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
Gate the Mllama CUDA forward test on has_real_cuda (#12143)
Daniel Han839cdbfb519a
diff --git a/tests/test_mllama_flash_attention_excluded.py b/tests/test_mllama_flash_attention_excluded.pyindex ce8f9e613df..86adb5ae95a 100644--- a/tests/test_mllama_flash_attention_excluded.py+++ b/tests/test_mllama_flash_attention_excluded.py@@ -5,6 +5,7 @@ import unsloth # noqa: F401import transformersfrom unsloth.models import _utils+from real_accelerator import has_real_cudapytestmark = pytest.mark.skipif(not hasattr(transformers, "MllamaConfig"), reason = "needs transformers with Mllama"@@ -64,7 +65,10 @@ def test_flash_entries_of_a_mapping_fall_back(monkeypatch):assert impl == {"": "sdpa", "vision_config": "eager"}[email protected](not torch.cuda.is_available(), reason = "needs CUDA")+# has_real_cuda(), not torch.cuda.is_available(): tests/_zoo_aggressive_cuda_spoof.py patches+# the latter to True process-wide, which would un-skip this on a CPU-only box. The CUDA probe+# rather than has_real_accelerator() because the body allocates on "cuda".[email protected](not has_real_cuda(), reason = "needs CUDA")def test_mllama_image_forward_runs_on_the_resolved_implementation(monkeypatch):monkeypatch.setattr(_utils, "HAS_FLASH_ATTENTION", True)model_class, config = _tiny_mllama()