diff --git a/ggml/src/ggml-cuda/ggml-cuda.cu b/ggml/src/ggml-cuda/ggml-cuda.cu
index e76ff3128fdd..4baf49e1fbc3 100644
--- a/ggml/src/ggml-cuda/ggml-cuda.cu
+++ b/ggml/src/ggml-cuda/ggml-cuda.cu
@@ -305,7 +305,11 @@ static ggml_cuda_device_info ggml_cuda_init() {
 
         info.default_tensor_split[id] = total_vram;
         total_vram += device_vram;
+#if defined(GGML_USE_HIP)
+        info.devices[id].integrated = prop.integrated;
+#else
         info.devices[id].integrated = false; // Temporarily disabled due to issues with corrupted output (e.g. #15034)
+#endif
         info.devices[id].nsm        = prop.multiProcessorCount;
         info.devices[id].smpb       = prop.sharedMemPerBlock;
         info.devices[id].warp_size  = prop.warpSize;
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 8675f6087336..aea52c724e38 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -424,6 +424,25 @@ llama_context::llama_context(
 
         LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
 
+        // graph inputs live in the CPU backend's buffer. if a GPU can use that buffer directly, the scheduler
+        // does not copy the inputs to the device, so nothing orders the next set_inputs after the previous compute
+        for (size_t i = 0; i < backend_ptrs.size(); ++i) {
+            if (ggml_backend_dev_type(ggml_backend_get_device(backend_ptrs[i])) != GGML_BACKEND_DEVICE_TYPE_CPU) {
+                continue;
+            }
+            for (auto * backend : backend_ptrs) {
+                auto * dev = ggml_backend_get_device(backend);
+                auto dev_type = ggml_backend_dev_type(dev);
+                if ((dev_type == GGML_BACKEND_DEVICE_TYPE_GPU || dev_type == GGML_BACKEND_DEVICE_TYPE_IGPU) &&
+                    ggml_backend_dev_supports_buft(dev, backend_buft[i])) {
+                    sync_host_inputs = true;
+                }
+            }
+        }
+        if (sync_host_inputs) {
+            LLAMA_LOG_INFO("%s: GPU reads host input buffer in place, synchronizing before set_inputs\n", __func__);
+        }
+
         // TODO: move these checks to ggml_backend_sched
         // enabling pipeline parallelism in the scheduler increases memory usage, so it is only done when necessary
         bool pipeline_parallel =
@@ -1445,6 +1464,11 @@ llm_graph_result * llama_context::process_ubatch(const llama_ubatch & ubatch, ll
     {
         //const auto t_start_us = ggml_time_us();
 
+        // the previous graph_compute_async may still be reading the inputs that are about to be overwritten
+        if (sync_host_inputs) {
+            ggml_backend_sched_synchronize(sched.get());
+        }
+
         // FIXME this call causes a crash if any model inputs were not used in the graph and were therefore not allocated
         res->set_inputs(&ubatch);
 
diff --git a/src/llama-context.h b/src/llama-context.h
index b403b099b76f..c09ec7c77258 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -389,6 +389,10 @@ struct llama_context {
     // env: LLAMA_GRAPH_REUSE_DISABLE
     bool graph_reuse_disable = false;
 
+    // true if a GPU backend reads the CPU backend's (host) buffer in place, e.g. integrated GPUs
+    // in that case set_inputs must not overwrite graph inputs while a previous compute is in flight
+    bool sync_host_inputs = false;
+
     // perf
     mutable int64_t t_start_us  = 0;
     mutable int64_t t_load_us   = 0;