OpenAI compatible transcription api

This commit is contained in:
2025-01-17 01:02:40 +01:00
parent b9b9292821
commit 99ff06ff17
4 changed files with 209 additions and 134 deletions

View File

@@ -1,89 +1,64 @@
"""
Reflector GPU backend - transcriber
===================================
"""
import os
import tempfile
import threading
from modal import Image, Secret, App, asgi_app, method, enter
import modal
from pydantic import BaseModel
# Whisper
WHISPER_MODEL: str = "large-v2"
WHISPER_COMPUTE_TYPE: str = "float16"
WHISPER_NUM_WORKERS: int = 1
MODELS_DIR = "/models"
MODEL_NAME = "large-v2"
MODEL_COMPUTE_TYPE: str = "float16"
MODEL_NUM_WORKERS: int = 1
MINUTES = 60 # seconds
volume = modal.Volume.from_name("models", create_if_missing=True)
app = modal.App("reflector-transcriber")
WHISPER_MODEL_DIR = "/root/transcription_models"
def download_model():
from faster_whisper import download_model
app = App(name="reflector-transcriber")
volume.reload()
download_model(MODEL_NAME, cache_dir=MODELS_DIR)
volume.commit()
def download_whisper():
from faster_whisper.utils import download_model
print("Downloading Whisper model")
download_model(WHISPER_MODEL, cache_dir=WHISPER_MODEL_DIR)
print("Whisper model downloaded")
def migrate_cache_llm():
"""
XXX The cache for model files in Transformers v4.22.0 has been updated.
Migrating your old cache. This is a one-time only operation. You can
interrupt this and resume the migration later on by calling
`transformers.utils.move_cache()`.
"""
from transformers.utils.hub import move_cache
print("Moving LLM cache")
move_cache(cache_dir=WHISPER_MODEL_DIR, new_cache_dir=WHISPER_MODEL_DIR)
print("LLM cache moved")
transcriber_image = (
Image.debian_slim(python_version="3.10.8")
.apt_install("git")
.apt_install("wget")
.apt_install("libsndfile-dev")
image = (
modal.Image.debian_slim(python_version="3.12")
.pip_install(
"faster-whisper",
"requests",
"torch",
"transformers==4.34.0",
"sentencepiece",
"protobuf",
"huggingface_hub==0.16.4",
"gitpython",
"torchaudio",
"fairseq2",
"pyyaml",
"hf-transfer~=0.1"
"huggingface_hub==0.27.1",
"hf-transfer==0.1.9",
"torch==2.5.1",
"faster-whisper==1.1.1",
)
.run_function(download_whisper)
.run_function(migrate_cache_llm)
.env(
{
"HF_HUB_ENABLE_HF_TRANSFER": "1",
"LD_LIBRARY_PATH": (
"/usr/local/lib/python3.10/site-packages/nvidia/cudnn/lib/:"
"/opt/conda/lib/python3.10/site-packages/nvidia/cublas/lib/"
)
"/usr/local/lib/python3.12/site-packages/nvidia/cudnn/lib/:"
"/opt/conda/lib/python3.12/site-packages/nvidia/cublas/lib/"
),
}
)
.run_function(download_model, volumes={MODELS_DIR: volume})
)
@app.cls(
gpu="A10G",
timeout=60 * 5,
container_idle_timeout=60 * 5,
timeout=5 * MINUTES,
container_idle_timeout=5 * MINUTES,
allow_concurrent_inputs=6,
image=transcriber_image,
image=image,
volumes={MODELS_DIR: volume},
)
class Transcriber:
@enter()
@modal.enter()
def enter(self):
import faster_whisper
import torch
@@ -92,21 +67,20 @@ class Transcriber:
self.use_gpu = torch.cuda.is_available()
self.device = "cuda" if self.use_gpu else "cpu"
self.model = faster_whisper.WhisperModel(
WHISPER_MODEL,
MODEL_NAME,
device=self.device,
compute_type=WHISPER_COMPUTE_TYPE,
num_workers=WHISPER_NUM_WORKERS,
download_root=WHISPER_MODEL_DIR,
local_files_only=True
compute_type=MODEL_COMPUTE_TYPE,
num_workers=MODEL_NUM_WORKERS,
download_root=MODELS_DIR,
local_files_only=True,
)
@method()
@modal.method()
def transcribe_segment(
self,
audio_data: str,
audio_suffix: str,
source_language: str,
timestamp: float = 0
language: str,
):
with tempfile.NamedTemporaryFile("wb+", suffix=f".{audio_suffix}") as fp:
fp.write(audio_data)
@@ -114,40 +88,23 @@ class Transcriber:
with self.lock:
segments, _ = self.model.transcribe(
fp.name,
language=source_language,
language=language,
beam_size=5,
word_timestamps=True,
vad_filter=True,
vad_parameters={"min_silence_duration_ms": 500},
)
multilingual_transcript = {}
transcript_source_lang = ""
text = ""
words = []
if segments:
segments = list(segments)
for segment in segments:
text += segment.text
words.extend(
{"word": word.word, "start": word.start, "end": word.end}
for word in segment.words
)
for segment in segments:
transcript_source_lang += segment.text
for word in segment.words:
words.append(
{
"text": word.word,
"start": round(timestamp + word.start, 3),
"end": round(timestamp + word.end, 3),
}
)
multilingual_transcript[source_language] = transcript_source_lang
return {
"text": multilingual_transcript,
"words": words
}
# -------------------------------------------------------------------
# Web API
# -------------------------------------------------------------------
return {"text": text, "words": words}
@app.function(
@@ -155,21 +112,23 @@ class Transcriber:
timeout=60,
allow_concurrent_inputs=40,
secrets=[
Secret.from_name("reflector-gpu"),
modal.Secret.from_name("reflector-gpu"),
],
volumes={MODELS_DIR: volume},
)
@asgi_app()
@modal.asgi_app()
def web():
from fastapi import Body, Depends, FastAPI, HTTPException, UploadFile, status
from fastapi.security import OAuth2PasswordBearer
from typing_extensions import Annotated
transcriberstub = Transcriber()
transcriber = Transcriber()
app = FastAPI()
oauth2_scheme = OAuth2PasswordBearer(tokenUrl="token")
supported_audio_file_types = ["wav", "mp3", "ogg", "flac"]
supported_file_types = ["mp3", "mp4", "mpeg", "mpga", "m4a", "wav", "webm"]
def apikey_auth(apikey: str = Depends(oauth2_scheme)):
if apikey != os.environ["REFLECTOR_GPU_APIKEY"]:
@@ -182,21 +141,20 @@ def web():
class TranscriptResponse(BaseModel):
result: dict
@app.post("/transcribe", dependencies=[Depends(apikey_auth)])
@app.post("/v1/audio/transcriptions", dependencies=[Depends(apikey_auth)])
def transcribe(
file: UploadFile,
source_language: Annotated[str, Body(...)] = "en",
timestamp: Annotated[float, Body()] = 0.0
model: str = "whisper-1",
language: Annotated[str, Body(...)] = "en",
) -> TranscriptResponse:
audio_data = file.file.read()
audio_suffix = file.filename.split(".")[-1]
assert audio_suffix in supported_audio_file_types
assert audio_suffix in supported_file_types
func = transcriberstub.transcribe_segment.spawn(
func = transcriber.transcribe_segment.spawn(
audio_data=audio_data,
audio_suffix=audio_suffix,
source_language=source_language,
timestamp=timestamp
language=language,
)
result = func.get()
return result