From e7aa89a48c9ea815a90699b59baa863b92b6e9e5 Mon Sep 17 00:00:00 2001 From: =?UTF-8?q?Stanis=C5=82aw=20Szymczyk?= Date: Sun, 22 Mar 2026 13:42:19 +0100 Subject: [PATCH] model : replaced ggml_argsort_top_k with ggml_top_k in DeepSeek V3.2 indexer implementation since the former fails for large tensors even when using CCCL. --- src/models/deepseek32.cpp | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/src/models/deepseek32.cpp b/src/models/deepseek32.cpp index 23bb45c534..372eace710 100644 --- a/src/models/deepseek32.cpp +++ b/src/models/deepseek32.cpp @@ -185,7 +185,7 @@ llm_build_deepseek32::llm_build_deepseek32(const llama_model & model, const llm_ cb(indexer_score, "indexer_score", il); uint32_t n_top_k = indexer_score->ne[0] < n_indexer_top_k ? indexer_score->ne[0] : n_indexer_top_k; - ggml_tensor * top_k = ggml_cont(ctx0, ggml_argsort_top_k(ctx0, indexer_score, n_top_k)); + ggml_tensor * top_k = ggml_cont(ctx0, ggml_top_k(ctx0, indexer_score, n_top_k)); cb(top_k, "top_k", il); // prepare new kq mask - starts filled with -INFINITY