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.

FP8 e4m3fn scale dequant: pass enable_fp_fusion like every other launch (#14969)

Daniel Hand95472002820
diff --git a/unsloth/kernels/fp8.py b/unsloth/kernels/fp8.py
index 5d52a15f13a..31f7b9fee6f 100644
--- a/unsloth/kernels/fp8.py
+++ b/unsloth/kernels/fp8.py
@@ -253,6 +253,7 @@ def _scale_dequant_one_pass(x: torch.Tensor, s: torch.Tensor, dtype: torch.dtype
ROUND_SCALE = dtype != torch.float16,
LONG_INDEXING = long_indexing(xb, out, block = BLOCK_M * n),
num_warps = num_warps,
+ enable_fp_fusion = True,
)
return out.t() if transposed else out