diff --git a/python/sglang/srt/layers/quantization/fp8.py b/python/sglang/srt/layers/quantization/fp8.py
index b1a1df37a..b75e78936 100644
--- a/python/sglang/srt/layers/quantization/fp8.py
+++ b/python/sglang/srt/layers/quantization/fp8.py
@@ -50,7 +50,7 @@ from sglang.srt.layers.quantization.fp8_kernel import (
     scaled_fp8_quant,
 )
 from sglang.srt.layers.quantization.fp8_utils import (
-    _use_aiter_gfx95,
+    _use_aiter_bpreshuffle_gfx95,
     apply_fp8_linear,
     can_auto_enable_marlin_fp8,
     cutlass_fp8_supported,
@@ -515,7 +515,7 @@ class Fp8LinearMethod(LinearMethodBase):
         layer.weight_scale_inv.data = weight_scale.data
 
         if (
-            _use_aiter_gfx95
+            _use_aiter_bpreshuffle_gfx95
             and self.w8a8_block_fp8_linear is aiter_w8a8_block_fp8_linear
         ):
             n, k = layer.weight.shape
diff --git a/python/sglang/srt/layers/quantization/fp8_utils.py b/python/sglang/srt/layers/quantization/fp8_utils.py
index 8d075b688..016fe71f5 100755
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -34,6 +34,7 @@ from sglang.srt.utils import (
     get_bool_env_var,
     get_cuda_version,
     get_device_capability,
+    get_hip_version,
     is_blackwell_supported,
     is_cuda,
     is_flashinfer_available,
@@ -59,6 +60,8 @@ _is_musa = is_musa()
 
 _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 _use_aiter_gfx95 = _use_aiter and _is_gfx95_supported
+# ROCm 7.0 hipcc miscompiles gemm_a8w8_blockscale_bpreshuffle on gfx95 (#23319).
+_use_aiter_bpreshuffle_gfx95 = _use_aiter_gfx95 and get_hip_version() >= (7, 2, 0)
 
 
 def use_aiter_triton_gemm_w8a8_tuned_gfx950(n: int, k: int) -> bool:
@@ -761,7 +764,7 @@ def aiter_w8a8_block_fp8_linear(
 
     n, k = weight.shape
 
-    if _use_aiter_gfx95:
+    if _use_aiter_bpreshuffle_gfx95:
         use_triton = use_aiter_triton_gemm_w8a8_tuned_gfx950(n, k)
     else:
         use_triton = True
diff --git a/python/sglang/srt/utils/common.py b/python/sglang/srt/utils/common.py
index 8c7cb4ae2..19f3a4838 100644
--- a/python/sglang/srt/utils/common.py
+++ b/python/sglang/srt/utils/common.py
@@ -3442,6 +3442,12 @@ def is_gfx95_supported():
         return False
 
 
+def get_hip_version():
+    if torch.version.hip:
+        return tuple(map(int, torch.version.hip.split("-")[0].split(".")))
+    return (0, 0, 0)
+
+
 # LoRA-related constants and utilities
 SUPPORTED_LORA_TARGET_MODULES = [
     "q_proj",