diff --git a/tests/v1/attention/test_dsv4_kernel_block_size.py b/tests/v1/attention/test_dsv4_kernel_block_size.py
new file mode 100644
index 000000000000..1cc0f0239a8d
--- /dev/null
+++ b/tests/v1/attention/test_dsv4_kernel_block_size.py
@@ -0,0 +1,36 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from vllm.models.deepseek_v4.nvidia.flashinfer_sparse import (
+    DeepseekV4FlashInferMLASparseBackend,
+)
+from vllm.models.deepseek_v4.sparse_mla import (
+    DeepseekV4SparseMLABackend,
+    dsv4_supported_kernel_block_sizes,
+)
+from vllm.platforms import current_platform
+from vllm.v1.attention.backends.mla.indexer import DeepseekV4IndexerBackend
+
+
+def test_dsv4_kernel_block_size_sm12x(monkeypatch):
+    monkeypatch.setattr(
+        current_platform, "is_device_capability_family", lambda family: family == 120
+    )
+    assert dsv4_supported_kernel_block_sizes() == [64]
+    assert DeepseekV4SparseMLABackend.get_supported_kernel_block_sizes() == [64]
+    assert DeepseekV4FlashInferMLASparseBackend.get_supported_kernel_block_sizes() == [
+        64
+    ]
+    assert DeepseekV4IndexerBackend.get_supported_kernel_block_sizes() == [64]
+
+
+def test_dsv4_kernel_block_size_not_sm12x(monkeypatch):
+    monkeypatch.setattr(
+        current_platform, "is_device_capability_family", lambda family: False
+    )
+    assert dsv4_supported_kernel_block_sizes() == [256]
+    assert DeepseekV4SparseMLABackend.get_supported_kernel_block_sizes() == [256]
+    assert DeepseekV4FlashInferMLASparseBackend.get_supported_kernel_block_sizes() == [
+        256
+    ]
+    assert DeepseekV4IndexerBackend.get_supported_kernel_block_sizes() == [256]
diff --git a/vllm/models/deepseek_v4/nvidia/flashinfer_sparse.py b/vllm/models/deepseek_v4/nvidia/flashinfer_sparse.py
index 883ef5ecafd4..81e5ea8e3303 100644
--- a/vllm/models/deepseek_v4/nvidia/flashinfer_sparse.py
+++ b/vllm/models/deepseek_v4/nvidia/flashinfer_sparse.py
@@ -25,7 +25,7 @@
 )
 from vllm.platforms.interface import DeviceCapability
 from vllm.utils.flashinfer import flashinfer_trtllm_batch_decode_sparse_mla_dsv4
-from vllm.v1.attention.backend import AttentionCGSupport, MultipleOf
+from vllm.v1.attention.backend import AttentionCGSupport
 from vllm.v1.attention.backends.mla.compressor_utils import (
     get_dspark_swa_index_width,
 )
@@ -110,10 +110,6 @@ class DeepseekV4FlashInferMLASparseBackend(DeepseekV4SparseMLABackend):
         "fp8_ds_mla",
     ]
 
-    @staticmethod
-    def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
-        return [256]
-
     @staticmethod
     def get_name() -> str:
         return "FLASHINFER_MLA_SPARSE_DSV4"
diff --git a/vllm/models/deepseek_v4/sparse_mla.py b/vllm/models/deepseek_v4/sparse_mla.py
index 4953bdc9256f..5f44c1465ae4 100644
--- a/vllm/models/deepseek_v4/sparse_mla.py
+++ b/vllm/models/deepseek_v4/sparse_mla.py
@@ -39,6 +39,19 @@
 _C128A_TOPK_ALIGNMENT = 128
 
 
+def dsv4_supported_kernel_block_sizes() -> list[int | MultipleOf]:
+    """Kernel page size for DeepSeek-V4 sparse MLA.
+
+    FlashInfer SM120 DSV4 decode is compiled for 64-token pages
+    (``_DECODE_DSV4_PAGE_BLOCK_SIZE = 64``). Keep manager ``--block-size 256``
+    so C128 storage is 2. Returning ``[64]`` lets ``select_common_block_size``
+    split each manager block into four kernel pages. SM100 stays at 256.
+    """
+    if current_platform.is_device_capability_family(120):
+        return [64]
+    return [256]
+
+
 class DeepseekV4SparseMLABackend(AttentionBackend):
     """DeepSeek-V4 sparse-MLA backend base.
 
@@ -58,7 +71,7 @@ class DeepseekV4SparseMLABackend(AttentionBackend):
 
     @staticmethod
     def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
-        return [256]
+        return dsv4_supported_kernel_block_sizes()
 
     @staticmethod
     def get_builder_cls() -> type["DeepseekV4SparseMLAMetadataBuilder"]:
diff --git a/vllm/v1/attention/backends/mla/indexer.py b/vllm/v1/attention/backends/mla/indexer.py
index 2321cf8d13d4..5729a9a2fbf5 100644
--- a/vllm/v1/attention/backends/mla/indexer.py
+++ b/vllm/v1/attention/backends/mla/indexer.py
@@ -251,7 +251,15 @@ def supported_kv_cache_layouts(cls) -> tuple[KVCacheLayout, ...]:
     @staticmethod
     def get_supported_kernel_block_sizes() -> list[int | MultipleOf]:
         # Block sizes count uncompressed tokens: C4 indexer pages hold 64 rows.
-        return [256]
+        # Imported lazily: vllm.models.deepseek_v4 transitively imports
+        # vllm._aiter_ops (via fused_moe), which would deadlock cold start
+        # with an import cycle when vllm._aiter_ops pulls in this module
+        # first (see vllm/v1/attention/ops/rocm_aiter_mla_sparse.py).
+        from vllm.models.deepseek_v4.sparse_mla import (
+            dsv4_supported_kernel_block_sizes,
+        )
+
+        return dsv4_supported_kernel_block_sizes()
 
 
 class DeepseekV41IndexerBackend(DeepseekV4IndexerBackend):