From 000fded1ea6cbfa14c2aaefe3c35a903b78f5a3a Mon Sep 17 00:00:00 2001 From: Yee Man Chan Date: Tue, 3 Feb 2026 18:42:17 +0800 Subject: [PATCH] add back ggml_cont for Vcur --- src/models/kimi-linear.cpp | 1 + 1 file changed, 1 insertion(+) diff --git a/src/models/kimi-linear.cpp b/src/models/kimi-linear.cpp index 1e533fa51b..0f037d1a39 100644 --- a/src/models/kimi-linear.cpp +++ b/src/models/kimi-linear.cpp @@ -291,6 +291,7 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll ggml_row_size(kv->type, kv_per_head), ggml_row_size(kv->type, kv_per_head * n_head), ggml_row_size(kv->type, n_embd_head_qk_nope)); + Vcur = ggml_cont(ctx0, Vcur); cb(Vcur, "mla_V", il); // Concatenate k_nope + k_pe (broadcast k_pe to all heads)