[NPU] Adapt hicache for K3 hybrid models (#39415)

test/manual/hicache/test_npu_kimi_k3_hicache.py

@@ -0,0 +1,53 @@
+import unittest
+
+from sglang.test.ascend.gsm8k_ascend_mixin import GSM8KAscendMixin
+from sglang.test.ascend.test_ascend_utils import KIMI_K3_W4A8_INT_MOE_WEIGHTS_PATH
+from sglang.test.test_utils import CustomTestCase
+
+
+class TestKimiK3MixedWithHiCacheL2(GSM8KAscendMixin, CustomTestCase):
+    """Testcase: Verify the inference accuracy of Kimi-K3 (MLA + KDA hybrid) on GSM8K
+    with mixed (non-PD) serving and HiCache L2 cache on NPU.
+
+    [Test Category] HiCache
+    [Test Target] Kimi-K3 (MLA + KDA hybrid, mamba layers)
+    [Test Config] Mixed deployment, NPU, HiCache L2 (kernel_ascend IO backend)
+    """
+
+    model = KIMI_K3_W4A8_INT_MOE_WEIGHTS_PATH
+    accuracy = 0.9
+    other_args = [
+        "--trust-remote-code",
+        "--device",
+        "npu",
+        "--attention-backend",
+        "ascend",
+        "--quantization",
+        "modelslim",
+        "--dtype",
+        "bfloat16",
+        "--tp-size",
+        "64",
+        "--enable-dp-attention",
+        "--dp-size",
+        "4",
+        "--enable-dp-lm-head",
+        "--moe-a2a-backend",
+        "deepep",
+        "--deepep-mode",
+        "auto",
+        "--mem-fraction-static",
+        "0.75",
+        "--max-mamba-cache-size",
+        "240",
+        "--enable-hierarchical-cache",
+        "--hicache-io-backend",
+        "kernel_ascend",
+        "--enable-cache-report",
+        "--hicache-ratio",
+        "4.0",
+    ]
+
+
+if __name__ == "__main__":
+    unittest.main()

test/registered/unit/disaggregation/test_decode_queue_cleanup.py

@@ -236,6 +236,10 @@ class TestDecodeQueueCleanup(CustomTestCase):
         queue._pre_alloc = MagicMock()
         queue.req_to_token_pool = MagicMock()
         queue.req_to_token_pool.available_size.return_value = 1
+        # Non-hybrid pools have no mamba allocator; MagicMock would otherwise
+        # auto-create one and break the `available_size() <= 0` comparison in
+        # pop_preallocated.
+        queue.req_to_token_pool.mamba_allocator = None
         queue.req_to_metadata_buffer_idx_allocator = MagicMock()
         queue.req_to_metadata_buffer_idx_allocator.available_size.return_value = 1
 

test/registered/unit/managers/test_priority_scheduling_disaggregation.py

@@ -148,6 +148,10 @@ class TestDecodePreallocQueuePriority(unittest.TestCase):
 
         queue.req_to_token_pool = MagicMock()
         queue.req_to_token_pool.available_size.return_value = 100
+        # Non-hybrid pools have no mamba allocator; MagicMock would otherwise
+        # auto-create one and break the `available_size() <= 0` comparison in
+        # pop_preallocated.
+        queue.req_to_token_pool.mamba_allocator = None
         queue.req_to_token_pool.req_to_token = torch.arange(
             8 * 16, dtype=torch.int64
         ).reshape(8, 16)

test/registered/unit/mem_cache/test_decode_radix_lock_ref.py

@@ -442,6 +442,10 @@ class TestDecodeLockRefScenarios(CustomTestCase):
         queue.tree_cache.dec_lock_ref = MagicMock()
         queue.req_to_token_pool = MagicMock()
         queue.req_to_token_pool.available_size.return_value = 1
+        # Non-hybrid pools have no mamba allocator; MagicMock would otherwise
+        # auto-create one and break the `available_size() <= 0` comparison in
+        # pop_preallocated.
+        queue.req_to_token_pool.mamba_allocator = None
         queue.req_to_metadata_buffer_idx_allocator = MagicMock()
         queue.req_to_metadata_buffer_idx_allocator.available_size.return_value = 1
         queue.token_to_kv_pool = MagicMock()

test/registered/unit/mem_cache/test_npu_mamba_async_layout.py

@@ -156,6 +156,7 @@ class TestNPUMambaAsyncConfig(unittest.TestCase):
             MambaPoolHost.__dict__["_copy_tensor_all_layers_lf_pf"], staticmethod
         )
 
+    @patch.object(mamba_pool_host, "transfer_mamba_state", None)
     def test_conv_only_load_skips_empty_temporal_component(self):
         pool = MambaPoolHost.__new__(MambaPoolHost)
         pool.layout = "page_first_direct"