From 5d16e81dd9896355d36b363bbf786abaa8f6995f Mon Sep 17 00:00:00 2001 From: ynankani Date: Tue, 11 Aug 2026 13:19:05 +0000 Subject: [PATCH] convert : keep quantization scales for nemotron --mtp export (#26903) Signed-off-by: ynankani --- conversion/nemotron.py | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/conversion/nemotron.py b/conversion/nemotron.py index e5075020c1..c46cec1438 100644 --- a/conversion/nemotron.py +++ b/conversion/nemotron.py @@ -275,10 +275,18 @@ class NemotronHModel(GraniteHybridModel): return None elif cls.mtp_only: # --mtp: export the MTP head plus the tensors it shares with the target model + # Include lm_head scale sidecars so NVFP4 packing sees them. keep = name in ( "backbone.embeddings.weight", "backbone.norm_f.weight", "lm_head.weight", + "lm_head.weight_scale", + "lm_head.weight_scale_2", + "lm_head.weight_scale_inv", + "lm_head.input_scale", + "lm_head.input_global_scale", + "lm_head.weight_global_scale", + "lm_head.weight_packed", ) if not keep: return None