You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
langchain/libs/community/langchain_community/utilities/nvidia_riva.py

668 lines
21 KiB
Python

"""A common module for NVIDIA Riva Runnables."""
import asyncio
import logging
import pathlib
import queue
import tempfile
import threading
import wave
from enum import Enum
from typing import (
TYPE_CHECKING,
Any,
AsyncGenerator,
AsyncIterator,
Dict,
Generator,
Iterator,
List,
Optional,
Tuple,
Union,
cast,
)
from langchain_core.messages import AnyMessage, BaseMessage
from langchain_core.prompt_values import PromptValue
from langchain_core.pydantic_v1 import (
AnyHttpUrl,
BaseModel,
Field,
parse_obj_as,
root_validator,
validator,
)
from langchain_core.runnables import RunnableConfig, RunnableSerializable
if TYPE_CHECKING:
import riva.client
import riva.client.proto.riva_asr_pb2 as rasr
_LOGGER = logging.getLogger(__name__)
_QUEUE_GET_TIMEOUT = 0.5
_MAX_TEXT_LENGTH = 400
_SENTENCE_TERMINATORS = ("\n", ".", "!", "?", "¡", "¿")
# COMMON utilities used by all Riva Runnables
def _import_riva_client() -> "riva.client":
"""Import the riva client and raise an error on failure."""
try:
# pylint: disable-next=import-outside-toplevel # this client library is optional
import riva.client
except ImportError as err:
raise ImportError(
"Could not import the NVIDIA Riva client library. "
"Please install it with `pip install nvidia-riva-client`."
) from err
return riva.client
class SentinelT: # pylint: disable=too-few-public-methods
"""An empty Sentinel type."""
HANGUP = SentinelT()
_TRANSFORM_END = SentinelT()
class RivaAudioEncoding(str, Enum):
"""An enum of the possible choices for Riva audio encoding.
The list of types exposed by the Riva GRPC Protobuf files can be found
with the following commands:
```python
import riva.client
print(riva.client.AudioEncoding.keys()) # noqa: T201
```
"""
ALAW = "ALAW"
ENCODING_UNSPECIFIED = "ENCODING_UNSPECIFIED"
FLAC = "FLAC"
LINEAR_PCM = "LINEAR_PCM"
MULAW = "MULAW"
OGGOPUS = "OGGOPUS"
@classmethod
def from_wave_format_code(cls, format_code: int) -> "RivaAudioEncoding":
"""Return the audio encoding specified by the format code in the wave file.
ref: https://mmsp.ece.mcgill.ca/Documents/AudioFormats/WAVE/WAVE.html
"""
try:
return {1: cls.LINEAR_PCM, 6: cls.ALAW, 7: cls.MULAW}[format_code]
except KeyError as err:
raise NotImplementedError(
"The following wave file format code is "
f"not supported by Riva: {format_code}"
) from err
@property
def riva_pb2(self) -> "riva.client.AudioEncoding":
"""Returns the Riva API object for the encoding."""
riva_client = _import_riva_client()
return getattr(riva_client.AudioEncoding, self)
class RivaAuthMixin(BaseModel):
"""Configuration for the authentication to a Riva service connection."""
url: Union[AnyHttpUrl, str] = Field(
AnyHttpUrl("http://localhost:50051", scheme="http"),
description="The full URL where the Riva service can be found.",
examples=["http://localhost:50051", "https://user@pass:riva.example.com"],
)
ssl_cert: Optional[str] = Field(
None,
description="A full path to the file where Riva's public ssl key can be read.",
)
@property
def auth(self) -> "riva.client.Auth":
"""Return a riva client auth object."""
riva_client = _import_riva_client()
url = cast(AnyHttpUrl, self.url)
use_ssl = url.scheme == "https" # pylint: disable=no-member # false positive
url_no_scheme = str(self.url).split("/")[2]
return riva_client.Auth(
ssl_cert=self.ssl_cert, use_ssl=use_ssl, uri=url_no_scheme
)
@validator("url", pre=True, allow_reuse=True)
@classmethod
def _validate_url(cls, val: Any) -> AnyHttpUrl:
"""Do some initial conversations for the URL before checking."""
if isinstance(val, str):
return cast(AnyHttpUrl, parse_obj_as(AnyHttpUrl, val))
return cast(AnyHttpUrl, val)
class RivaCommonConfigMixin(BaseModel):
"""A collection of common Riva settings."""
encoding: RivaAudioEncoding = Field(
default=RivaAudioEncoding.LINEAR_PCM,
description="The encoding on the audio stream.",
)
sample_rate_hertz: int = Field(
default=8000, description="The sample rate frequency of audio stream."
