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.

Tests: skip the QAT MX and FP8 block GEMM GPU cases on the recorded real CUDA, not the spoofable probe (#14785)

Daniel Hanc993bf3e2b23
diff --git a/tests/test_fp8_torchao_import_route.py b/tests/test_fp8_torchao_import_route.py
index c1d1e71435a..a837d0d72f0 100644
--- a/tests/test_fp8_torchao_import_route.py
+++ b/tests/test_fp8_torchao_import_route.py
@@ -10,6 +10,8 @@ import pytest
import torch
from real_accelerator import has_real_cuda
+from real_accelerator import has_real_cuda
+
from unsloth.kernels import fp8 as F8
OLD = "torchao.prototype.blockwise_fp8_inference.blockwise_quantization"
diff --git a/tests/test_qat_microscaling_schemes.py b/tests/test_qat_microscaling_schemes.py
index 9f68efd85d5..9ab7885965c 100644
--- a/tests/test_qat_microscaling_schemes.py
+++ b/tests/test_qat_microscaling_schemes.py
@@ -179,7 +179,7 @@ def test_mx_qat_linear_takes_float32_activations_under_autocast(weight_dtype):
# Autocast feeds float32 activations into bfloat16 (or float32) linears.
# torchao's CPU fake-quant backward cannot mix a float32 weight with bfloat16 activations.
device = "cpu" if weight_dtype == torch.bfloat16 else "cuda"
- if device == "cuda" and not torch.cuda.is_available():
+ if device == "cuda" and not has_real_cuda():
pytest.skip("float32 weights under autocast need the CUDA MX path")
model = torch.nn.Sequential(torch.nn.Linear(64, 128, bias = False)).to(device, weight_dtype)
model = _prepare_model_for_qat(model, _mx_torchao_config())