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
Unsloth Studio (AMD/ROCm): don't turn on cudnn.benchmark for image and video generation (#11732)
Leo Borcherdingf9bffe265889
diff --git a/studio/backend/core/inference/diffusion_speed.py b/studio/backend/core/inference/diffusion_speed.pyindex 9cc666eed98..28fb7561eb8 100644--- a/studio/backend/core/inference/diffusion_speed.py+++ b/studio/backend/core/inference/diffusion_speed.py@@ -728,6 +728,12 @@ def _compile_vae_decode(pipe: Any, logger: Any) -> bool:def _enable_cudnn_benchmark(logger: Any) -> bool:try:import torch++ from core._torchao_stub import _module_is_rocm++ # On ROCm this is MIOpen's exhaustive search: a first VAE decode tuned for 10 to 23 minutes and crashed a gfx1030.+ if _module_is_rocm(torch):+ return Falsetorch.backends.cudnn.benchmark = Truereturn Trueexcept Exception as exc: # noqa: BLE001 - optimisation onlydiff --git a/studio/backend/tests/test_diffusion_speed.py b/studio/backend/tests/test_diffusion_speed.pyindex e525116857d..260f974de07 100644--- a/studio/backend/tests/test_diffusion_speed.py+++ b/studio/backend/tests/test_diffusion_speed.py@@ -357,6 +357,23 @@ def test_speed_default_cudnn_benchmark_only_on_cuda(monkeypatch):assert applied["cudnn_benchmark"] is False # not CUDA -> no autotune flip+ "hip, version",+ [("7.13.0", "2.11.0+rocm7.13.0"), (None, "2.9.1+rocmsdk20251116")],+ ids = ["version_hip", "rocm_tag_only"],+)+def test_speed_default_skips_cudnn_benchmark_on_rocm(monkeypatch, hip, version):+ torch = _stub_torch(monkeypatch)+ torch.version = types.SimpleNamespace(hip = hip)+ torch.__version__ = version+ pipe = _Pipe(with_compile = True)+ applied = apply_speed_optims(+ pipe, _target(), is_gguf = True, family = _family(), speed_mode = SPEED_DEFAULT+ )+ assert applied["cudnn_benchmark"] is False+ assert torch.backends.cudnn.benchmark is False++def test_speed_max_enables_tf32_and_fused_qkv(monkeypatch):torch = _stub_torch(monkeypatch)pipe = _Pipe(with_compile = True, with_fuse = True)