From c3e3f9e5338e8eb3d5b13d20f20d38b53d9d19ce Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?Sigbj=C3=B8rn=20Skj=C3=A6ret?= <sigbjorn.skjaeret@scala.com>
Date: Thu, 12 Mar 2026 10:04:20 +0100
Subject: [PATCH] convert : better mtp check and fix return [no ci] (#20419)

---
 convert_hf_to_gguf.py | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/convert_hf_to_gguf.py b/convert_hf_to_gguf.py
index 6e3d3ad1dc..37834c78b8 100755
--- a/convert_hf_to_gguf.py
+++ b/convert_hf_to_gguf.py
@@ -10092,9 +10092,9 @@ class NemotronHModel(GraniteHybridModel):
             # Skip Multi-Token Prediction (MTP) tensors. These are used for
             # for speculative decoding but we don't include them in this model
             # conversion. See https://github.com/ggml-org/llama.cpp/pull/18886
-            if "mtp" in name:
+            if name.startswith("mtp."):
                 logger.info(f"gguf: Skipping MTP (Speculative) layer: {name}")
-                return []
+                return
 
             if name.endswith("mixer.gate.e_score_correction_bias"):
                 new_name = name.replace("e_score_correction_bias", "e_score_correction.bias")