diff --git a/python/sglang/srt/environ.py b/python/sglang/srt/environ.py
index 589447e07..3a724e7d6 100644
--- a/python/sglang/srt/environ.py
+++ b/python/sglang/srt/environ.py
@@ -683,6 +683,7 @@ class Envs:
 
     # RoPE cache configuration
     SGLANG_SPEC_EXPANSION_SAFETY_FACTOR = EnvInt(2)
+    SGLANG_ROPE_CACHE_FP32 = EnvBool(False)
     SGLANG_ROPE_CACHE_SAFETY_MARGIN = EnvInt(256)
     SGLANG_ROPE_CACHE_ALIGN = EnvInt(128)
 
diff --git a/python/sglang/srt/layers/rotary_embedding/base.py b/python/sglang/srt/layers/rotary_embedding/base.py
index 39cb8f7e2..e04cd876f 100644
--- a/python/sglang/srt/layers/rotary_embedding/base.py
+++ b/python/sglang/srt/layers/rotary_embedding/base.py
@@ -7,6 +7,7 @@ from typing import TYPE_CHECKING, Dict, List, Optional, Tuple, Union
 
 import torch
 
+from sglang.srt.environ import envs
 from sglang.srt.layers.rotary_embedding.utils import apply_rotary_emb
 from sglang.srt.layers.utils import MultiPlatformOp
 from sglang.srt.platforms import current_platform
@@ -93,8 +94,8 @@ class RotaryEmbedding(MultiPlatformOp):
         self.dtype = dtype
 
         cache = self._compute_cos_sin_cache()
-        # NOTE(ByronHsu): cache needs to be in FP32 for numerical stability
-        if not _is_cuda:
+        # NOTE(ByronHsu): cache needs to be in FP32 for numerical stability.
+        if not (_is_cuda or envs.SGLANG_ROPE_CACHE_FP32.get()):
             cache = cache.to(dtype)
 
         if (
@@ -178,8 +179,6 @@ class RotaryEmbedding(MultiPlatformOp):
 
     def _ensure_cos_sin_cache_length(self, needed_max_pos: int):
         """Ensure cos_sin_cache length > needed_max_pos."""
-        from sglang.srt.environ import envs
-
         cur_len = int(self.cos_sin_cache.shape[0])
         if needed_max_pos < cur_len:
             return