)
language_code: str = Field(
default="en-US",
description=(
"The [BCP-47 language code]"
"(https://www.rfc-editor.org/rfc/bcp/bcp47.txt) for "
"the target language."
),
)
class _Event:
"""A combined event that is threadsafe and async safe."""
_event: threading.Event
_aevent: asyncio.Event
def __init__(self) -> None:
"""Initialize the event."""
self._event = threading.Event()
self._aevent = asyncio.Event()
def set(self) -> None:
"""Set the event."""
self._event.set()
self._aevent.set()
def clear(self) -> None:
"""Set the event."""
self._event.clear()
self._aevent.clear()
def is_set(self) -> bool:
"""Indicate if the event is set."""
return self._event.is_set()
def wait(self) -> None:
"""Wait for the event to be set."""
self._event.wait()
async def async_wait(self) -> None:
"""Async wait for the event to be set."""
await self._aevent.wait()
def _mk_wave_file(
output_directory: Optional[str], sample_rate: float
) -> Tuple[Optional[str], Optional[wave.Wave_write]]:
"""Create a new wave file and return the wave write object and filename."""
if output_directory:
with tempfile.NamedTemporaryFile(
mode="bx", suffix=".wav", delete=False, dir=output_directory
) as f:
wav_file_name = f.name
wav_file = wave.open(wav_file_name, "wb")
wav_file.setnchannels(1)
wav_file.setsampwidth(2)
wav_file.setframerate(sample_rate)
return (wav_file_name, wav_file)
return (None, None)
def _coerce_string(val: "TTSInputType") -> str:
"""Attempt to coerce the input value to a string.
This is particularly useful for converting LangChain message to strings.
"""
if isinstance(val, PromptValue):
return val.to_string()
if isinstance(val, BaseMessage):
return str(val.content)
return str(val)
def _process_chunks(inputs: Iterator["TTSInputType"]) -> Generator[str, None, None]:
"""Filter the input chunks are return strings ready for TTS."""
buffer = ""
for chunk in inputs:
chunk = _coerce_string(chunk)
# return the buffer if an end of sentence character is detected
for terminator in _SENTENCE_TERMINATORS:
while terminator in chunk:
last_sentence, chunk = chunk.split(terminator, 1)
yield buffer + last_sentence + terminator
buffer = ""
buffer += chunk
# return the buffer if is too long
if len(buffer) > _MAX_TEXT_LENGTH:
for idx in range(0, len(buffer), _MAX_TEXT_LENGTH):
yield buffer[idx : idx + 5]
buffer = ""
# return remaining buffer
if buffer:
yield buffer
# Riva AudioStream Type
StreamInputType = Union[bytes, SentinelT]
StreamOutputType = str
class AudioStream:
"""A message containing streaming audio."""
_put_lock: threading.Lock
_queue: queue.Queue
output: queue.Queue
hangup: _Event
user_talking: _Event
user_quiet: _Event
_worker: Optional[threading.Thread]
def __init__(self, maxsize: int = 0) -> None:
"""Initialize the queue."""
self._put_lock = threading.Lock()
self._queue = queue.Queue(maxsize=maxsize)
self.output = queue.Queue()
self.hangup = _Event()
self.user_quiet = _Event()
self.user_talking = _Event()
self._worker = None
def __iter__(self) -> Generator[bytes, None, None]:
"""Return an error."""
while True:
# get next item
try:
next_val = self._queue.get(True, _QUEUE_GET_TIMEOUT)
except queue.Empty:
continue
# hangup when requested
if next_val == HANGUP:
break
# yield next item
yield next_val
self._queue.task_done()
async def __aiter__(self) -> AsyncIterator[StreamInputType]:
"""Iterate through all items in the queue until HANGUP."""
while True:
# get next item
try:
next_val = await asyncio.get_event_loop().run_in_executor(
None, self._queue.get, True, _QUEUE_GET_TIMEOUT
)
except queue.Empty:
continue
# hangup when requested
if next_val == HANGUP:
break
# yield next item
yield next_val
self._queue.task_done()
@property
def hungup(self) -> bool:
"""Indicate if the audio stream has hungup."""
return self.hangup.is_set()
@property
def empty(self) -> bool:
"""Indicate in the input stream buffer is empty."""
return self._queue.empty()
@property
def complete(self) -> bool:
"""Indicate if the audio stream has hungup and been processed."""
input_done = self.hungup and self.empty
output_done = (
self._worker is not None
and not self._worker.is_alive()
and self.output.empty()
)
return input_done and output_done
@property
def running(self) -> bool:
"""Indicate if the ASR stream is running."""
if self._worker:
return self._worker.is_alive()
return False
def put(self, item: StreamInputType, timeout: Optional[int] = None) -> None:
"""Put a new item into the queue."""
with self._put_lock:
if self.hungup:
raise RuntimeError(
"The audio stream has already been hungup. Cannot put more data."
