diff --git a/vllm/model_executor/models/gemma4.py b/vllm/model_executor/models/gemma4.py
index f5528f5c34e9..4a1e891dc28b 100644
--- a/vllm/model_executor/models/gemma4.py
+++ b/vllm/model_executor/models/gemma4.py
@@ -1490,6 +1490,11 @@ def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
                 quant_params = router_quant_params[router_name]
                 if len(quant_params) == 3:
                     weight_name = f"{router_name}.weight"
+                    if is_pp_missing_parameter(weight_name, self):
+                        del router_quant_params[router_name]
+                        continue
+                    if weight_name not in params_dict:
+                        raise KeyError(weight_name)
                     param = params_dict[weight_name]
                     weight_loader = getattr(
                         param, "weight_loader", default_weight_loader
diff --git a/vllm/tokenizers/registry.py b/vllm/tokenizers/registry.py
index 3241aa1abbd0..9d1d812c2d19 100644
--- a/vllm/tokenizers/registry.py
+++ b/vllm/tokenizers/registry.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 import contextlib
+import copy
 from dataclasses import dataclass, field
 from functools import lru_cache
 from pathlib import Path
@@ -104,6 +105,7 @@ def _maybe_patch_gemma4_gguf_tokenizer(
     if not special_ids:
         return tokenizer
 
+    patched_tokenizer = copy.copy(tokenizer)
     token_attrs = {
         "padding_token_id": "pad_token",
         "bos_token_id": "bos_token",
@@ -114,12 +116,12 @@ def _maybe_patch_gemma4_gguf_tokenizer(
         token_id = special_ids.get(id_attr)
         if token_id is None:
             continue
-        token = tokenizer.convert_ids_to_tokens(token_id)
+        token = patched_tokenizer.convert_ids_to_tokens(token_id)
         if token is None:
             continue
-        setattr(tokenizer, token_attr, token)
+        setattr(patched_tokenizer, token_attr, token)
 
-    return tokenizer
+    return patched_tokenizer
 
 
 def resolve_tokenizer_args(
diff --git a/vllm/transformers_utils/processor.py b/vllm/transformers_utils/processor.py
index 3b5397113997..b4b5d30f65ac 100644
--- a/vllm/transformers_utils/processor.py
+++ b/vllm/transformers_utils/processor.py
@@ -1,6 +1,7 @@
 # SPDX-License-Identifier: Apache-2.0
 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
+import copy
 import importlib
 import inspect
 from functools import lru_cache
@@ -367,7 +368,9 @@ def cached_processor_from_config(
             model_config.model,
             getattr(model_config.hf_config, "model_type", None),
         )
-        processor.tokenizer = tokenizer
+        if tokenizer is not processor.tokenizer:
+            processor = copy.copy(processor)
+            processor.tokenizer = tokenizer
     return processor