Preserve pending CP HiCache backups through final insert

Deferred non-chunked inserts can happen while a CP HiCache per-layer backup is still in flight. Rolling back the unattached prepared backup at unfinished-request time drops the only reservation that final insertion can attach, so cache_finished_req has to start a post-forward backup and catch up all layers synchronously.\n\nKeep the prepared backup for non-chunked deferred inserts and continue rolling it back for chunked middle inserts, where the suffix may be extended by later chunks and the old backup would cover the wrong range. Document the failure mode so future changes do not rediscover the same fallback path.\n\nConstraint: CP HiCache radix split cannot mutate in-flight backup nodes.\nConstraint: Chunked prefill middle inserts still need rollback because their backup range is not final.\nRejected: Always rollback unattached backups | causes post-forward catch_up_all_layers for non-chunked deferred inserts.\nRejected: Always preserve unattached backups | can attach stale backup ranges for chunked middle inserts.\nConfidence: high\nScope-risk: narrow\nDirective: Do not clear req.cp_hicache_prepared_backup on non-chunked deferred insert without proving final insert no longer needs it.\nTested: remote cjy-glm5-new py_compile radix_cache.py\nTested: remote cjy-glm5-new pytest test_cp_hicache_metadata.py::{nonchunked preserve,chunked rollback} => 2 passed\nNot-tested: live ETE replay after restarting prefill with this exact commit
This commit is contained in:
laoyao0822
2026-06-13 03:44:23 +08:00
parent ceb5345410
commit 1d168d061f
3 changed files with 102 additions and 1 deletions

View File

@@ -5489,3 +5489,48 @@ Validation:
- Remote targeted regression set passed (`4 passed`): disabled sync materialize,
consume miss warning, first-layer no-warning, and create-skip warning when
enabled.
### C128 — 2026-06-13 Non-chunked deferred insert must preserve prepared CP backup
Finding:
- Remote log `sglang_cp_hicache_20260612_183417.log` still showed
`[CP_HICACHE_FALLBACK][post_forward_catch_up_backup]` followed by
`[CP_HICACHE_FALLBACK][catch_up_all_layers]` for a small set of nodes.
- The fallback always followed `insert_deferred_pending_backup_split`, usually
with `reason=node_pending_backup` and non-empty `ongoing/pending` write state.
- This means the request had already completed forward and tried to update the
radix tree, but the insert could not safely split a node whose CP HiCache
backup was still in flight.
- `cache_unfinished_req()` treated every deferred insert the same way: it rolled
back an unattached `cp_hicache_prepared_backup` and cleared it from the request.
- For disaggregated non-chunked prefill this rollback is too early. The request
is still kept in the inflight transfer queue and later reaches
`cache_finished_req()` when the transfer completes. If the prepared backup is
preserved, that final insert can attach the already-registered per-layer
backup. If it is cleared, the final insert has no prepared reservation and
immediately falls back to post-forward all-layer catch-up backup.
Correction:
- Preserve unattached prepared CP HiCache backup across deferred
`cache_unfinished_req(chunked=False)` so the later final insert can attach it.
- Keep the previous rollback behavior for `chunked=True` middle chunks. A
chunked request may execute another forward for the same request before final
insert; retaining an old prepared reservation would cover the wrong suffix and
block the next chunk from preparing its own backup.
Validation:
- Remote RED/GREEN targeted tests in `cjy-glm5-new`:
- `test_cache_unfinished_req_deferred_insert_preserves_unattached_backup_for_final_insert`
- `test_cache_unfinished_chunked_deferred_insert_rolls_back_unattached_backup`
- Remote full metadata suite passed:
`PYTHONPATH=python python -m pytest -q test/registered/unit/mem_cache/test_cp_hicache_metadata.py`
=> `127 passed, 21 warnings`.
Remaining validation gap:
- Needs a fresh ETE/replay run after restart to confirm the runtime
`CP_HICACHE_FALLBACK][catch_up_all_layers]` events disappear or reduce to only
genuinely unsupported cases.

View File

@@ -725,7 +725,15 @@ class RadixCache(BasePrefixCache):
prepared_cp_backup is not None
and not getattr(prepared_cp_backup, "attached", False)
and hasattr(self, "_rollback_prepared_cp_backup")
and chunked
):
# A chunked middle insert can be followed by another forward for
# the same request, so an unattached prepared reservation would
# cover the wrong suffix and block the next chunk from preparing
# its own backup. Non-chunked disagg prefill still has a later
# cache_finished_req() pass after KV transfer; keep the prepared
# backup alive so that final insert can attach it instead of
# falling back to post-forward all-layer catch-up backup.
self._rollback_prepared_cp_backup(
prepared_cp_backup, "unfinished_pending_backup_split_deferred"
)

View File

@@ -2019,7 +2019,9 @@ class TestHiRadixCacheCPBackup(CustomTestCase):
self.assertEqual(req.prefix_indices.tolist(), list(range(6)))
self.assertEqual(cache.token_to_kv_pool_allocator.freed, [])
def test_cache_unfinished_req_deferred_insert_rolls_back_unattached_backup(self):
def test_cache_unfinished_req_deferred_insert_preserves_unattached_backup_for_final_insert(
self,
):
cache = HiRadixCache.__new__(HiRadixCache)
cache.disable = False
cache._uses_cp_hicache = True
@@ -2059,6 +2061,52 @@ class TestHiRadixCacheCPBackup(CustomTestCase):
cache.cache_unfinished_req(req, chunked=False)
self.assertEqual(rollbacks, [])
self.assertIs(req.cp_hicache_prepared_backup, prepared)
self.assertEqual(req.prefix_indices.tolist(), list(range(6)))
def test_cache_unfinished_chunked_deferred_insert_rolls_back_unattached_backup(
self,
):
cache = HiRadixCache.__new__(HiRadixCache)
cache.disable = False
cache._uses_cp_hicache = True
cache.is_eagle = False
cache.page_size = 4
deferred_node = TreeNode()
rollbacks = []
class Pool:
req_to_token = torch.arange(8, dtype=torch.int64).view(1, 8)
cache.req_to_token_pool = Pool()
cache.token_to_kv_pool_allocator = RecordingTokenAllocator()
cache.insert = lambda params: types.SimpleNamespace(
prefix_len=1,
pending_backup_deferred_node=deferred_node,
)
cache.match_prefix = lambda params: (_ for _ in ()).throw(
AssertionError("deferred insert must not rematch")
)
cache._free_kv_indices_range = lambda *args, **kwargs: (_ for _ in ()).throw(
AssertionError("deferred unfinished insert must keep request KV")
)
cache._rollback_prepared_cp_backup = lambda prepared, reason: rollbacks.append(
(prepared, reason)
)
prepared = types.SimpleNamespace(attached=False)
req = types.SimpleNamespace(
fill_ids=list(range(6)),
req_pool_idx=0,
extra_key=None,
cache_protected_len=1,
last_node=TreeNode(),
prefix_indices=torch.empty((0,), dtype=torch.int64),
cp_hicache_prepared_backup=prepared,
)
cache.cache_unfinished_req(req, chunked=True)
self.assertEqual(
rollbacks, [(prepared, "unfinished_pending_backup_split_deferred")]
)