From dacfa01c0dd0d3e86ea058a5bcbe641869c77c6c Mon Sep 17 00:00:00 2001 From: ttomsa Date: Mon, 23 Feb 2026 01:01:57 +0000 Subject: [PATCH] enable float16 and unaligned vector load/store --- tinygrad/codegen/late/devectorizer.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/tinygrad/codegen/late/devectorizer.py b/tinygrad/codegen/late/devectorizer.py index c0812dcfe1..c4e79584f4 100644 --- a/tinygrad/codegen/late/devectorizer.py +++ b/tinygrad/codegen/late/devectorizer.py @@ -157,8 +157,8 @@ def split_load_store(ctx:Renderer|None, ls:UOp, idx:UOp): elif isinstance(buf.dtype, ImageDType): lengths = [4] elif ctx is not None and CPU_X86: - lengths = [4,2] if buf.dtype.base == dtypes.float32 else [] - #must_divide = False + lengths = [8,4,2] if buf.dtype.base == dtypes.half else [4,2] + must_divide = False elif ctx is not None and ctx.supports_float4: # TODO: a better way to get this than ctx lengths = [8,4,2] if buf.dtype.base == dtypes.half and getenv("ALLOW_HALF8") else ([16,8,4,2] if AMX else [4,2])