Fix per-tier cached-token re-count for chunked prefill

python/sglang/srt/managers/schedule_policy.py

@@ -442,6 +442,10 @@ def __init__(
         self.can_run_list = []
         self.preempt_list = []
         self.new_chunked_req = None
+        # Continuation chunks of an already-running chunked req. Their prefix
+        # was credited to log_hit_tokens on the first chunk, so per-tier
+        # cached-token stats must not count them again.
+        self.continuation_chunk_reqs = []
         self.log_hit_tokens = 0
         self.reprocessed_log_hit_tokens = 0
         # TODO(lsyin): report the real input tokens excluding page alignment
@@ -698,6 +702,7 @@ def add_chunked_req(self, req: Req):
         req.set_extend_input_len(min(req.extend_input_len, _rem_tokens))
         req.fill_ids = req.fill_ids[: len(req.prefix_indices) + req.extend_input_len]
         self.can_run_list.append(req)
+        self.continuation_chunk_reqs.append(req)
         self._update_prefill_budget(
             0,
             req.extend_input_len,

python/sglang/srt/managers/scheduler_components/metrics_reporter.py

@@ -60,6 +60,15 @@ class PrefillStats:
     new_token_ratio: float
     num_running_reqs: QueueCount
     num_new_seqs: int  # len(can_run_list)
+
+    # Detailed prefix-cache hit split.
+    # log_hit_tokens should equal:
+    #   log_hit_tokens_device + log_hit_tokens_host + log_hit_tokens_storage
+    # when storage accounting is enabled. Without L3 storage, storage is 0.
+    log_hit_tokens_device: int = 0
+    log_hit_tokens_host: int = 0
+    log_hit_tokens_storage: int = 0
+
     reprocessed_log_input_tokens: int = 0
     reprocessed_log_hit_tokens: int = 0
     num_pending_tokens: int = 0
@@ -72,11 +81,30 @@ def from_adder(
         enable_priority_scheduling: bool = False,
         num_pending_tokens: int = 0,
     ):
+        # Skip continuation chunks of a chunked prefill: their prefix was
+        # already credited (to both log_hit_tokens and the per-tier split) on
+        # the first chunk, and re-reading req.cached_tokens_* here would count
+        # the same hit once per chunk.
+        continuation_reqs = getattr(adder, "continuation_chunk_reqs", [])
+        split_reqs = [req for req in adder.can_run_list if req not in continuation_reqs]
+        log_hit_tokens_device = sum(
+            getattr(req, "cached_tokens_device", 0) for req in split_reqs
+        )
+        log_hit_tokens_host = sum(
+            getattr(req, "cached_tokens_host", 0) for req in split_reqs
+        )
+        log_hit_tokens_storage = sum(
+            getattr(req, "cached_tokens_storage", 0) for req in split_reqs
+        )
+
         return cls(
             log_input_tokens=adder.log_input_tokens,
             log_hit_tokens=adder.log_hit_tokens,
             reprocessed_log_input_tokens=adder.reprocessed_log_input_tokens,
             reprocessed_log_hit_tokens=adder.reprocessed_log_hit_tokens,
+            log_hit_tokens_device=log_hit_tokens_device,
+            log_hit_tokens_host=log_hit_tokens_host,
+            log_hit_tokens_storage=log_hit_tokens_storage,
             new_token_ratio=adder.new_token_ratio,
             num_running_reqs=QueueCount.from_reqs(
                 running_reqs, enable_priority_scheduling
@@ -523,11 +551,25 @@ def report_prefill_stats(
         )
         iter_msg = f" [{batch_iter}]" if LOG_FORWARD_ITERS else ""
 
+        cache_split_msg = (
+            f"#cached-token-device: {prefill_stats.log_hit_tokens_device}, "
+            f"#cached-token-host: {prefill_stats.log_hit_tokens_host}, "
+        )
+
+        if (
+            getattr(self.scheduler, "enable_hicache_storage", False)
+            or prefill_stats.log_hit_tokens_storage > 0
+        ):
+            cache_split_msg += (
+                f"#cached-token-storage: {prefill_stats.log_hit_tokens_storage}, "
+            )
+
         msg = (
             f"Prefill batch{iter_msg}, "
             f"#new-seq: {prefill_stats.num_new_seqs}, "
             f"#new-token: {prefill_stats.log_input_tokens}, "
             f"#cached-token: {prefill_stats.log_hit_tokens}, "
+            f"{cache_split_msg}"
             f"{token_usage_msg}"
             f"#running-req: {prefill_stats.num_running_reqs.total}, "
             f"#queue-req: {len(self.scheduler.waiting_queue)}, "