Keep chunked CP prefills solo during bs>1 admission

Revert the tail-chunk co-batching gate from 54c056af because allowing a continued chunk to share the next CP bs>1 batch reopens the mixed chunk/page-tail scheduler risks we are currently avoiding. Keep the independent real-prefix budget accounting so chunked requests still contribute their carried prefix to CP cached-token and buffer estimates.\n\nConstraint: Chunked-prefill requests must remain solo until the CP split/page-tail contract is revalidated for mixed batches.\nRejected: Full revert of 54c056af | it would also drop true-prefix budget accounting and under-estimate cache/buffer pressure for admitted chunks.\nConfidence: high\nScope-risk: moderate\nDirective: Do not reintroduce tail-chunk co-batching without tests covering page-tail split, CP buffer admission, and ETE chunked+cache-hit replay.\nTested: Local py_compile for environ.py, schedule_policy.py, cp_shared_kv_compose.py, test_prefill_adder.py, test_cp_shared_kv_compose_v2_8rank.py.\nTested: Remote cjy-glm5-new PYTHONPATH=python pytest -q test/registered/unit/managers/test_prefill_adder.py -> 25 passed.\nTested: Remote cjy-glm5-new PYTHONPATH=python pytest -q test/registered/unit/mem_cache/test_cp_shared_kv_runtime.py -> 157 passed, 2 subtests passed.\nNot-tested: Full mixed replay with chunked-prefill traffic after service restart.

test/manual/test_cp_shared_kv_compose_v2_8rank.py

@@ -42,17 +42,11 @@ class _CpGroupShim:
 
 
 def _build_scenario(rank: int, cp_size: int, device: torch.device):
-    """bs=5 with mixed prefix/extend lengths, extends crossing page bounds.
-
-    The last request is CHUNK-SHAPED (large page-aligned carried prefix +
-    a tail-chunk-sized extend): with mix-chunked co-batching (plan doc S1)
-    a continued chunk is exactly such a (prefix, extend) pair to the CP
-    compose machinery, so this scenario byte-validates the mixed batch.
-    """
+    """bs=4 with mixed prefix/extend lengths, extends crossing page bounds."""
     page_size = 64
-    prefix_lens = [640, 1280, 320, 640, 2048]
-    extend_lens = [95, 130, 64, 200, 512]
-    batch_size = len(prefix_lens)
+    batch_size = 4
+    prefix_lens = [640, 1280, 320, 640]
+    extend_lens = [95, 130, 64, 200]
     kv_dim = 656  # fp8 layout bytes/token
 
     from sglang.srt.mem_cache.cp_shared_kv_compute_owner import (

test/registered/unit/managers/test_prefill_adder.py

@@ -6,8 +6,6 @@ from unittest.mock import MagicMock
 
 import torch
 
-from sglang.srt.environ import envs
-
 if "sgl_kernel" not in sys.modules:
     sys.modules["sgl_kernel"] = types.ModuleType("sgl_kernel")
 sys.modules["sgl_kernel"].__file__ = "sgl_kernel_stub.py"
@@ -987,81 +985,6 @@ class TestPrefillAdder(CustomTestCase):
         self.assertEqual(adder.cp_shared_kv_prefill_total_cached_tokens, 256)
         self.assertEqual(adder.cp_shared_kv_prefill_total_extend_tokens, 128)
 
-    def _mix_chunked_adder(self, *, rem_chunk_tokens, extend_cap):
-        set_global_server_args_for_scheduler(
-            ServerArgs(
-                model_path="dummy",
-                enable_nsa_prefill_context_parallel=True,
-                nsa_prefill_cp_mode="in-seq-split",
-            )
-        )
-        self.mock_token_allocator.available_size.return_value = 100000
-        return self.create_adder(
-            self.create_running_batch(),
-            page_size=64,
-            rem_input_tokens=8192,
-            rem_chunk_tokens=rem_chunk_tokens,
-            enable_cp_shared_kv_prefill_bs_gt1=True,
-            cp_shared_kv_prefill_max_batch_requests=8,
-            cp_shared_kv_prefill_max_total_extend_tokens=extend_cap,
-        )
-
-    def test_cp_prefill_mix_chunked_tail_chunk_admits_following_requests(self):
-        # Plan doc S1: with the flag on, a TAIL chunk (extend below the chunk
-        # budget, page-aligned carried prefix) leaves extend-cap headroom and
-        # the following short-extend request co-batches with it.
-        with envs.SGLANG_CP_PREFILL_MIX_CHUNKED.override(True):
-            adder = self._mix_chunked_adder(rem_chunk_tokens=4096, extend_cap=4096)
-            chunked = self.create_prefill_req("chunked", extend_input_len=1280)
-            chunked.prefix_indices = torch.zeros((4096,), dtype=torch.int64)
-            self.assertIsNone(adder.add_chunked_req(chunked))  # tail chunk
-            normal = self.create_prefill_req("normal", extend_input_len=128)
-            self.assertEqual(
-                adder.add_one_req(
-                    normal, has_chunked_req=True, truncation_align_size=None
-                ),
-                AddReqResult.CONTINUE,
-            )
-            self.assertEqual(
-                [req.rid for req in adder.can_run_list], ["chunked", "normal"]
-            )
-
-    def test_cp_prefill_mix_chunked_full_chunk_stays_solo_by_budget(self):
-        # A FULL chunk consumes the whole (chunk-clamped) extend cap, so the
-        # first following request is rejected by the extend gate — no special
-        # code, the budget arithmetic ends the scan.
-        with envs.SGLANG_CP_PREFILL_MIX_CHUNKED.override(True):
-            adder = self._mix_chunked_adder(rem_chunk_tokens=256, extend_cap=4096)
-            chunked = self.create_prefill_req("chunked", extend_input_len=512)
-            chunked.prefix_indices = torch.zeros((4096,), dtype=torch.int64)
-            self.assertIs(adder.add_chunked_req(chunked), chunked)  # truncated
-            self.assertEqual(chunked.extend_input_len, 256)
-            normal = self.create_prefill_req("normal", extend_input_len=128)
-            self.assertEqual(
-                adder.add_one_req(
-                    normal, has_chunked_req=True, truncation_align_size=None
-                ),
-                AddReqResult.OTHER,
-            )
-            self.assertEqual([req.rid for req in adder.can_run_list], ["chunked"])
-
-    def test_cp_prefill_mix_chunked_non_aligned_prefix_stays_solo(self):
-        # I1 guard: a chunked prefix that is not a page multiple would break
-        # the CP page-aligned split in a multi-request batch — keep it solo.
-        with envs.SGLANG_CP_PREFILL_MIX_CHUNKED.override(True):
-            adder = self._mix_chunked_adder(rem_chunk_tokens=4096, extend_cap=4096)
-            chunked = self.create_prefill_req("chunked", extend_input_len=1280)
-            chunked.prefix_indices = torch.zeros((100,), dtype=torch.int64)
-            self.assertIsNone(adder.add_chunked_req(chunked))
-            normal = self.create_prefill_req("normal", extend_input_len=128)
-            self.assertEqual(
-                adder.add_one_req(
-                    normal, has_chunked_req=True, truncation_align_size=None
-                ),
-                AddReqResult.OTHER,
-            )
-            self.assertEqual([req.rid for req in adder.can_run_list], ["chunked"])
-
     def test_cp_prefill_total_cached_limit_stops_second_cached_request(self):
         set_global_server_args_for_scheduler(
             ServerArgs(