]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commitdiff
llama-fit-params: fix step size for last device (#18415)
authorJohannes Gäßler <redacted>
Sun, 28 Dec 2025 09:52:09 +0000 (10:52 +0100)
committerGitHub <redacted>
Sun, 28 Dec 2025 09:52:09 +0000 (10:52 +0100)
src/llama.cpp

index 93a9c408ba4790712ac0e6ee27e8835e2dae3b36..76b3acbadb62906dc42474f380ee2b70b2bd99ac 100644 (file)
@@ -512,6 +512,9 @@ static void llama_params_fit_impl(
             if (mem_high[id] > targets[id]) {
                 assert(ngl_per_device_high[id].n_layer > ngl_per_device[id].n_layer);
                 uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
+                if (hp_nex > 0 && size_t(id) == nd - 1) {
+                    delta--;
+                }
                 LLAMA_LOG_DEBUG("%s: start filling device %" PRIu32 ", delta=%" PRIu32 "\n", __func__, id, delta);
                 while (delta > 1) {
                     uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);