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: allow the zoo's padded logits row stride in the fused linear adapter check (#15803)
Daniel Han2d4ac1cb2894
diff --git a/tests/python/test_fused_linear_trl_routes.py b/tests/python/test_fused_linear_trl_routes.pyindex 4c0bd33bde9..1d708905e72 100644--- a/tests/python/test_fused_linear_trl_routes.py+++ b/tests/python/test_fused_linear_trl_routes.py@@ -455,7 +455,10 @@ def test_adapter_bounds_its_logits_buffer_by_chunk_size(trainable_head, grad, mosizes.clear()with torch.set_grad_enabled(grad):adapter.apply(hidden, weight, None, targets, 1.0, 32768, None, 1.0, ("entropy",))- assert kwargs[-1]["max_rows_per_chunk"] is None and max(sizes) == n * v * 4+ # One [N, V] buffer; zoo may pad its row stride (unslothai/unsloth-zoo#2190 pads to 16 for Triton).+ pad = getattr(fused_linear, "_FWD_PAD", 1)+ assert kwargs[-1]["max_rows_per_chunk"] is None+ assert max(sizes) == n * (-(-v // pad) * pad) * 4, (sizes, n, v, pad)@pytest.mark.parametrize("grad", [False, True])