[UnifiedTree] Support DeepSeek V4 host pool with multiple layouts. (#25282)

Co-authored-by: hzh0425 <hzh0425@apache.org>

test/registered/radix_cache/test_unified_radix_cache_kl.py

@@ -49,6 +49,8 @@ class UnifiedRadixTreeTestMixin:
     prefill_cache_assert = None
     decode_cache_assert = None
     sampling_temperature: float = 1
+    decode_hit_request_batch_size: int | None = None
+    decode_hit_inter_batch_delay_s: float = 0
 
     gsm8k_threshold: float = 0.93
     mmlu_threshold: float = 0.8
@@ -163,6 +165,8 @@ class UnifiedRadixTreeTestMixin:
             branches_per_group=branches,
             max_new_tokens=self.max_new_tokens,
             sampling_temperature=self.sampling_temperature,
+            request_batch_size=self.decode_hit_request_batch_size,
+            inter_batch_delay_s=self.decode_hit_inter_batch_delay_s,
         )
 
 

test/registered/radix_cache/test_unified_radix_cache_kl_hicache.py

@@ -93,8 +93,13 @@ def _assert_dsv4_decode_cached_tokens(result, history_len, output_len, label):
 class TestUnifiedDeepSeekV4FlashHiCache(UnifiedRadixTreeTestMixin, CustomTestCase):
     """DeepSeek V4 Flash FP8 + HiCache + UnifiedRadixCache."""
 
+    hicache_io_backend = "direct"
+    hicache_mem_layout = "page_first_direct"
+    max_running_requests = 4
     kl_threshold = 0.005
     sampling_temperature = 0
+    decode_hit_request_batch_size = 3
+    decode_hit_inter_batch_delay_s = 0.5
     decode_cache_assert = staticmethod(_assert_dsv4_decode_cached_tokens)
     gsm8k_threshold = 0.90
     num_gsm8k_questions = 100
@@ -130,15 +135,15 @@ class TestUnifiedDeepSeekV4FlashHiCache(UnifiedRadixTreeTestMixin, CustomTestCas
                 "--hicache-write-policy",
                 "write_through",
                 "--hicache-io-backend",
-                "direct",
+                cls.hicache_io_backend,
                 "--hicache-mem-layout",
-                "page_first_direct",
+                cls.hicache_mem_layout,
                 "--swa-full-tokens-ratio",
                 "0.25",
                 "--max-total-tokens",
                 "20000",
                 "--max-running-requests",
-                "2",
+                str(cls.max_running_requests),
             ],
             env={
                 "SGLANG_DSV4_FP4_EXPERTS": "0",
@@ -152,5 +157,14 @@ class TestUnifiedDeepSeekV4FlashHiCache(UnifiedRadixTreeTestMixin, CustomTestCas
         kill_process_tree(cls.process.pid)
 
 
+class TestUnifiedDeepSeekV4FlashHiCachePageFirstDirect(
+    TestUnifiedDeepSeekV4FlashHiCache
+):
+    """DeepSeek V4 Flash HiCache layout smoke: page_first_direct + direct."""
+
+    hicache_io_backend = "kernel"
+    hicache_mem_layout = "layer_first"
+
+
 if __name__ == "__main__":
     unittest.main()