CP HiCache metrics: fix tp_rank AttributeError under CP + label by cp_rank

_cp_maybe_collect_metrics read cache_controller.tp_rank/dp_rank, but those were
resolved only inside _generate_storage_config -- which never runs under CP
shared-L2 (storage backends are forbidden there), so the first metrics emit
crashed the prefill scheduler with:
  AttributeError: 'HiCacheController' object has no attribute 'tp_rank'

- cache_controller.py: resolve tp_rank/tp_size/dp_rank once in __init__ (always
  available, not storage-gated); drop the now-redundant derivation in
  _generate_storage_config.
- hiradix_cache.py: label the CP HiCache metrics by cp_rank/cp_size -- the
  identity that actually distinguishes the per-rank L2 lanes / L3 slabs these
  metrics measure -- resolved from the CP layout (_cp_hicache_cp_*), alongside
  the now-reliable tp/dp/pp coordinates.
- metrics_collector.py: docstring -> per-rank via cp_rank.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

python/sglang/srt/managers/cache_controller.py

@@ -427,6 +427,19 @@ class HiCacheController:
         self.pp_size = pp_size
         self.enable_storage_metrics = enable_storage_metrics
 
+        # Resolve this rank's parallel identity once, unconditionally -- not only
+        # when a storage backend is configured. CP shared-L2 attaches no storage
+        # backend, yet per-rank consumers (e.g. the CP HiCache metrics collector)
+        # still need tp_rank/tp_size/dp_rank.
+        if is_dp_attention_enabled():
+            self.tp_rank = get_attention_tp_rank()
+            self.tp_size = get_attention_tp_size()
+            self.dp_rank = get_attention_dp_rank()
+        else:
+            self.tp_rank = get_tensor_model_parallel_rank()
+            self.tp_size = get_tensor_model_parallel_world_size()
+            self.dp_rank = 0
+
         # Default storage page IO functions (may be overridden by attach).
         self.page_get_func = self._generic_page_get
         self.page_set_func = self._generic_page_set
@@ -788,15 +801,8 @@ class HiCacheController:
         if storage_backend_extra_config is None:
             storage_backend_extra_config = {}
 
-        if is_dp_attention_enabled():
-            self.tp_rank = get_attention_tp_rank()
-            self.tp_size = get_attention_tp_size()
-            self.dp_rank = get_attention_dp_rank()
-        else:
-            self.tp_rank = get_tensor_model_parallel_rank()
-            self.tp_size = get_tensor_model_parallel_world_size()
-            self.dp_rank = 0
-
+        # tp_rank/tp_size/dp_rank are resolved once in __init__ (always available,
+        # including under CP shared-L2 where no storage backend is attached).
         # Currently, NPUMLATokenToKVPool is the subclass of MLATokenToKVPool.
         is_mla_backend = isinstance(self.mem_pool_device, MLATokenToKVPool)
         # Least Common Multiple among heterogeneous tp size

python/sglang/srt/mem_cache/hiradix_cache.py

@@ -3910,7 +3910,13 @@ class HiRadixCache(RadixCache):
             from sglang.srt.observability.metrics_collector import CpHiCacheMetricsCollector
 
             cc = self.cache_controller
+            # cp_rank/cp_size are the identity that actually distinguishes the
+            # per-rank L2 lanes / L3 slabs (what these metrics measure); tp/dp/pp
+            # give the full parallel coordinate. cp_* come from the CP layout
+            # (_cp_hicache_cp_*), the rest from the controller (always set).
             labels = {
+                "cp_rank": self._cp_hicache_cp_rank(),
+                "cp_size": self._cp_hicache_cp_size(),
                 "tp_rank": cc.tp_rank, "dp_rank": cc.dp_rank,
                 "pp_rank": cc.pp_rank, "pp_size": cc.pp_size,
             }

python/sglang/srt/observability/metrics_collector.py

@@ -1464,7 +1464,8 @@ class CpHiCacheMetricsCollector:
 
     Gauges for instantaneous state (occupancy, inflight, queue depths); Counters for cumulative work
     (spill/reload pages, GC reclaims, write errors, L2 commit/evict). Fed periodically off the allocator's
-    occupancy_by_payload()/stats() + CpL3Store.stats(). Per-rank via the tp_rank label. The mainline
+    occupancy_by_payload()/stats() + CpL3Store.stats(). Per-rank via the cp_rank label (each CP rank owns
+    its own L2 lane + L3 slabs); tp/dp/pp labels give the full parallel coordinate. The mainline
     StorageMetricsCollector is dead under CP (enable_storage=False), so this is the only L2/L3 visibility.
     """