`torch.empty` can create issues; use `torch.zeros` (24892520) · Commits · github_fork / Stable Diffusion Webui

For MPS, using a tensor created with `torch.empty()` can cause `torch.baddbmm()` to include NaNs in the tensor it returns, even though `beta=0`. However, with a tensor of shape [1,1,1], there should be a negligible performance difference between `torch.empty()` and `torch.zeros()` anyway, so it's better to just use `torch.zeros()` for this and avoid unnecessarily creating issues.

modules/sub_quadratic_attention.py

+2 −2

Original line number	Diff line number	Diff line
		@@ -58,7 +58,7 @@ def _summarize_chunk(
		scale: float,
		) -> AttnChunk:
		attn_weights = torch.baddbmm(
		torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
		torch.zeros(1, 1, 1, device=query.device, dtype=query.dtype),
		query,
		key.transpose(1,2),
		alpha=scale,
		@@ -121,7 +121,7 @@ def _get_attention_scores_no_kv_chunking(
		scale: float,
		) -> Tensor:
		attn_scores = torch.baddbmm(
		torch.empty(1, 1, 1, device=query.device, dtype=query.dtype),
		torch.zeros(1, 1, 1, device=query.device, dtype=query.dtype),
		query,
		key.transpose(1,2),
		alpha=scale,

Admin message