fix embeddings when using CUDA (#3657)

diff --git a/llama.cpp b/llama.cpp
index 37df88779ab35f57c9da4a9588917a21ae4930f4..04a779e041ac100c5a3ee25d84a4ae90788155c3 100644 (file)
--- a/llama.cpp
+++ b/llama.cpp
@@ -5903,6 +5903,13 @@ static int llama_decode_internal(
 
     ggml_allocr_alloc_graph(lctx.alloc, gf);
 
+    struct ggml_tensor * res        = gf->nodes[gf->n_nodes - 1];
+    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
+
+    GGML_ASSERT(strcmp(res->name,        "result_output") == 0);
+    GGML_ASSERT(strcmp(embeddings->name, "result_norm")   == 0);
+
+
 #ifdef GGML_USE_CUBLAS
     for (int i = 0; i < gf->n_leafs; i++) {
         ggml_tensor * node = gf->leafs[i];
@@ -5920,6 +5927,12 @@ static int llama_decode_internal(
     }
 
     ggml_cuda_set_mul_mat_q(cparams.mul_mat_q);
+
+    // HACK: ggml-alloc may change the tensor backend when reusing a parent, so force output to be on the CPU here if needed
+    if (!lctx.embedding.empty()) {
+        embeddings->backend = GGML_BACKEND_CPU;
+    }
+    res->backend = GGML_BACKEND_CPU;
 #endif
 
     // LLAMA_LOG_INFO("graph build time: %.3f ms (%d nodes, %d leafs)\n", (ggml_time_us() - t_start_us)/1000.0, gf->n_nodes, gf->n_leafs);
@@ -5944,12 +5957,6 @@ static int llama_decode_internal(
         n_threads = 1;
     }
 
-    struct ggml_tensor * res        = gf->nodes[gf->n_nodes - 1];
-    struct ggml_tensor * embeddings = gf->nodes[gf->n_nodes - 2];
-
-    GGML_ASSERT(strcmp(res->name,        "result_output") == 0);
-    GGML_ASSERT(strcmp(embeddings->name, "result_norm")   == 0);
-
 #if GGML_USE_MPI
     const int64_t n_layer = hparams.n_layer;
     ggml_mpi_graph_compute_pre(lctx.ctx_mpi, gf, n_layer);