diff --git a/example/ck_tile/02_layernorm2d/generate.py b/example/ck_tile/02_layernorm2d/generate.py index c4366f6662..ea77f16488 100644 --- a/example/ck_tile/02_layernorm2d/generate.py +++ b/example/ck_tile/02_layernorm2d/generate.py @@ -552,7 +552,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t, h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 256, 4, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 6, 1, 256, 2, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 1, True, False, True, True, False, 0, 0, 0)], - '4096' :[ h_traits('x', 'y', 'xs', 'ys', 1, 2, 1, 256, 8, True, False, True, True, False, 0, 0, 0), + '4096' :[ h_traits('x', 'y', 'xs', 'ys', 2, 2, 1, 256, 8, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 4, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 2, 1,1024, 2, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 1, True, False, True, True, False, 0, 0, 0)], @@ -560,7 +560,7 @@ float layernorm2d_fwd(layernorm2d_fwd_traits t, h_traits('x', 'y', 'xs', 'ys', 1, 3, 1, 512, 4, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 3, 1,1024, 2, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 6, 1,1024, 1, True, False, True, True, False, 0, 0, 0)], - '8192' :[ h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 256, 8, True, False, True, True, False, 0, 0, 0), + '8192' :[ h_traits('x', 'y', 'xs', 'ys', 2, 4, 1, 256, 8, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 4, 1, 512, 4, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 4, 1,1024, 2, True, False, True, True, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 1, 8, 1,1024, 1, True, False, True, True, False, 0, 0, 0)], diff --git a/example/ck_tile/10_rmsnorm2d/generate.py b/example/ck_tile/10_rmsnorm2d/generate.py index ea8dfdf9ce..3b431197b3 100644 --- a/example/ck_tile/10_rmsnorm2d/generate.py +++ b/example/ck_tile/10_rmsnorm2d/generate.py @@ -578,7 +578,7 @@ float rmsnorm2d_fwd(rmsnorm2d_fwd_traits t, h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 3, 1, 512, 4, True, False, False, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 3, 1,1024, 2, True, False, False, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 6, 1,1024, 1, True, False, False, False, 0, 0, 0)], - '8192' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 4, 1, 256, 8, True, False, False, False, 0, 0, 0), + '8192' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 2, 16, 1, 256, 2, True, False, False, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 4, 1, 512, 4, True, False, False, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 4, 1,1024, 2, True, False, False, False, 0, 0, 0), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 8, 1,1024, 1, True, False, False, False, 0, 0, 0)], @@ -634,7 +634,7 @@ float rmsnorm2d_fwd(rmsnorm2d_fwd_traits t, h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 3, 1, 512, 4, True, False, False, False, 0, 0, 1), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 3, 1,1024, 2, True, False, False, False, 0, 0, 1), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 6, 1,1024, 1, True, False, False, False, 0, 0, 1)], - '8192' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 4, 1, 256, 8, True, False, False, False, 0, 0, 1), + '8192' :[ h_traits('x', 'y', 'xs', 'ys', 'uqy', 2, 16, 1, 256, 2, True, False, False, False, 0, 0, 1), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 4, 1, 512, 4, True, False, False, False, 0, 0, 1), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 4, 1,1024, 2, True, False, False, False, 0, 0, 1), h_traits('x', 'y', 'xs', 'ys', 'uqy', 1, 8, 1,1024, 1, True, False, False, False, 0, 0, 1)],