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 diffusion: off-GIL CUDA prime does not latch before torch is imported (#14683)
Daniel Han0c2dfdaa1e05
diff --git a/studio/backend/tests/test_cuda_prime_off_gil.py b/studio/backend/tests/test_cuda_prime_off_gil.pyindex 8858e8815a1..c48e5d6fd78 100644--- a/studio/backend/tests/test_cuda_prime_off_gil.py+++ b/studio/backend/tests/test_cuda_prime_off_gil.py@@ -138,6 +138,28 @@ def test_prime_never_imports_torch(fake_cdll, monkeypatch):assert opened == [] and "torch" not in sys.modules+def test_a_call_before_torch_is_imported_does_not_latch(fake_cdll, monkeypatch):+ calls, opened, _gate = fake_cdll+ monkeypatch.delitem(sys.modules, "torch", raising = False)+ assert prime.prime_cuda_driver() is False+ assert opened == [] and not prime._done.is_set()+ assert prime.prime_cuda_driver(_fake_torch(calls)) is True+ assert calls == [("cuInit", 0)]+ assert prime.prime_cuda_driver(_fake_torch(calls)) is False+ assert calls == [("cuInit", 0)]+++def test_a_call_while_torch_is_still_importing_does_not_latch(fake_cdll, monkeypatch):+ calls, opened, _gate = fake_cdll+ half = SimpleNamespace(version = SimpleNamespace(cuda = "12.8", hip = None))+ monkeypatch.setitem(sys.modules, "torch", half)+ assert prime.prime_cuda_driver() is False+ assert opened == [] and not prime._done.is_set()+ monkeypatch.delitem(sys.modules, "torch")+ assert prime.prime_cuda_driver(_fake_torch(calls)) is True+ assert calls == [("cuInit", 0)]++def test_prime_never_raises_without_a_driver(monkeypatch):import ctypesdiff --git a/studio/backend/utils/cuda_driver_prime.py b/studio/backend/utils/cuda_driver_prime.pyindex c2f73ac493d..a5679955b0c 100644--- a/studio/backend/utils/cuda_driver_prime.py+++ b/studio/backend/utils/cuda_driver_prime.py@@ -66,6 +66,12 @@ def prime_cuda_driver(torch_module = None) -> bool:global _claimedif _done.is_set():return False+ if torch_module is None:+ # Never imports torch (would deadlock mid-import waiters). Torch absent or mid-import (no ``torch.cuda`` yet)+ # must not latch, so hardware detection's later call still primes.+ torch_module = sys.modules.get("torch")+ if torch_module is None or getattr(torch_module, "cuda", None) is None:+ return Falsewith _claim_lock:first = not _claimed_claimed = True@@ -75,11 +81,6 @@ def prime_cuda_driver(torch_module = None) -> bool:try:if not _driver_names():return False- if torch_module is None:- # Never imports torch: a caller still inside `import torch` would deadlock its waiters.- torch_module = sys.modules.get("torch")- if torch_module is None:- return Falseif not _should_prime(torch_module):return Falsereturn _cu_init()