)
if item is HANGUP:
self.hangup.set()
self._queue.put(item, timeout=timeout)
async def aput(self, item: StreamInputType, timeout: Optional[int] = None) -> None:
"""Async put a new item into the queue."""
loop = asyncio.get_event_loop()
await asyncio.wait_for(loop.run_in_executor(None, self.put, item), timeout)
def close(self, timeout: Optional[int] = None) -> None:
"""Send the hangup signal."""
self.put(HANGUP, timeout)
async def aclose(self, timeout: Optional[int] = None) -> None:
"""Async send the hangup signal."""
await self.aput(HANGUP, timeout)
def register(self, responses: Iterator["rasr.StreamingRecognizeResponse"]) -> None:
"""Drain the responses from the provided iterator and put them into a queue."""
if self.running:
raise RuntimeError("An ASR instance has already been registered.")
has_started = threading.Barrier(2, timeout=5)
def worker() -> None:
"""Consume the ASR Generator."""
has_started.wait()
for response in responses:
if not response.results:
continue
for result in response.results:
if not result.alternatives:
continue
if result.is_final:
self.user_talking.clear()
self.user_quiet.set()
transcript = cast(str, result.alternatives[0].transcript)
self.output.put(transcript)
elif not self.user_talking.is_set():
self.user_talking.set()
self.user_quiet.clear()
self._worker = threading.Thread(target=worker)
self._worker.daemon = True
self._worker.start()
has_started.wait()
# RivaASR Runnable
ASRInputType = AudioStream
ASROutputType = str
class RivaASR(
RivaAuthMixin,
RivaCommonConfigMixin,
RunnableSerializable[ASRInputType, ASROutputType],
):
"""A runnable that performs Automatic Speech Recognition (ASR) using NVIDIA Riva."""
name: str = "nvidia_riva_asr"
description: str = (
"A Runnable for converting audio bytes to a string."
"This is useful for feeding an audio stream into a chain and"
"preprocessing that audio to create an LLM prompt."
)
# riva options
audio_channel_count: int = Field(
1, description="The number of audio channels in the input audio stream."
)
profanity_filter: bool = Field(
True,
description=(
"Controls whether or not Riva should attempt to filter "
"profanity out of the transcribed text."
),
)
enable_automatic_punctuation: bool = Field(
True,
description=(
"Controls whether Riva should attempt to correct "
"senetence puncuation in the transcribed text."
),
)
@root_validator(pre=True)
@classmethod
def _validate_environment(cls, values: Dict[str, Any]) -> Dict[str, Any]:
"""Validate the Python environment and input arguments."""
_ = _import_riva_client()
return values
@property
def config(self) -> "riva.client.StreamingRecognitionConfig":
"""Create and return the riva config object."""
riva_client = _import_riva_client()
return riva_client.StreamingRecognitionConfig(
interim_results=True,
config=riva_client.RecognitionConfig(
encoding=self.encoding,
sample_rate_hertz=self.sample_rate_hertz,
audio_channel_count=self.audio_channel_count,
max_alternatives=1,
profanity_filter=self.profanity_filter,
enable_automatic_punctuation=self.enable_automatic_punctuation,
language_code=self.language_code,
),
)
def _get_service(self) -> "riva.client.ASRService":
"""Connect to the riva service and return the a client object."""
riva_client = _import_riva_client()
try:
return riva_client.ASRService(self.auth)
except Exception as err:
raise ValueError(
"Error raised while connecting to the Riva ASR server."
) from err
def invoke(
self,
input: ASRInputType,
_: Optional[RunnableConfig] = None,
) -> ASROutputType:
"""Transcribe the audio bytes into a string with Riva."""
# create an output text generator with Riva
if not input.running:
service = self._get_service()
responses = service.streaming_response_generator(
audio_chunks=input,
streaming_config=self.config,
)
input.register(responses)
# return the first valid result
full_response: List[str] = []
while not input.complete:
with input.output.not_empty:
ready = input.output.not_empty.wait(0.1)
if ready:
while not input.output.empty():
try:
full_response += [input.output.get_nowait()]
except queue.Empty:
continue
input.output.task_done()
_LOGGER.debug("Riva ASR returning: %s", repr(full_response))
return " ".join(full_response).strip()
return ""
# RivaTTS Runnable
# pylint: disable-next=invalid-name
TTSInputType = Union[str, AnyMessage, PromptValue]
TTSOutputType = bytes
class RivaTTS(
RivaAuthMixin,
RivaCommonConfigMixin,
RunnableSerializable[TTSInputType, TTSOutputType],
):
"""A runnable that performs Text-to-Speech (TTS) with NVIDIA Riva."""
name: str = "nvidia_riva_tts"
description: str = (
"A tool for converting text to speech."
