Skip to content

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.

Pin CUDA availability in the step graph capture-stream release test (#15015)

Daniel Hanf2ecfe15a19b
diff --git a/studio/backend/tests/test_diffusion_offload_step_graphs.py b/studio/backend/tests/test_diffusion_offload_step_graphs.py
index 12642272de7..e83c217c65f 100644
--- a/studio/backend/tests/test_diffusion_offload_step_graphs.py
+++ b/studio/backend/tests/test_diffusion_offload_step_graphs.py
@@ -669,6 +669,10 @@ def test_releasing_an_offloaded_step_graph_frees_its_capture_stream_workspace(mo
torch._C, "_cuda_clearCublasWorkspaces", lambda: calls.append(1), raising = False
)
monkeypatch.setattr(torch.cuda, "empty_cache", lambda: calls.append(2))
+ # The flush only runs where CUDA is available (#14777). Pin it here: on a CPU runner the answer otherwise
+ # depends on whether an earlier test in the worker installed the CUDA spoof.
+ monkeypatch.setattr(torch.cuda, "is_available", lambda: True)
+ monkeypatch.setattr(torch.cuda, "synchronize", lambda *a, **k: None)
handle = cg.GraphedForward.__new__(cg.GraphedForward)
handle.placement = _Placement(streams = True)
handle._release()