normal file
View File

@@ -0,0 +1,137 @@
# Local Development GPU Setup
Run transcription and diarization locally for development/testing.
> **For production deployment**, see the [Self-Hosted GPU Setup Guide](../../docs/docs/installation/self-hosted-gpu-setup.md).
## Prerequisites
1. **Python 3.12+** and **uv** package manager
2. **FFmpeg** installed and on PATH
3. **HuggingFace account** with access to pyannote models
### Accept Pyannote Licenses (Required)
Before first run, accept licenses for these gated models (logged into HuggingFace):
- https://hf.co/pyannote/speaker-diarization-3.1
- https://hf.co/pyannote/segmentation-3.0
## Quick Start
### 1. Install dependencies
```bash
cd gpu/self_hosted
uv sync
```
### 2. Start the GPU service
```bash
cd gpu/self_hosted
HF_TOKEN=<your-huggingface-token> \
REFLECTOR_GPU_APIKEY=dev-key-12345 \
.venv/bin/uvicorn main:app --host 0.0.0.0 --port 8000
```
Note: The `.env` file is NOT auto-loaded. Pass env vars explicitly or use:
```bash
export HF_TOKEN=<your-token>
export REFLECTOR_GPU_APIKEY=dev-key-12345
.venv/bin/uvicorn main:app --host 0.0.0.0 --port 8000
```
### 3. Configure Reflector to use local GPU
Edit `server/.env`:
```bash
# Transcription - local GPU service
TRANSCRIPT_BACKEND=modal
TRANSCRIPT_URL=http://host.docker.internal:8000
TRANSCRIPT_MODAL_API_KEY=dev-key-12345
# Diarization - local GPU service
DIARIZATION_BACKEND=modal
DIARIZATION_URL=http://host.docker.internal:8000
DIARIZATION_MODAL_API_KEY=dev-key-12345
```
Note: Use `host.docker.internal` because Reflector server runs in Docker.
### 4. Restart Reflector server
```bash
cd server
docker compose restart server worker
```
## Testing
### Test transcription
```bash
curl -s -X POST http://localhost:8000/v1/audio/transcriptions \
-H "Authorization: Bearer dev-key-12345" \
-F "file=@/path/to/audio.wav" \
-F "language=en"
```
### Test diarization
```bash
curl -s -X POST "http://localhost:8000/diarize?audio_file_url=<audio-url>" \
-H "Authorization: Bearer dev-key-12345"
```
## Platform Notes
### macOS (ARM)
Docker build fails - CUDA packages are x86_64 only. Use local Python instead:
```bash
uv sync
HF_TOKEN=xxx REFLECTOR_GPU_APIKEY=xxx .venv/bin/uvicorn main:app --host 0.0.0.0 --port 8000
```
### Linux with NVIDIA GPU
Docker works with CUDA acceleration:
```bash
docker compose up -d
```
### CPU-only
Works on any platform, just slower. PyTorch auto-detects and falls back to CPU.
## Switching Back to Modal.com
Edit `server/.env`:
```bash
TRANSCRIPT_BACKEND=modal
TRANSCRIPT_URL=https://monadical-sas--reflector-transcriber-parakeet-web.modal.run
TRANSCRIPT_MODAL_API_KEY=<modal-api-key>
DIARIZATION_BACKEND=modal
DIARIZATION_URL=https://monadical-sas--reflector-diarizer-web.modal.run
DIARIZATION_MODAL_API_KEY=<modal-api-key>
```
## Troubleshooting
### "Could not download pyannote pipeline"
- Accept model licenses at HuggingFace (see Prerequisites)
- Verify HF_TOKEN is set and valid
### Service won't start
- Check port 8000 is free: `lsof -i :8000`
- Kill orphan processes if needed
### Transcription returns empty text
- Ensure audio contains speech (not just tones/silence)
- Check audio format is supported (wav, mp3, etc.)
### Deprecation warnings from torchaudio/pyannote
- Safe to ignore - doesn't affect functionality

8
gpu/self_hosted/README.md
View File

@@ -56,9 +56,13 @@ Docker
- Not yet provided in this directory. A Dockerfile will be added later. For now, use Local run above
Conformance tests
# Setup
# From this directory
[SETUP.md](SETUP.md)
# Conformance tests
## From this directory
TRANSCRIPT_URL=http://localhost:8000 \
TRANSCRIPT_API_KEY=dev-key \

9
gpu/self_hosted/app/services/transcriber.py
View File

@@ -129,6 +129,11 @@ class WhisperService:
audio = np.frombuffer(proc.stdout, dtype=np.float32)
return audio
# IMPORTANT: This VAD segment logic is duplicated in multiple files for deployment isolation.
# If you modify this function, you MUST update all copies:
# - gpu/modal_deployments/reflector_transcriber.py
# - gpu/modal_deployments/reflector_transcriber_parakeet.py
# - gpu/self_hosted/app/services/transcriber.py (this file)
def vad_segments(
audio_array,
sample_rate: int = SAMPLE_RATE,
@@ -153,6 +158,10 @@ class WhisperService:
end = speech["end"]
yield (start / float(SAMPLE_RATE), end / float(SAMPLE_RATE))
start = None
# Handle case where audio ends while speech is still active
if start is not None:
audio_duration = len(audio_array) / float(sample_rate)
yield (start / float(SAMPLE_RATE), audio_duration)
iterator.reset_states()
audio_array = load_audio_via_ffmpeg(file_path, SAMPLE_RATE)

8
gpu/self_hosted/app/utils.py
View File

@@ -34,6 +34,12 @@ def ensure_dirs():
UPLOADS_PATH.mkdir(parents=True, exist_ok=True)
# IMPORTANT: This function is duplicated in multiple files for deployment isolation.
# If you modify the audio format detection logic, you MUST update all copies:
# - gpu/self_hosted/app/utils.py (this file)
# - gpu/modal_deployments/reflector_transcriber.py (2 copies)
# - gpu/modal_deployments/reflector_transcriber_parakeet.py
# - gpu/modal_deployments/reflector_diarizer.py
def detect_audio_format(url: str, headers: Mapping[str, str]) -> str:
url_path = urlparse(url).path
for ext in SUPPORTED_FILE_EXTENSIONS:
@@ -47,6 +53,8 @@ def detect_audio_format(url: str, headers: Mapping[str, str]) -> str:
return "wav"
if "audio/mp4" in content_type:
return "mp4"
if "audio/webm" in content_type or "video/webm" in content_type:
return "webm"
raise HTTPException(
status_code=400,