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
Block compile: keep a caller's flex BACKEND instead of adding FORCE_USE_FLEX_ATTENTION (#15044)
Daniel Hand3e3edc1e9c2
diff --git a/tests/test_block_compile.py b/tests/test_block_compile.pyindex ec1361a8466..9110d7ee8e0 100644--- a/tests/test_block_compile.py+++ b/tests/test_block_compile.py@@ -1463,3 +1463,23 @@ def test_router_logits_capture_compiles_one_graph_for_every_layer(monkeypatch):assert len(got) == len(want)for g, w in zip(got, want):torch.testing.assert_close(g.detach(), w)++[email protected]("backend", ["AUTO", "TRITON"])+def test_training_flex_options_keep_a_callers_backend(backend):+ """BACKEND supersedes FORCE_USE_FLEX_ATTENTION, and torch 2.10+ raises when both are set."""+ M = _load()+ options = {"BACKEND": backend}+ assert M._training_flex_options(options) == options+ assert M._training_flex_options({"ROWS_GUARANTEED_SAFE": True}) == {+ "ROWS_GUARANTEED_SAFE": True,+ "FORCE_USE_FLEX_ATTENTION": True,+ }+ flex = pytest.importorskip("torch.nn.attention.flex_attention")+ if (+ not hasattr(flex, "_apply_kernel_options")+ or "BACKEND" not in getattr(flex, "FlexKernelOptions", {}).__annotations__+ ):+ return+ q = torch.randn(1, 1, 16, 16)+ flex._apply_kernel_options(q, q, q, False, M._training_flex_options(options))diff --git a/unsloth/models/_block_compile.py b/unsloth/models/_block_compile.pyindex 8047ab0229e..84bcb34720f 100644--- a/unsloth/models/_block_compile.py+++ b/unsloth/models/_block_compile.py@@ -573,7 +573,8 @@ def _training_flex_options(options):(GQA, head_dim 256), demoting every block. A caller's explicit choice is kept."""if options is None:return {"FORCE_USE_FLEX_ATTENTION": True}- if isinstance(options, dict) and "FORCE_USE_FLEX_ATTENTION" not in options:+ # torch 2.10+ rejects BACKEND together with the legacy FORCE_USE_FLEX_ATTENTION.+ if isinstance(options, dict) and not ({"FORCE_USE_FLEX_ATTENTION", "BACKEND"} & options.keys()):return {**options, "FORCE_USE_FLEX_ATTENTION": True}return options