diff --git a/tests/v1/attention/test_rocm_glm5next_sparse.py b/tests/v1/attention/test_rocm_glm5next_sparse.py
index 509d8f1fae..ccc9e5c617 100644
--- a/tests/v1/attention/test_rocm_glm5next_sparse.py
+++ b/tests/v1/attention/test_rocm_glm5next_sparse.py
@@ -76,7 +76,9 @@ def test_fit_kpool_indices_rejects_narrow_input():
         ("auto", 512, 0, 2, 2, 1, True),
         ("fp8", 512, 1, 0, 0, 32, False),
         ("auto", 576, 1, 0, 0, 32, False),
-        ("auto", 512, 0, 2, 4, 2, False),
+        ("auto", 512, 0, 2, 4, 2, True),
+        ("auto", 512, 0, 2, 12, 6, True),
+        ("auto", 512, 0, 0, 0, 0, False),
     ],
 )
 def test_rocm_sparse_triton_route(
@@ -88,6 +90,7 @@ def test_rocm_sparse_triton_route(
     max_query_len,
     expected,
 ):
+    """Validate Triton routing for prefill, decode, and MTP verification."""
     assert (
         _use_rocm_sparse_triton(
             kv_cache_dtype=kv_cache_dtype,
diff --git a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py
index 3a9a4c92b3..45a413aab3 100644
--- a/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py
+++ b/vllm/v1/attention/backends/mla/rocm_aiter_mla_sparse.py
@@ -53,13 +53,16 @@ def _use_rocm_sparse_triton(
     num_decode_tokens: int,
     max_query_len: int,
 ) -> bool:
-    """Select the rope-free BF16 path not supported by AITER sparse MLA."""
-    plain_decode = num_decode_tokens == num_decodes
+    """Select the rope-free BF16 path not supported by AITER sparse MLA.
+
+    The ragged Triton kernel indexes metadata per query token, so multi-token
+    speculative verification rows have the same capability requirements as
+    plain decode rows.
+    """
     return (
         not kv_cache_dtype.startswith("fp8")
         and head_size == kv_lora_rank
-        and plain_decode
-        and (num_prefills > 0 or (num_decodes > 0 and max_query_len == 1))
+        and (num_prefills > 0 or num_decodes > 0)
     )