[AMD] [GLM-5.3-Flash Day 0] Enable zero-RoPE MHA prefill on ROCm (#39338)

Co-authored-by: Thomas Wang <thomawan@amd.com>
Co-authored-by: Kevin Mi <45493463+kevin-mii@users.noreply.github.com>

python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha_rocm.py

@@ -279,8 +279,14 @@ class DeepseekMHARocmForwardMixin:
     def _concat_and_cast_mha_k_rocm(
         self: DeepseekV2AttentionMLA,
         k_nope: torch.Tensor,
-        k_pe: torch.Tensor,
+        k_pe: torch.Tensor | None,
     ):
+        if self.qk_rope_head_dim == 0:
+            assert k_pe is None or k_pe.shape[-1] == 0
+            # No RoPE tail to append, so k is k_nope as-is. The concat branch
+            # below keeps k_nope's dtype, so no cast is needed here either.
+            return k_nope.contiguous()
+
         k_shape = (k_nope.shape[0], self.num_local_heads, self.qk_head_dim)
         k = k_nope.new_empty(*k_shape)
         if self.current_attention_backend == "aiter":