[HiCache] Rework the buffer-mode storage prefetch pipeline and retry bookkeeping

python/sglang/srt/arg_groups/fields/memory.py

@@ -177,17 +177,26 @@ class Memory(msgspec.Struct):
         int,
         Arg(
             help=(
-                "Scheduling passes a queued request waits after a storage "
-                "prefetch miss before the availability check is retried "
-                "(under load the first check can run before the needed "
-                "backup commits). 0 disables retries."
+                "Scheduling passes a queued request waits before its storage "
+                "availability check is re-issued, when the prefetch found "
+                "nothing and a backup may still be committing (under load the "
+                "first check can run before it does). A re-issue that waits on "
+                "staging or a moved match instead goes out on the next pass. "
+                "Only passes that reach prefill scheduling count. 0 disables "
+                "miss retries; known-hit deferrals are always re-issued."
             ),
         ),
-    ] = 0
+    ] = 8
     hicache_storage_prefetch_retry_max_attempts: A[
         int,
-        "Maximum storage prefetch retries per request when --hicache-storage-prefetch-retry-poll-interval is set.",
-    ] = 4
+        Arg(
+            help=(
+                "Storage availability re-issues a queued request may make, paced "
+                "miss polls and immediate re-issues alike; past the cap it is "
+                "admitted with whatever the device holds. 0 disables re-issues."
+            ),
+        ),
+    ] = 8
 
     # -------------------------------------------------------------------------
     # Unified Radix Cache

python/sglang/srt/disaggregation/prefill.py

@@ -592,6 +592,7 @@ def get_next_disagg_prefill_batch_to_run(
         last_batch: Optional[ScheduleBatch],
     ) -> NextBatchPlan:
         self.process_pending_chunked_abort()
+        self._process_hicache_events()
 
         # HACK (byronhsu): reset the batch_is_full flag because we never enter update_running_batch which resets it
         # Otherwise, it hangs under high concurrency

python/sglang/srt/hardware_backend/npu/dsv4/c128_sidecar_component.py

@@ -402,6 +402,7 @@ def build_hicache_transfers(
         host_indices: Optional[torch.Tensor] = None,
         token_ids: Optional[Sequence[int]] = None,
         prefetch_tokens: int = 0,
+        staging_tokens: int = 0,
         last_hash: Optional[str] = None,
     ) -> Optional[list[PoolTransfer]]:
         ct = self.component_type

python/sglang/srt/managers/schedule_batch.py

@@ -149,6 +149,7 @@
     from sglang.srt.configs.model_config import ModelConfig
     from sglang.srt.managers.hisparse_coordinator import HiSparseCoordinator
     from sglang.srt.managers.scheduler_components.metrics_reporter import PrefillStats
+    from sglang.srt.mem_cache.storage_prefetch import StagedPrefetchPlan
     from sglang.srt.session.session_controller import Session
     from sglang.srt.speculative.spec_info import SpecInput, SpeculativeAlgorithm
 
@@ -1127,14 +1128,14 @@ def __init__(
         self.host_loaded_length = 0
         # Buffer-mode host memory is transport staging, not an L2 cache tier.
         self.host_hit_is_storage = False
-        # Storage prefetch retry state while queued
-        # (see Scheduler._retry_missed_storage_prefetches).
-        self.storage_prefetch_retry_pending = False
-        self.storage_prefetch_retry_wait_polls = 0
         self.storage_prefetch_retry_attempts = 0
+        self.staged_prefetch_plan: Optional[StagedPrefetchPlan] = None
         # Receipt of the tree lock held on last_node (anchor, SWA boundary,
         # skipped components); every release replays it unchanged.
         self.lock_receipt: DecLockRefParams = DecLockRefParams()
+        # Device/host prefix used to plan the latest L3 lookup. Admission uses
+        # it to detect newly exposed storage demand after queue-time eviction.
+        self.storage_prefetch_last_match_len: Optional[int] = None
         # Whether the prefill-time SWA tree lock has been released early
         self.swa_prefix_lock_released: bool = False
         # Logical-page KV sharding: rotation base of the chain this request