diff --git a/vllm/utils/deep_gemm.py b/vllm/utils/deep_gemm.py
index 38d92f01192b1..4083193d7650e 100644
--- a/vllm/utils/deep_gemm.py
+++ b/vllm/utils/deep_gemm.py
@@ -135,7 +135,7 @@ DEFAULT_BLOCK_SIZE = [128, 128]
 
 
 # Taken from https://github.com/deepseek-ai/DeepGEMM/blob/dd6ed14acbc7445dcef224248a77ab4d22b5f240/deep_gemm/utils/math.py#L38
-# TODO(wentao): optimize this function, using triton or cuda kernel
+@torch.compile(dynamic=True, backend=current_platform.simple_compile_backend)
 def per_block_cast_to_fp8(
         x: torch.Tensor,
         block_size: list[int] = DEFAULT_BLOCK_SIZE,
@@ -187,4 +187,4 @@ __all__ = [
     "is_deep_gemm_e8m0_used",
     "is_deep_gemm_supported",
     "should_use_deepgemm_for_fp8_linear",
-]
+]
\ No newline at end of file