Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
cuDNN SDPA DTensor test: gate on a real CUDA device (#15582)
Daniel Hanb3c7886dc6a3
diff --git a/tests/test_cudnn_sdpa_grad_layout_dtensor.py b/tests/test_cudnn_sdpa_grad_layout_dtensor.pyindex b02705a9ff6..5b68697df8e 100644--- a/tests/test_cudnn_sdpa_grad_layout_dtensor.py+++ b/tests/test_cudnn_sdpa_grad_layout_dtensor.py@@ -13,6 +13,8 @@ import textwrapimport pytest+from real_accelerator import has_real_cuda # tests/_shared, on sys.path via tests/conftest.py+torch = pytest.importorskip("torch")_PROBE = textwrap.dedent(@@ -70,7 +72,7 @@ def _free_port():@pytest.mark.gpu[email protected](not torch.cuda.is_available(), reason = "needs CUDA")[email protected](not has_real_cuda(), reason = "needs CUDA")@pytest.mark.skipif(not (torch.distributed.is_available() and torch.distributed.is_nccl_available()),reason = "needs NCCL",