amd multigpu full layer offload w/o vram scratch

LostRuins · Jul 24, 2023 · 5b838d4 · 5b838d4
1 parent 9bfb2fd
commit 5b838d4
Showing 1 changed file with 14 additions and 0 deletions.
diff --git a/llama.cpp b/llama.cpp
@@ -1274,18 +1274,32 @@ static void llama_model_load_internal(
 
 #ifdef GGML_USE_CUBLAS
         const int max_backend_supported_layers = hparams.n_layer + 3;
+#if defined(GGML_USE_HIPBLAS)
+        const int max_offloadable_layers = low_vram ? hparams.n_layer + 3 : hparams.n_layer + 3;
+#else
         const int max_offloadable_layers = low_vram ? hparams.n_layer + 1 : hparams.n_layer + 3;
+#endif
         if (n_gpu_layers > (int) hparams.n_layer + 1) {
             if (low_vram) {
+#if defined(GGML_USE_HIPBLAS)
+                fprintf(stderr, "%s: offloading v cache to GPU\n", __func__);
+                vram_kv_cache += hparams.kv_size() / 2;
+#else
                 fprintf(stderr, "%s: cannot offload v cache to GPU due to low VRAM option\n", __func__);
+#endif
             } else {
                 fprintf(stderr, "%s: offloading v cache to GPU\n", __func__);
                 vram_kv_cache += hparams.kv_size() / 2;
             }
         }
         if (n_gpu_layers > (int) hparams.n_layer + 2) {
             if (low_vram) {
+#if defined(GGML_USE_HIPBLAS)
+                fprintf(stderr, "%s: offloading k cache to GPU\n", __func__);
+                vram_kv_cache += hparams.kv_size() / 2;
+#else
                 fprintf(stderr, "%s: cannot offload k cache to GPU due to low VRAM option\n", __func__);
+#endif
             } else {
                 fprintf(stderr, "%s: offloading k cache to GPU\n", __func__);
                 vram_kv_cache += hparams.kv_size() / 2;