From: Diego Devesa <redacted>
Date: Tue, 7 Jan 2025 11:38:05 +0000 (+0100)
Subject: ggml-backend : only offload from host buffers (llama/11120)
X-Git-Tag: upstream/1.7.4+33~17
X-Git-Url: https://git.djapps.eu/?a=commitdiff_plain;h=3988d6396b7b10e015467a64f9bcf549b6b41cb5;p=pkg%2Fggml%2Fsources%2Fwhisper.cpp

ggml-backend : only offload from host buffers (llama/11120)
---

diff --git a/ggml/src/ggml-backend.cpp b/ggml/src/ggml-backend.cpp
index e2d6c405..d034f8b7 100644
--- a/ggml/src/ggml-backend.cpp
+++ b/ggml/src/ggml-backend.cpp
@@ -761,7 +761,7 @@ 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) {
+        if (tensor->op != GGML_OP_ROPE && src->buffer != NULL && src->buffer->usage == GGML_BACKEND_BUFFER_USAGE_WEIGHTS && ggml_backend_buffer_is_host(src->buffer)) {
             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) {