Fix DeepSeek V4 PP HiCache SWA allocation and layer mapping (#29106)

Co-authored-by: hjzhang <zhanghjzzz@qq.com>
Co-authored-by: hzh0425 <hzh0425@apache.org>

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dsv4.py

@@ -34,6 +34,8 @@ def _assert_dsv4_decode_cached_tokens(result, history_len, output_len, label):
 class TestUnifiedDeepSeekV4FlashHiCache(UnifiedRadixTreeTestMixin, CustomTestCase):
     """DeepSeek V4 Flash FP8 + HiCache + UnifiedRadixCache."""
 
+    tp_size = 4
+    pp_size = 1
     hicache_io_backend = "direct"
     hicache_mem_layout = "page_first_direct"
     max_running_requests = 4
@@ -49,6 +51,43 @@ class TestUnifiedDeepSeekV4FlashHiCache(UnifiedRadixTreeTestMixin, CustomTestCas
     def test_multiturn_logprobs_match(self):
         pass
 
+    @classmethod
+    def _server_args(cls):
+        args = [
+            "--trust-remote-code",
+            "--tp-size",
+            str(cls.tp_size),
+        ]
+        if cls.pp_size != 1:
+            args += ["--pp-size", str(cls.pp_size)]
+        args += [
+            "--attention-backend",
+            "compressed",
+            "--page-size",
+            "256",
+            "--chunked-prefill-size",
+            "8192",
+            "--mem-fraction-static",
+            "0.9",
+            "--disable-shared-experts-fusion",
+            "--enable-hierarchical-cache",
+            "--hicache-ratio",
+            "4",
+            "--hicache-write-policy",
+            "write_through",
+            "--hicache-io-backend",
+            cls.hicache_io_backend,
+            "--hicache-mem-layout",
+            cls.hicache_mem_layout,
+            "--swa-full-tokens-ratio",
+            "0.25",
+            "--max-total-tokens",
+            "20000",
+            "--max-running-requests",
+            str(cls.max_running_requests),
+        ]
+        return args
+
     @classmethod
     def setUpClass(cls):
         cls.model = DSV4_FLASH_MODEL
@@ -57,35 +96,7 @@ class TestUnifiedDeepSeekV4FlashHiCache(UnifiedRadixTreeTestMixin, CustomTestCas
             cls.model,
             cls.base_url,
             timeout=DSV4_FLASH_LAUNCH_TIMEOUT,
-            other_args=[
-                "--trust-remote-code",
-                "--tp-size",
-                "4",
-                "--attention-backend",
-                "compressed",
-                "--page-size",
-                "256",
-                "--chunked-prefill-size",
-                "8192",
-                "--mem-fraction-static",
-                "0.9",
-                "--disable-shared-experts-fusion",
-                "--enable-hierarchical-cache",
-                "--hicache-ratio",
-                "4",
-                "--hicache-write-policy",
-                "write_through",
-                "--hicache-io-backend",
-                cls.hicache_io_backend,
-                "--hicache-mem-layout",
-                cls.hicache_mem_layout,
-                "--swa-full-tokens-ratio",
-                "0.25",
-                "--max-total-tokens",
-                "20000",
-                "--max-running-requests",
-                str(cls.max_running_requests),
-            ],
+            other_args=cls._server_args(),
             env={
                 "SGLANG_DSV4_FP4_EXPERTS": "0",
                 "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1",

test/registered/radix_cache/unified_radix_tree/test_unified_radix_cache_kl_dsv4_pp.py

@@ -0,0 +1,24 @@
+import unittest
+
+import test_unified_radix_cache_kl_dsv4 as dsv4_kl
+
+from sglang.test.ci.ci_register import register_cuda_ci
+
+register_cuda_ci(est_time=900, stage="extra-b", runner_config="8-gpu-h200")
+
+
+class TestUnifiedDeepSeekV4FlashHiCachePP4TP2(
+    dsv4_kl.TestUnifiedDeepSeekV4FlashHiCache
+):
+    """DeepSeek V4 Flash FP8 + HiCache + UnifiedRadixCache under PP4 TP2."""
+
+    pp_size = 4
+    tp_size = 2
+
+    @unittest.skip("PP4TP2 coverage uses accuracy and cache-hit KL cases.")
+    def test_multiturn_logprobs_match(self):
+        pass
+
+
+if __name__ == "__main__":
+    unittest.main()