From 1c1484d7f8ff33a1ec1dd62d8ce0c1e1537c57ee Mon Sep 17 00:00:00 2001 From: Martin Jurkovic Date: Thu, 15 Oct 2026 14:23:26 +0300 Subject: [PATCH] Make returned attention key/value projections contiguous Signed-off-by: Martin Jurkovic --- sdm/nn/attention.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/sdm/nn/attention.py b/sdm/nn/attention.py index 7abd49ad22..7048817559 100644 --- a/sdm/nn/attention.py +++ b/sdm/nn/attention.py @@ -383,8 +383,8 @@ def forward( # CUDA autocast runs RMSNorm in fp32, so cast explicitly before # caching: https://github.com/pytorch/pytorch/blob/v2.13.0/aten/src/ATen/autocast_mode.h#L875 return out, KVCacheEntry( - key=key.to(value.dtype), - value=value, + key=key.to(value.dtype).contiguous(), + value=value.contiguous(), ) return out