diff --git a/python/sglang/srt/models/gemma4_causal.py b/python/sglang/srt/models/gemma4_causal.py
index 190452fcd124..c406f12a2b6c 100644
--- a/python/sglang/srt/models/gemma4_causal.py
+++ b/python/sglang/srt/models/gemma4_causal.py
@@ -1147,7 +1147,8 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
             ("experts.w13_weight", "experts.gate_up_proj", ("w1", "w3")),
             ("experts.w2_weight", "experts.down_proj", ("w2",)),
         ]
-        num_experts = self.config.num_experts
+        # Dense subclasses (e.g. the Gemma4 MTP assistant) reuse this.
+        num_experts = getattr(self.config, "num_experts", None) or 0
 
         # Per-expert checkpoint format used by compressed-tensors / FP8
         # (e.g. RedHatAI/*-FP8-Dynamic) and by ModelOpt NVFP4
@@ -1159,11 +1160,15 @@ def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         # in a trailing dot, so the standard `name.replace(weight_name,
         # param_name)` collapses every suffix uniformly to the fused
         # FusedMoE params (experts.w13_*, experts.w2_*).
-        per_expert_params_mapping = FusedMoE.make_expert_params_mapping(
-            ckpt_gate_proj_name="gate_proj",
-            ckpt_down_proj_name="down_proj",
-            ckpt_up_proj_name="up_proj",
-            num_experts=num_experts,
+        per_expert_params_mapping = (
+            FusedMoE.make_expert_params_mapping(
+                ckpt_gate_proj_name="gate_proj",
+                ckpt_down_proj_name="down_proj",
+                ckpt_up_proj_name="up_proj",
+                num_experts=num_experts,
+            )
+            if num_experts
+            else []
         )
 
         k_eq_v_layers = self._get_k_eq_v_layers()
diff --git a/python/sglang/srt/models/gemma4_mtp.py b/python/sglang/srt/models/gemma4_mtp.py
index 1cb87b7c2e99..ade10ce5b990 100644
--- a/python/sglang/srt/models/gemma4_mtp.py
+++ b/python/sglang/srt/models/gemma4_mtp.py
@@ -21,6 +21,7 @@
 from torch import nn
 from transformers import PretrainedConfig, PreTrainedModel
 
+from sglang.srt.distributed import get_pp_group
 from sglang.srt.layers.linear import ReplicatedLinear
 from sglang.srt.layers.logits_processor import (
     LogitsMetadata,
@@ -72,6 +73,7 @@ def __init__(
         self.assistant_config = config
         self.config = text_config
         self.quant_config = quant_config
+        self.pp_group = get_pp_group()
 
         self.vocab_size = text_config.vocab_size
         self.hidden_size = text_config.hidden_size
diff --git a/python/sglang/srt/server_args.py b/python/sglang/srt/server_args.py
index 1d1b8d29959d..2d203ad03cfd 100644
--- a/python/sglang/srt/server_args.py
+++ b/python/sglang/srt/server_args.py
@@ -2232,11 +2232,13 @@ def _handle_model_specific_adjustments(self):
             )
 
             if is_sm100_supported() and self.moe_runner_backend == "auto":
-
-                self.moe_runner_backend = "flashinfer_trtllm"
-                logger.info(
-                    "Use flashinfer_trtllm as MoE runner backend on SM100 for Gemma-4 NVFP4"
-                )
+                if self.get_model_config().quantization == "modelopt_fp4":
+                    self.quantization = "modelopt_fp4"
+                    self.moe_runner_backend = "flashinfer_trtllm"
+                    logger.info(
+                        "Use flashinfer_trtllm as MoE runner backend on "
+                        "SM100 for Gemma-4 (modelopt_fp4)"
+                    )
         elif model_arch == "MossVLForConditionalGeneration":
             if self.is_attention_backend_not_set():
                 self.prefill_attention_backend = "flashinfer"