From eea6a82b25fcf16f9409c55f114c56935bf9b907 Mon Sep 17 00:00:00 2001
From: Kawrakow <iwankawrakow@gmail.com>
Date: Mon, 8 Jun 2026 16:51:05 +0200
Subject: [PATCH] Fix bf16 graph reduce type (#1938)

---
 ggml/src/ggml-cuda/binbcast.cu | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/ggml/src/ggml-cuda/binbcast.cu b/ggml/src/ggml-cuda/binbcast.cu
index 5953746b..898640dd 100644
--- a/ggml/src/ggml-cuda/binbcast.cu
+++ b/ggml/src/ggml-cuda/binbcast.cu
@@ -476,6 +476,18 @@ void ggml_cuda_op_add(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
             else if (dst->src[0]->type == GGML_TYPE_F32 && dst->src[1]->type == GGML_TYPE_F32) {
                 k_fast_add_2<<<nblocks, kBlockSize, 0, ctx.stream()>>>(dst->ne[0], nelem,
                         (const float *)dst->src[0]->data, (const float *)dst->src[1]->data, (float *)dst->data);
+            }
+            else if (dst->src[0]->type == GGML_TYPE_BF16 && dst->src[1]->type == GGML_TYPE_BF16) {
+                k_fast_add_2<<<nblocks, kBlockSize, 0, ctx.stream()>>>(dst->ne[0], nelem,
+                        (const nv_bfloat16 *)dst->src[0]->data, (const nv_bfloat16 *)dst->src[1]->data, (float *)dst->data);
+            }
+            else if (dst->src[0]->type == GGML_TYPE_BF16 && dst->src[1]->type == GGML_TYPE_F32) {
+                k_fast_add_2<<<nblocks, kBlockSize, 0, ctx.stream()>>>(dst->ne[0], nelem,
+                        (const nv_bfloat16 *)dst->src[0]->data, (const float *)dst->src[1]->data, (float *)dst->data);
+            }
+            else if (dst->src[0]->type == GGML_TYPE_F32 && dst->src[1]->type == GGML_TYPE_BF16) {
+                k_fast_add_2<<<nblocks, kBlockSize, 0, ctx.stream()>>>(dst->ne[0], nelem,
+                        (const float *)dst->src[0]->data, (const nv_bfloat16 *)dst->src[1]->data, (float *)dst->data);
             } else {
                 k_fast_add_2<<<nblocks, kBlockSize, 0, ctx.stream()>>>(dst->ne[0], nelem,
                         (const float *)dst->src[0]->data, (const half *)dst->src[1]->data, (float *)dst->data);