diff --git a/python/sglang/srt/models/bailing_moe.py b/python/sglang/srt/models/bailing_moe.py
index fef8818a9dba..c0b79f4eb4c4 100644
--- a/python/sglang/srt/models/bailing_moe.py
+++ b/python/sglang/srt/models/bailing_moe.py
@@ -170,10 +170,7 @@ def __init__(
             self.expert_bias = None
 
     def forward(self, hidden_states):
-        logits = F.linear(hidden_states.to(self.weight.dtype), self.weight, None).to(
-            hidden_states.dtype
-        )
-        return logits
+        return F.linear(hidden_states.to(self.weight.dtype), self.weight, None)
 
 
 class BailingMoESparseMoeBlock(nn.Module):
diff --git a/python/sglang/srt/models/bailing_moe_linear.py b/python/sglang/srt/models/bailing_moe_linear.py
index ce8f8770589e..353d880ae55d 100644
--- a/python/sglang/srt/models/bailing_moe_linear.py
+++ b/python/sglang/srt/models/bailing_moe_linear.py
@@ -224,10 +224,7 @@ def __init__(
             self.expert_bias = None
 
     def forward(self, hidden_states):
-        logits = F.linear(hidden_states.to(self.weight.dtype), self.weight, None).to(
-            hidden_states.dtype
-        )
-        return logits
+        return F.linear(hidden_states.to(self.weight.dtype), self.weight, None)
 
 
 class BailingMoE(nn.Module):
diff --git a/python/sglang/srt/models/ernie4.py b/python/sglang/srt/models/ernie4.py
index 5d30b1e77eba..6ddf450e5d2e 100644
--- a/python/sglang/srt/models/ernie4.py
+++ b/python/sglang/srt/models/ernie4.py
@@ -60,7 +60,7 @@ def __init__(
             torch.empty((config.moe_num_experts, config.hidden_size))
         )
         self.e_score_correction_bias = nn.Parameter(
-            torch.empty((1, config.moe_num_experts))
+            torch.empty((1, config.moe_num_experts), dtype=torch.float32)
         )
 
     def forward(self, hidden_states):
diff --git a/python/sglang/srt/models/kimi_linear.py b/python/sglang/srt/models/kimi_linear.py
index b0feaa7998d6..27b33df580f3 100644
--- a/python/sglang/srt/models/kimi_linear.py
+++ b/python/sglang/srt/models/kimi_linear.py
@@ -110,7 +110,9 @@ def __init__(
             prefix=f"{prefix}.gate",
         )
 
-        self.gate.e_score_correction_bias = nn.Parameter(torch.empty(num_experts))
+        self.gate.e_score_correction_bias = nn.Parameter(
+            torch.empty(num_experts, dtype=torch.float32)
+        )
 
         self.experts = get_moe_impl_class(quant_config)(
             num_experts=config.n_routed_experts,
diff --git a/python/sglang/srt/models/mimo_v2.py b/python/sglang/srt/models/mimo_v2.py
index f6f40c22f366..801576a70085 100644
--- a/python/sglang/srt/models/mimo_v2.py
+++ b/python/sglang/srt/models/mimo_v2.py
@@ -53,10 +53,7 @@
     resolve_linear_parallel_group,
 )
 from sglang.srt.layers.logits_processor import LogitsProcessor
-from sglang.srt.layers.moe import (
-    get_moe_a2a_backend,
-    get_moe_runner_backend,
-)
+from sglang.srt.layers.moe import get_moe_a2a_backend
 from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class
 from sglang.srt.layers.moe.topk import TopK, TopKOutputFormat
 from sglang.srt.layers.moe.utils import is_deepep_class_backend
@@ -360,15 +357,8 @@ def __init__(
             torch.empty((config.n_routed_experts, config.hidden_size), dtype=self.dtype)
         )
         if config.topk_method == "noaux_tc":
-            correction_bias_dtype = (
-                torch.bfloat16
-                if quant_config is not None
-                and quant_config.get_name() == "modelopt_fp4"
-                and get_moe_runner_backend().is_flashinfer_trtllm()
-                else torch.float32
-            )
             self.e_score_correction_bias = nn.Parameter(
-                torch.empty((config.n_routed_experts), dtype=correction_bias_dtype)
+                torch.empty((config.n_routed_experts), dtype=torch.float32)
             )
         else:
             self.e_score_correction_bias = None