[PD] Enable optimistic prefill with buffer-only L3 write-through HiCache (#40043)

Co-authored-by: cctry <csycfl@gmail.com>

python/sglang/srt/arg_groups/serving_hook.py

@@ -466,13 +466,20 @@ def handle_other_validations(server_args: Any):
                 "_handle_other_validations",
                 optimistic_prefill_attempts=0,
             )
-        elif cfg.enable_hierarchical_cache and (
-            cfg.hicache_storage_backend is not None
-            or cfg.hicache_write_policy != "write_back"
+        elif cfg.enable_hierarchical_cache and not (
+            (
+                cfg.hicache_storage_backend is None
+                and cfg.hicache_write_policy == "write_back"
+            )
+            or (
+                cfg.hicache_storage_backend is not None
+                and cfg.hicache_host_memory_mode == "buffer_only"
+                and cfg.hicache_write_policy == "write_through"
+            )
         ):
             logger.warning(
-                "Optimistic prefill only supports L2 hierarchical cache "
-                "with write-back policy"
+                "Optimistic prefill supports L2 write-back or L3 buffer-only "
+                "write-through hierarchical cache"
             )
             declare_resolution(
                 server_args,

python/sglang/srt/disaggregation/prefill.py

@@ -1496,6 +1496,17 @@ class SchedulerDisaggregationPrefillMixin:
         """Release KV cache and requeue an optimistic prefill request."""
         max_attempts = get_disagg().optimistic_prefill_attempts
         maybe_cache_unfinished_req(req, self.tree_cache)
+        # The cached prefix is evictable once the KV is released. Its length
+        # (capped at what a retry can match) seeds the retry's storage baseline,
+        # so an evicted prefix is looked up in L3 once before it is recomputed.
+        yielded_prefix_len = (
+            0
+            if req.skip_radix_cache_insert
+            else min(
+                req.kv.cache_protected_len,
+                req._compute_max_prefix_len(len(req.full_untruncated_fill_ids)),
+            )
+        )
         self._release_aborted_request(req)
         release_kv_cache(req, self.tree_cache)
         req.reset_for_retract()
@@ -1510,6 +1521,9 @@ class SchedulerDisaggregationPrefillMixin:
         req.pending_bootstrap = True
         req.time_stats.reset_prefill_retry_time()
         req.advance_cache_request_handle()
+        # A fresh lookup budget for the new attempt, as after a retraction.
+        req.storage_prefetch_retry_attempts = 0
+        req.storage_prefetch_last_match_len = yielded_prefix_len or None
         if req.prefill_attempt_count >= max_attempts:
             logger.info(
                 f"Req {req.rid} exhausted optimistic prefill attempts "