From fdc1260e99191717b0aa0a48117d4b758a24a513 Mon Sep 17 00:00:00 2001
From: Victor-Loos <loosvictor0@gmail.com>
Date: Sat, 18 Jul 2026 14:47:23 +0200
Subject: [PATCH 1/2] ggml-cuda: avoid ROCm_Host compute on HIP integrated GPUs

---
 ggml/src/ggml-cuda/ggml-cuda.cu | 25 +++++++++++++++++++++++--
 1 file changed, 23 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index 4aaa4f367afc..e868a54a1d95 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -4661,6 +4661,20 @@ static enum ggml_backend_dev_type ggml_backend_cuda_device_get_type(ggml_backend
         : GGML_BACKEND_DEVICE_TYPE_GPU;
 }
 
+static bool ggml_backend_cuda_host_buffer_supported() {
+    return getenv("GGML_CUDA_NO_PINNED") == nullptr;
+}
+
+static bool ggml_backend_cuda_device_supports_cuda_host_buft(int device) {
+#if defined(GGML_USE_HIP)
+    if (ggml_cuda_info().devices[device].integrated) {
+        return false;
+    }
+#endif
+
+    return ggml_backend_cuda_host_buffer_supported();
+}
+
 static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_backend_dev_props * props) {
     ggml_backend_cuda_device_context * ctx = (ggml_backend_cuda_device_context *)dev->context;
 
@@ -4670,7 +4684,7 @@ static void ggml_backend_cuda_device_get_props(ggml_backend_dev_t dev, ggml_back
     props->device_id   = ctx->pci_bus_id.empty() ? nullptr : ctx->pci_bus_id.c_str();
     ggml_backend_cuda_device_get_memory(dev, &props->memory_free, &props->memory_total);
 
-    bool host_buffer = getenv("GGML_CUDA_NO_PINNED") == nullptr;
+    bool host_buffer = ggml_backend_cuda_host_buffer_supported();
 #ifdef GGML_CUDA_NO_PEER_COPY
     bool events = false;
 #else
@@ -4698,6 +4712,10 @@ static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_buffer_type(ggml_
 
 static ggml_backend_buffer_type_t ggml_backend_cuda_device_get_host_buffer_type(ggml_backend_dev_t dev) {
     GGML_UNUSED(dev);
+    if (!ggml_backend_cuda_host_buffer_supported()) {
+        return nullptr;
+    }
+
     return ggml_backend_cuda_host_buffer_type();
 }
 
@@ -5131,7 +5149,10 @@ static bool ggml_backend_cuda_device_supports_op(ggml_backend_dev_t dev, const g
 static bool ggml_backend_cuda_device_supports_buft(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft) {
     ggml_backend_cuda_device_context * dev_ctx = (ggml_backend_cuda_device_context *) dev->context;
     const bool integrated = ggml_cuda_info().devices[dev_ctx->device].integrated;
-    return (ggml_backend_buft_is_cuda(buft) && buft->device == dev) || (integrated && ggml_backend_buft_is_cuda_host(buft));
+    return (ggml_backend_buft_is_cuda(buft) && buft->device == dev) ||
+        (integrated &&
+         ggml_backend_buft_is_cuda_host(buft) &&
+         ggml_backend_cuda_device_supports_cuda_host_buft(dev_ctx->device));
 }
 
 static int64_t get_op_batch_size(const ggml_tensor * op) {

From ce82541acbaf5c532c0727d6ccb6de2b0b0c948d Mon Sep 17 00:00:00 2001
From: Victor-Loos <loosvictor0@gmail.com>
Date: Tue, 21 Jul 2026 19:09:27 +0200
Subject: [PATCH 2/2] ggml-cuda: fix unused parameter in non-HIP builds

---
 ggml/src/ggml-cuda/ggml-cuda.cu | 2 ++
 1 file changed, 2 insertions(+)
 mode change 100644 => 100755 ggml/src/ggml-cuda/ggml-cuda.cu

diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
old mode 100644
new mode 100755
index e868a54a1d95..83eb1a42d443
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -4670,6 +4670,8 @@ static bool ggml_backend_cuda_device_supports_cuda_host_buft(int device) {
     if (ggml_cuda_info().devices[device].integrated) {
         return false;
     }
+#else
+    GGML_UNUSED(device);
 #endif
 
     return ggml_backend_cuda_host_buffer_supported();