From 1528d4e18224460109abe2e854eafd2941053fe3 Mon Sep 17 00:00:00 2001
From: biondizzle <biondizzle@gmail.com>
Date: Mon, 18 May 2026 22:46:04 +0000
Subject: [PATCH] Fix NVFP4 mapper: strip model. prefix from checkpoint keys

The NVFP4 checkpoint uses model.layers.* but vLLM's AutoWeightsLoader
expects layers.* (relative to the model module). Strip the model. prefix
instead of adding it.
---
 vllm/patches/deepseek_v4.py | 5 ++---
 1 file changed, 2 insertions(+), 3 deletions(-)

diff --git a/vllm/patches/deepseek_v4.py b/vllm/patches/deepseek_v4.py
index 31eb7f44..5e7ea781 100644
--- a/vllm/patches/deepseek_v4.py
+++ b/vllm/patches/deepseek_v4.py
@@ -1667,9 +1667,8 @@ def _make_deepseek_v4_nvfp4_weights_mapper() -> WeightsMapper:
 
     return WeightsMapper(
         orig_to_new_prefix={
-            "layers.": "model.layers.",
-            "embed.": "model.embed.",
-            "norm.": "model.norm.",
+            # Strip model. prefix (checkpoint has it, model params don't)
+            "model.": "",
         },
         orig_to_new_regex=expert_rename_regex,
         orig_to_new_suffix=suffix_renames,