"This is useful for converting LLM output into audio bytes."
)
# riva options
voice_name: str = Field(
"English-US.Female-1",
description=(
"The voice model in Riva to use for speech. "
"Pre-trained models are documented in "
"[the Riva documentation]"
"(https://docs.nvidia.com/deeplearning/riva/user-guide/docs/tts/tts-overview.html)."
),
)
output_directory: Optional[str] = Field(
None,
description=(
"The directory where all audio files should be saved. "
"A null value indicates that wave files should not be saved. "
"This is useful for debugging purposes."
),
)
@root_validator(pre=True)
@classmethod
def _validate_environment(cls, values: Dict[str, Any]) -> Dict[str, Any]:
"""Validate the Python environment and input arguments."""
_ = _import_riva_client()
return values
@validator("output_directory")
@classmethod
def _output_directory_validator(cls, v: str) -> str:
if v:
dirpath = pathlib.Path(v)
dirpath.mkdir(parents=True, exist_ok=True)
return str(dirpath.absolute())
return v
def _get_service(self) -> "riva.client.SpeechSynthesisService":
"""Connect to the riva service and return the a client object."""
riva_client = _import_riva_client()
try:
return riva_client.SpeechSynthesisService(self.auth)
except Exception as err:
raise ValueError(
"Error raised while connecting to the Riva TTS server."
) from err
def invoke(
self, input: TTSInputType, _: Union[RunnableConfig, None] = None
) -> TTSOutputType:
"""Perform TTS by taking a string and outputting the entire audio file."""
return b"".join(self.transform(iter([input])))
def transform(
self,
input: Iterator[TTSInputType],
config: Optional[RunnableConfig] = None,
**kwargs: Optional[Any],
) -> Iterator[TTSOutputType]:
"""Perform TTS by taking a stream of characters and streaming output bytes."""
service = self._get_service()
# create an output wave file
wav_file_name, wav_file = _mk_wave_file(
self.output_directory, self.sample_rate_hertz
)
# split the input text and perform tts
for chunk in _process_chunks(input):
_LOGGER.debug("Riva TTS chunk: %s", chunk)
# start riva tts streaming
responses = service.synthesize_online(
text=chunk,
voice_name=self.voice_name,
language_code=self.language_code,
encoding=self.encoding.riva_pb2,
sample_rate_hz=self.sample_rate_hertz,
)
# stream audio bytes out
for resp in responses:
audio = cast(bytes, resp.audio)
if wav_file:
wav_file.writeframesraw(audio)
yield audio
# close the wave file when we are done
if wav_file:
wav_file.close()
_LOGGER.debug("Riva TTS wrote file: %s", wav_file_name)
async def atransform(
self,
input: AsyncIterator[TTSInputType],
config: Optional[RunnableConfig] = None,
**kwargs: Optional[Any],
) -> AsyncGenerator[TTSOutputType, None]:
"""Intercept async transforms and route them to the synchronous transform."""
loop = asyncio.get_running_loop()
input_queue: queue.Queue = queue.Queue()
out_queue: asyncio.Queue = asyncio.Queue()
async def _producer() -> None:
"""Produce input into the input queue."""
async for val in input:
input_queue.put_nowait(val)
input_queue.put_nowait(_TRANSFORM_END)
def _input_iterator() -> Iterator[TTSInputType]:
"""Iterate over the input_queue."""
while True:
try:
val = input_queue.get(timeout=0.5)
except queue.Empty:
continue
if val == _TRANSFORM_END:
break
yield val
def _consumer() -> None:
"""Consume the input with transform."""
for val in self.transform(_input_iterator()):
out_queue.put_nowait(val)
out_queue.put_nowait(_TRANSFORM_END)
async def _consumer_coro() -> None:
"""Coroutine that wraps the consumer."""
await loop.run_in_executor(None, _consumer)
producer = loop.create_task(_producer())
consumer = loop.create_task(_consumer_coro())
while True:
try:
val = await asyncio.wait_for(out_queue.get(), 0.5)
except asyncio.exceptions.TimeoutError:
continue
out_queue.task_done()
if val is _TRANSFORM_END:
break
yield val
await producer
await consumer