Open-source State-of-the-art TTS model which runs on a CPU 😻
Python
"""Shared test setup."""
import pytest
import torch
@pytest.fixture(autouse=True)
def _nvidia_platform_for_nvidia_tests(request, monkeypatch):
"""The vLLM tests describe NVIDIA GPUs (none of them use a GPU). On a ROCm build of PyTorch
they would take the AMD paths, so make it look like a CUDA build for them;
tests/test_vllm_rocm.py sets the platform itself."""
if getattr(torch.version, "hip", None) and not request.module.__name__.endswith("test_vllm_rocm"):
monkeypatch.setattr(torch.version, "hip", None)
@pytest.fixture(autouse=True)
def _no_real_gpu_memory(monkeypatch):
"""Unit tests never size vLLM from the real GPU of the machine they run on (on a GPU host
the small-GPU fitting would otherwise read its free memory and size the tests' stub model
configs). Tests that need free/total numbers patch _free_gpu_memory themselves."""
def no_gpu():
raise RuntimeError("unit tests do not read the real GPU's free memory")
try:
import kittenml.kittentts2.vllm.backend as backend
except ImportError:
return
monkeypatch.setattr(backend, "_free_gpu_memory", no_gpu)