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
SDPA packed-segment test: gate the CUDA memory test on a real GPU (#12761)
Daniel Han268f88e209d2
diff --git a/tests/utils/test_sdpa_packed_segments.py b/tests/utils/test_sdpa_packed_segments.pyindex cc852ee09b7..a6ce28f9df1 100644--- a/tests/utils/test_sdpa_packed_segments.py+++ b/tests/utils/test_sdpa_packed_segments.py@@ -3,6 +3,7 @@import pytestimport torch+from real_accelerator import has_real_cuda # tests/_shared, on sys.path via tests/conftest.pyimport unsloth # noqa: F401from unsloth.utils import attention_dispatch as ad@@ -95,7 +96,9 @@ def test_segments_never_build_the_dense_mask(monkeypatch):assert out.shape == (1, 9, 2, 8)[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, and this body allocates on "cuda" (test_accelerator_skip_guards).[email protected](not has_real_cuda(), reason = "needs CUDA")def test_peak_memory_is_linear_in_tokens(monkeypatch):# 2 x 2048-token segments, 16 heads over 8 KV heads, as Qwen3-0.6B under padding-free SFT.total, n_heads, n_kv = 4096, 16, 8