From 67ad7616e07eb9e52a90a4a785c57f9e05929db1 Mon Sep 17 00:00:00 2001
From: LeiWang1999 <leiwang1999@outlook.com>
Date: Mon, 19 Aug 2024 12:52:34 +0000
Subject: [PATCH] fix scale only lop3 tensorize instructions.

---
 bitblas/gpu/intrin/lop3.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/bitblas/gpu/intrin/lop3.py b/bitblas/gpu/intrin/lop3.py
index a6a7011a0..aee3eac8b 100644
--- a/bitblas/gpu/intrin/lop3.py
+++ b/bitblas/gpu/intrin/lop3.py
@@ -1127,7 +1127,7 @@ def fast_decode_desc(compressed: T.handle, decompressed: T.handle, scale: T.hand
                     1,
                 ],
                 dtype=target_dtype,
-                scope="global",
+                scope="local",
             )
             with T.block("root"):
                 T.reads(Compressed[0:n_storage_elems], Scale[0:1])
@@ -1173,7 +1173,7 @@ def fast_decode_impl(compressed: T.handle, decompressed: T.handle, scale: T.hand
                 dtype=target_dtype,
                 offset_factor=1,
                 strides=[s0],
-                scope="global",
+                scope="local",
             )
             with T.block("root"):
                 T.reads(Compressed[0:n_storage_elems], Scale[0:1])