Skip to content

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.

MXFP4 low memory test: gate the chunked dX budget test on a real CUDA device (#15516)

Daniel Hanbaa95ba2e7b2
diff --git a/tests/test_mxfp4_packed_low_memory.py b/tests/test_mxfp4_packed_low_memory.py
index b3ce0dfb7d6..13cedb71343 100644
--- a/tests/test_mxfp4_packed_low_memory.py
+++ b/tests/test_mxfp4_packed_low_memory.py
@@ -425,7 +425,7 @@ def test_chunked_dx_does_not_overflow_in_a_16bit_chunk_product():
assert torch.equal(y, torch.zeros_like(y))
[email protected](not torch.cuda.is_available(), reason = "needs a CUDA device")
[email protected](not has_real_cuda(), reason = "needs a CUDA device")
def test_chunked_dx_stays_inside_its_own_budget(monkeypatch):
"""Each dX chunk row holds its 16-bit decode and the fp32 copy the fp32 product reads."""
import unsloth # noqa: F401