diff --git a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
index 5c33248f7..62d407387 100644
--- a/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
+++ b/python/sglang/srt/models/deepseek_common/attention_forward_methods/forward_mha.py
@@ -656,9 +656,14 @@ class DeepseekMHAForwardMixin:
             # reads cached prefix KV crashes with "576 != 656".
             kv_indices = filter_dcp_local_kv_indices(kv_indices=kv_indices)
             # Read door: the pool never translates, so the production site does.
-            kv_indices = get_attn_backend().kv_index_translator.translate_dcp_read_ids(
-                kv_indices
-            )
+            # Only the FlashAttention/FlashInfer backends bind a translator; the
+            # base class documents its None default as "no translate", and the DSA
+            # backend the EAGLE draft model runs on keeps that default. With no
+            # translator the ids never went VIRTUAL, so there is nothing to
+            # collapse.
+            translator = get_attn_backend().kv_index_translator
+            if translator is not None:
+                kv_indices = translator.translate_dcp_read_ids(kv_indices)
             kv_a, k_pe = get_token_to_kv_pool().get_mla_kv_buffer(
                 self.attn_mha, kv_indices, torch.bfloat16
             )