From 19064083bd3696682dbf886e23160f103137613e Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Tue, 23 Jun 2026 22:34:34 +0800
Subject: [PATCH] another fix for drafting

---
 gpttype_adapter.cpp | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/gpttype_adapter.cpp b/gpttype_adapter.cpp
index 245718c15..a355eddd7 100644
--- a/gpttype_adapter.cpp
+++ b/gpttype_adapter.cpp
@@ -2929,6 +2929,11 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
 
         llama_ctx_params.offload_kqv = !inputs.low_vram;
         llama_ctx_params.kv_unified = true;
+        if((inputs.use_mtp || draftmodel_filename != "") && inputs.draft_amount > 0)
+        {
+            // Match llama-server's target rollback slots for speculative verification.
+            llama_ctx_params.n_rs_seq = inputs.draft_amount;
+        }
         model_params.use_mmap = inputs.use_mmap;
         model_params.use_mlock = inputs.use_mlock;
         model_params.use_direct_io = false; //no direct io for now until stable