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: pin bf16 support and the HF cache location in the GGUF preflight fixtures (#15967)
Daniel Han921a6ceb9f2f
diff --git a/tests/saving/test_gguf_disk_preflight.py b/tests/saving/test_gguf_disk_preflight.pyindex 45412816520..9df6d18dc56 100644--- a/tests/saving/test_gguf_disk_preflight.py+++ b/tests/saving/test_gguf_disk_preflight.py@@ -112,6 +112,8 @@ def stub_sizing(monkeypatch):monkeypatch.setattr(S, "estimate_gguf_export_bytes", fake_estimate)monkeypatch.setattr(S, "free_bytes", fake_free)monkeypatch.setattr(S, "kaggle_tmp_redirect", fake_redirect)+ # Charge the cache with the export, wherever the host keeps it (a Windows runner: D: vs C:).+ monkeypatch.setattr(S, "_hub_cache_directory", lambda: None)monkeypatch.delenv("UNSLOTH_DISK_PREFLIGHT", raising = False)monkeypatch.delenv("UNSLOTH_PREWARM_HUB_CACHE", raising = False)@@ -2081,6 +2083,8 @@ class TestADisposableMergeIsNotChargedForAllThreeAtOnce:monkeypatch.setattr(S, "PeftModel", _FakeAdapterModel)monkeypatch.delenv("UNSLOTH_DISK_PREFLIGHT", raising = False)monkeypatch.delenv("UNSLOTH_PREWARM_HUB_CACHE", raising = False)+ # Without bf16 the preflight lowers a bf16 first conversion to f16, which adds a quantize pass.+ monkeypatch.setattr(S.torch.cuda, "is_bf16_supported", lambda *a, **k: True)state["directory"] = str(tmp_path / "model")return state