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: skip the compiled half of the fp8 codes layer test on Windows without MSVC (#17073)
Daniel Hana1c604722de7
diff --git a/studio/backend/tests/test_diffusion_native_quant.py b/studio/backend/tests/test_diffusion_native_quant.pyindex 9c886dd14c5..341f8a4d33b 100644--- a/studio/backend/tests/test_diffusion_native_quant.py+++ b/studio/backend/tests/test_diffusion_native_quant.py@@ -5,6 +5,7 @@from __future__ import annotations+import shutilimport sysimport types@@ -1317,6 +1318,11 @@ def test_fp8_codes_layer_matches_the_dequant_formula_eager_and_compiled(fmt):torch._dynamo.reset()explained = torch._dynamo.explain(layer)(x)assert explained.graph_break_count == 0+ if sys.platform == "win32" and shutil.which("cl") is None:+ # Inductor's CPU backend builds C++ with MSVC, which a plain Windows runner does not have on PATH.+ pytest.skip(+ "compiled half needs MSVC (cl) on Windows; the eager and graph-break checks above ran"+ )compiled = torch.compile(layer, fullgraph = True)assert torch.equal(compiled(x), layer(x))