Let tail chunks co-batch with cache-hit requests (SGLANG_CP_PREFILL_MIX_CHUNKED)

专题 S1 (design: docs_internal/perf/prefill-compute-intensity-plan.md
S1.0-S1.11).  A batch containing a chunked prefill request has been
forced to bs=1 by the CP gate, so every chunk of a long prompt
monopolizes a forward while short-extend cache-hit continuations queue
— the direct cause of the replay TTFT tail (p90 19.3s / p99 49.2s at
91.8% cache hit).  Yet mixed chunk batches already occur today (a
freshly-chunked request keeps earlier-admitted small requests), proving
the CP forward path is mixed-chunk-safe; only admission was asymmetric.

Three changes, the first flag-independent:
- add_chunked_req now seeds the budget with the chunk's TRUE prefix
  (was 0), so the CP cached tally and the buffer estimator's mqa_logits
  k_rows see the chunk's footprint before any later request is admitted
  (landmine D1).
- New SGLANG_CP_PREFILL_MIX_CHUNKED (default OFF): with it on, the gate
  admits requests after a chunked one and lets the existing CP caps
  (extend / cached / buffer, now correctly seeded) bound the batch — a
  FULL chunk still ends the scan by consuming the chunk-clamped extend
  cap; only a tail chunk leaves headroom.  A chunked prefix that is not
  page-aligned (rare sub-page final-chunk tail) keeps its batch solo
  (the CP page-aligned split would fail-fast otherwise).
- The symm staging capacity identity (admission extend cap + request
  slack == staging pages) is asserted when the flag is on, locking the
  coupling the design relies on (plan doc S1.4 I2).

Tests: 4 new adder units (budget seeding; tail chunk admits followers;
full chunk solo by budget; non-aligned prefix solo); the 8-rank
byte-exactness scenario gains a chunk-shaped request (2048-token
page-aligned carried prefix + 512 extend) — all four phases
(legacy/v2/symm/prefetch) byte-identical on g0033.  Known pre-existing
cross-file pollution noted in problems.md P16.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

test/manual/test_cp_shared_kv_compose_v2_8rank.py

@@ -42,11 +42,17 @@ class _CpGroupShim:
 
 
 def _build_scenario(rank: int, cp_size: int, device: torch.device):
-    """bs=4 with mixed prefix/extend lengths, extends crossing page bounds."""
+    """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.
+    """
     page_size = 64
-    batch_size = 4
-    prefix_lens = [640, 1280, 320, 640]
-    extend_lens = [95, 130, 64, 200]
+    prefix_lens = [640, 1280, 320, 640, 2048]
+    extend_lens = [95, 130, 64, 200, 512]
+    batch_size = len(prefix_lens)
     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

@@ -6,6 +6,8 @@ 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"
@@ -845,6 +847,107 @@ class TestPrefillAdder(CustomTestCase):
         )
         self.assertEqual([req.rid for req in adder.can_run_list], ["chunked"])
 
+    def test_add_chunked_req_seeds_true_prefix_into_cp_budget(self):
+        # C1 (plan doc S1.1-1a): the chunk's carried prefix must count toward
+        # the CP cached tally so later admission gates see its footprint.
+        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 = 10000
+        adder = self.create_adder(
+            self.create_running_batch(),
+            page_size=64,
+            rem_input_tokens=4096,
+            rem_chunk_tokens=4096,
+            enable_cp_shared_kv_prefill_bs_gt1=True,
+            cp_shared_kv_prefill_max_batch_requests=8,
+            cp_shared_kv_prefill_max_total_extend_tokens=4096,
+        )
+        chunked = self.create_prefill_req("chunked", extend_input_len=128)
+        chunked.prefix_indices = torch.zeros((256,), dtype=torch.int64)
+        adder.add_chunked_req(chunked)
+        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(