From f38de163419d9381464b9185bca55c54207b3e43 Mon Sep 17 00:00:00 2001 From: Georgi Gerganov Date: Fri, 2 Jan 2026 19:02:18 +0200 Subject: [PATCH] metal : adjust extra size for FA buffer to avoid reallocations (#18545) --- ggml/src/ggml-metal/ggml-metal-ops.cpp | 6 +++++- 1 file changed, 5 insertions(+), 1 deletion(-) diff --git a/ggml/src/ggml-metal/ggml-metal-ops.cpp b/ggml/src/ggml-metal/ggml-metal-ops.cpp index acf2aa9184..a50b12b6f3 100644 --- a/ggml/src/ggml-metal/ggml-metal-ops.cpp +++ b/ggml/src/ggml-metal/ggml-metal-ops.cpp @@ -2181,7 +2181,11 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const ggml_tensor * op) { const bool has_mask = op->src[3] != nullptr; - if (ggml_metal_op_flash_attn_ext_use_vec(op)) { + // note: the non-vec kernel requires more extra memory, so always reserve for it + GGML_ASSERT(OP_FLASH_ATTN_EXT_NCPSG >= OP_FLASH_ATTN_EXT_VEC_NCPSG); + + //if (ggml_metal_op_flash_attn_ext_use_vec(op)) { + if (false) { // note: always reserve the padding space to avoid graph reallocations //const bool has_kvpad = ne11 % OP_FLASH_ATTN_EXT_VEC_NCPSG != 0; const bool has_kvpad = true;