scott
169e003a34
All checks were successful
Build ROCm Image / build (push) Successful in 3m35s
- Warmup now uses a ~170-char representative sentence so torch.compile JIT-compiles for typical token sequence lengths. Previously "Warmup." compiled for very short shapes, causing a full re-compile (17s) on the first real HA request and pushing total synthesis past 30s. - Compile model.ve (voice encoder) in addition to s3gen — both are convolutional and hit the MIOpen workspace=0 bug. - Fix _patch_timing: attribute is model.ve not model.voice_encoder, so the timing wrap was silently skipping the speaker embedding. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
66 lines
2.1 KiB
Python
66 lines
2.1 KiB
Python
import asyncio
|
|
import logging
|
|
import sys
|
|
from functools import partial
|
|
|
|
from wyoming.server import AsyncServer
|
|
|
|
import engine
|
|
from config import get_wyoming_host, get_wyoming_port, load_config
|
|
from wyoming_handler import ChatterboxWyomingHandler
|
|
from wyoming_voices import create_wyoming_info, load_voices
|
|
|
|
logging.basicConfig(
|
|
level=logging.INFO,
|
|
format="%(asctime)s %(levelname)s %(name)s: %(message)s",
|
|
stream=sys.stdout,
|
|
)
|
|
logger = logging.getLogger(__name__)
|
|
|
|
|
|
def _warmup(voices: dict) -> None:
|
|
from wyoming_voices import resolve_voice
|
|
audio_prompt = resolve_voice(None, voices) if voices else None
|
|
logger.info("Running warmup synthesis to populate MIOpen kernel cache...")
|
|
try:
|
|
engine.synthesize(
|
|
text=(
|
|
"This is a warmup synthesis request used to pre-compile neural network kernels "
|
|
"for typical text lengths, so that the first real request runs at full speed."
|
|
),
|
|
audio_prompt_path=audio_prompt,
|
|
)
|
|
logger.info("Warmup complete — MIOpen cache populated")
|
|
except Exception:
|
|
logger.warning("Warmup synthesis failed (non-fatal)", exc_info=True)
|
|
|
|
|
|
async def main() -> None:
|
|
load_config()
|
|
|
|
logger.info("Loading TTS model...")
|
|
if not engine.load_model():
|
|
logger.error("Failed to load model, exiting")
|
|
sys.exit(1)
|
|
|
|
voices = load_voices()
|
|
wyoming_info = create_wyoming_info(engine.get_sample_rate(), voices)
|
|
|
|
# Run a warmup synthesis before accepting connections so MIOpen benchmarks
|
|
# and caches the best convolution algorithms for all layer shapes. Without
|
|
# this, the first real HA request triggers benchmarking (hundreds of runs)
|
|
# and times out before any audio is returned.
|
|
_warmup(voices)
|
|
|
|
host = get_wyoming_host()
|
|
port = get_wyoming_port()
|
|
uri = f"tcp://{host}:{port}"
|
|
|
|
logger.info(f"Starting Wyoming server on {uri}")
|
|
server = AsyncServer.from_uri(uri)
|
|
await server.run(partial(ChatterboxWyomingHandler, wyoming_info, voices))
|
|
|
|
|
|
if __name__ == "__main__":
|
|
asyncio.run(main())
|