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
Tests: baseline the torchao import probe from #15850, gate the PrefixGrouper CUDA test on a real card (#15948)
Daniel Han754fc0ea8aef
diff --git a/scripts/risky_loaders_baseline.json b/scripts/risky_loaders_baseline.jsonindex 30775832da4..ca1c9845204 100644--- a/scripts/risky_loaders_baseline.json+++ b/scripts/risky_loaders_baseline.json@@ -519,8 +519,8 @@"rule": "dynamic-import","sink": "import_module","digest": "59e72c01edb14c86",- "count": 3,- "reason": "Module names come from fixed in-file tables (_PEFT_TORCHAO_TENSOR_SUBCLASSES and the transformers symbol homes); remote code can only name an attribute, which resolves inside an already installed transformers module."+ "count": 4,+ "reason": "Module names come from fixed in-file tables (_PEFT_TORCHAO_TENSOR_SUBCLASSES, _TORCHAO_IMPORT_TIME_MODULES and the transformers symbol homes); remote code can only name an attribute, which resolves inside an already installed transformers module."},{"file": "unsloth/import_fixes.py",diff --git a/tests/test_prefix_grouper_layout.py b/tests/test_prefix_grouper_layout.pyindex 450e59d1978..9f93d4511ad 100644--- a/tests/test_prefix_grouper_layout.py+++ b/tests/test_prefix_grouper_layout.py@@ -12,6 +12,7 @@ from pathlib import Pathimport pytestimport torch+from real_accelerator import has_real_cuda_UTILS = Path(__file__).resolve().parents[1] / "unsloth" / "utils"@@ -196,7 +197,8 @@ def test_device_tensor_creations_do_not_grow_with_groups():assert _factory_calls(1, 8) == _factory_calls(6, 8) == _factory_calls(6, 2)[email protected](not torch.cuda.is_available(), reason = "needs CUDA")+# Not torch.cuda.is_available(): the CUDA spoof makes it answer True on CPU runners.[email protected](not has_real_cuda(), reason = "needs CUDA")def test_cuda_layout_reads_the_device_once_and_copies_once():import warnings