From 9beb6aca6995989fb31bae18e725443e83ede12a Mon Sep 17 00:00:00 2001
From: "Piotr Wilkin (ilintar)" <ilintar@gmail.com>
Date: Fri, 10 Jul 2026 17:22:15 +0200
Subject: [PATCH] fix: avoid f16 overflow in Z-Image quantized matmuls on ROCm
 (#1771)

---
 src/model/diffusion/z_image.hpp | 4 +++-
 1 file changed, 3 insertions(+), 1 deletion(-)

diff --git a/src/model/diffusion/z_image.hpp b/src/model/diffusion/z_image.hpp
index 362192f1..9fb333cb 100644
--- a/src/model/diffusion/z_image.hpp
+++ b/src/model/diffusion/z_image.hpp
@@ -150,6 +150,8 @@ namespace ZImage {
 
             if (sd_backend_is(ctx->backend, "ROCm")) {
                 out_proj->set_scale(1.f / 16.f);
+                out_proj->set_force_prec_f32(true);
+                qkv_proj->set_force_prec_f32(true);
             }
 
             auto qkv = qkv_proj->forward(ctx, x);                                                                            // [N, n_token, (num_heads + num_kv_heads*2)*head_dim]
@@ -227,7 +229,7 @@ namespace ZImage {
             auto w2 = std::dynamic_pointer_cast<Linear>(blocks["w2"]);
             auto w3 = std::dynamic_pointer_cast<Linear>(blocks["w3"]);
 
-            if (sd_backend_is(ctx->backend, "Vulkan")) {
+            if (sd_backend_is(ctx->backend, "Vulkan") || sd_backend_is(ctx->backend, "ROCm")) {
                 w2->set_force_prec_f32(true);
             }