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
Studio backend tests: pin the logical CPU count in the oversubscribed-threads spill test (#11120)
Daniel Han653d4ec6c3f8
diff --git a/studio/backend/tests/test_offload_planner_seam.py b/studio/backend/tests/test_offload_planner_seam.pyindex eed5ada47eb..cdd6fd8ea3f 100644--- a/studio/backend/tests/test_offload_planner_seam.py+++ b/studio/backend/tests/test_offload_planner_seam.py@@ -1836,6 +1836,13 @@ def test_oversubscribed_decode_threads_decline_spill_planning(monkeypatch):monkeypatch.setattr(llama_mod.os, "sched_getaffinity", lambda _pid: set(range(16)), raising = False)+ # The affinity above has to read as UNRESTRICTED, which means it must match the+ # host's logical count, and _spilled_decode_threads takes that from psutil. Left+ # real it is whatever the CI box has: on anything wider than 16 threads the 16-CPU+ # affinity looks like a taskset, the pricing declines for that reason instead, and+ # the oversubscription this test is about is never reached. _SmtHost is the 16/8+ # host the surrounding affinity tests already pin for the same reason.+ monkeypatch.setitem(sys.modules, "psutil", _SmtHost)assert _plan(_Stub(), free_mib = 14 * 1024, extra_args = ["--threads", "16"]) is Noneplan = _plan(_Stub(),