fix(hicache): preserve sparse restore checkpoints under host pressure

Choose a common validated trailing-state endpoint before partial host allocation, retain that endpoint set across query synchronization, and size auxiliary staging from the selected length. Assumed trailing hits prove only their original endpoint.

Add controller/cache CPU regressions and a PP checkpoint-vote control. Validation: 14 targeted CPU tests and repository pre-commit checks passed.
This commit is contained in:
2026-10-30 11:35:29 +08:00
parent 7f6fe22445
commit 0a56a5f6e7
4 changed files with 431 additions and 8 deletions
@@ -811,6 +811,25 @@ class HybridCacheController(BaseHiCacheController):
else:
available = self.mem_pool_host.available_size()
alloc_len = min(hit_tokens, available - available % self.page_size)
if any(
transfer.hit_policy == PoolHitPolicy.TRAILING_PAGES
for transfer in operation.pool_transfers or []
):
pages = operation.pool_storage_result.restorable_prefix_pages
if pages is not None:
# Capacity is not a restorable boundary: a shorter prefix can
# lack the SWA/Mamba checkpoint that made the queried hit usable.
# Choose before allocating, including shared side-pool buffers.
alloc_len = max(
(
page * self.page_size
for page in pages
if page * self.page_size <= alloc_len
),
default=0,
)
if alloc_len == 0:
return None, 0
if alloc_len >= min_tokens:
host_indices = self.alloc_prefetch_host_buffers(operation, alloc_len)
return host_indices, alloc_len
@@ -1377,10 +1396,19 @@ class HybridCacheController(BaseHiCacheController):
if operation.assume_stored:
# A prior hit on a suffix of this span proved it stored, and writes
# are prefix-covered, so re-querying only adds a round trip. Nothing
# is verified now, so no pool gets a verdict to learn from.
# are prefix-covered, so re-querying only adds a round trip. For a
# trailing pool this proves only the original endpoint, not every
# shorter checkpoint. Nothing is verified now, so no pool gets a
# verdict to learn from.
kv_hit_pages = len(hash_value)
operation.pool_storage_result.update_kv_hit_pages(kv_hit_pages)
if any(
transfer.hit_policy == PoolHitPolicy.TRAILING_PAGES
for transfer in operation.pool_transfers or []
):
operation.pool_storage_result.restorable_prefix_pages = (
[kv_hit_pages] if kv_hit_pages else []
)
operation.query_pool_hit_pages = {}
return hash_value, kv_hit_pages * self.page_size
@@ -1422,10 +1450,11 @@ class HybridCacheController(BaseHiCacheController):
"""One MIN collective for the folded hit count and every pool's own
boundary, so per-pool beliefs stay rank-identical. A pool omitted
by any rank heals from the folded cut without learning presence."""
if any(
has_trailing_pool = any(
transfer.hit_policy == PoolHitPolicy.TRAILING_PAGES
for transfer in operation.pool_transfers or []
):
)
if has_trailing_pool:
# A trailing state can exist at 512 and 1024 but not at 768. Taking
# MIN of each rank's longest hit would choose a missing checkpoint,
# so reduce one restorable flag per prefix length instead. Use the
@@ -1463,10 +1492,14 @@ class HybridCacheController(BaseHiCacheController):
for pool, pages in zip(pools, reduced[len(hits) :], strict=True)
if pages != _NO_POOL_VERDICT
}
if len(hits) == 1:
if not has_trailing_pool:
return int(reduced[0])
restorable = reduced[: len(hits)]
return max(i for i, ok in enumerate(restorable) if ok) * self.page_size
# Allocation may later need a shorter hit. Retain the common set, not
# the rank-local query result, so it cannot choose a peer's missing
# checkpoint after this collective has agreed on the longest one.
restorable = [i for i, ok in enumerate(reduced[: len(hits)]) if i and ok]
operation.pool_storage_result.restorable_prefix_pages = restorable
return max(restorable, default=0) * self.page_size
def _move_pool_indices(
self, host_pool, host_indices, device_indices, *, write_back_jit: bool
@@ -3111,7 +3111,7 @@ class UnifiedRadixCache(BasePrefixCache):
max(alloc_len, aux_hit_tokens), "host_capacity"
)
return True
if not self._alloc_prefetch_aux_staging(info, aux_hit_tokens):
if not self._alloc_prefetch_aux_staging(info, alloc_len):
# Same outcome as a KV shortfall: nothing stays staged.
cc.append_host_mem_release(host_indices=host_indices)
self._finish_storage_prefetch(