ggml-backend : only offload from host buffers (fix) (llama/11124)

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index d034f8b7ffb9d0ae6d3b4555345c9b2f5b6c9520..dba7be33b88c07e3b63b624bd3b20f8bcb2f7dbd 100644 (file)
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -761,10 +761,10 @@ static int ggml_backend_sched_backend_id_from_cur(ggml_backend_sched_t sched, st
         }
         // skip ROPE since the rope freqs tensor is too small to choose a backend based on it
         // not an ideal solution
-        if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS && ggml_backend_buffer_is_host(src->buffer)) {
+        if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS) {
             int src_backend_id = ggml_backend_sched_backend_from_buffer(sched, src, tensor);
             // check if a backend with higher prio wants to offload the op
-            if (src_backend_id == sched->n_backends - 1) {
+            if (src_backend_id == sched->n_backends - 1 && ggml_backend_buffer_is_host(src->buffer)) {
                 for (int b = 0; b < src_backend_id; b++) {
                     if (ggml_backend_supports_op(sched->backends[b], tensor) && ggml_backend_offload_op(sched->backends[b], tensor)) {
                         SET_CAUSE(tensor, "1.off");

diff --git a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
index 622c63f1f8ebfbdb87ad229e499bcc560d4a44f0..b311a5b1c4b73f810ca4c9a239ec2daa4c6da6ab 100644 (file)
--- a/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
+++ b/ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
@@ -4169,6 +4169,8 @@ static ggml_backend_buffer_t ggml_backend_cpu_aarch64_buffer_type_alloc_buffer(g
     buffer->buft              = buft;
     buffer->iface.init_tensor = ggml_backend_cpu_aarch64_buffer_init_tensor;
     buffer->iface.set_tensor  = ggml_backend_cpu_aarch64_buffer_set_tensor;
+    buffer->iface.get_tensor  = nullptr;
+    buffer->iface.cpy_tensor  = nullptr;
     return buffer;
 }