From 3e894f75872534b907b19175b6d2f3cdde1490a6 Mon Sep 17 00:00:00 2001 From: Gerald Corzo Perez Date: Fri, 21 Aug 2026 01:04:01 +0200 Subject: [PATCH] =?UTF-8?q?fix:=20gate=20K<4=20crash=20=E2=80=94=20views/a?= =?UTF-8?q?ggregation=20loops=20must=20use=20gated=20n=5Fexpert=5Fused,=20?= =?UTF-8?q?not=20hparams=20value=20(ggml=5Fview=5F2d=20out-of-bounds=20ass?= =?UTF-8?q?ert)?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- src/llama-graph.cpp | 13 +++++++------ 1 file changed, 7 insertions(+), 6 deletions(-) diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp index 253998c59..4c1161110 100644 --- a/src/llama-graph.cpp +++ b/src/llama-graph.cpp @@ -2263,25 +2263,26 @@ ggml_tensor * llm_graph_context::build_moe_ffn( assert(n_expert_used > 0); // order the views before the adds - for (uint32_t i = 0; i < hparams.n_expert_used; ++i) { + // REMORA: use the (possibly gated) n_expert_used so the views match the + // gated `experts` shape (native hparams.n_expert_used would exceed it) + for (uint32_t i = 0; i < (uint32_t) n_expert_used; ++i) { cur_experts[i] = ggml_view_2d(ctx0, experts, n_embd, n_tokens, experts->nb[2], i*experts->nb[1]); ggml_build_forward_expand(gf, cur_experts[i]); } // aggregate experts - // note: here we explicitly use hparams.n_expert_used instead of n_expert_used - // to avoid potentially a large number of add nodes during warmup - // ref: https://github.com/ggml-org/llama.cpp/pull/14753 + // note: the gate (REMORA_GATE_K) may have reduced n_expert_used below the + // hparams value, so the aggregation must iterate the gated count ggml_tensor * moe_out = cur_experts[0]; - for (uint32_t i = 1; i < hparams.n_expert_used; ++i) { + for (uint32_t i = 1; i < (uint32_t) n_expert_used; ++i) { moe_out = ggml_add(ctx0, moe_out, cur_experts[i]); ggml_build_forward_expand(gf, moe_out); } - if (hparams.n_expert_used == 1) { + if (n_expert_used == 1) { // avoid returning a non-contiguous tensor moe_out = ggml_cont(ctx0, moe_out); } -- 2.50.1 (Apple Git-155)