Clarify post-capture KV reservation logs (#33598)

python/sglang/srt/mem_cache/kv_vmm_backing.py

@@ -152,7 +152,8 @@ class KvVmmArena:
         # no_split so the caching allocator hands our bump pointers back verbatim.
         self.pool = torch.cuda.MemPool(self._allocator, no_split=True)
         logger.info(
-            "KvVmmArena[%s] ready: device=%d reserved=%.1f GiB granularity=%d KiB",
+            "KvVmmArena[%s] ready: device=%d reserved_va=%.1f GiB "
+            "granularity=%d KiB",
             self._sfx,
             self.device_id,
             self.reserved / (1024**3),

python/sglang/srt/mem_cache/memory_pool.py

@@ -1667,7 +1667,7 @@ class KVCache(abc.ABC):
             k_size_GB = k_size / GB
             v_size_GB = v_size / GB
             logger.info(
-                f"{cache_name} is allocated. dtype: {self.dtype}, "
+                f"{cache_name} {'VA upper bound' if self.post_capture_active else 'is allocated'}. dtype: {self.dtype}, "
                 f"#tokens: {num_tokens}, K size: {k_size_GB:.2f} GB, "
                 f"V size: {v_size_GB:.2f} GB"
             )
@@ -1675,7 +1675,7 @@ class KVCache(abc.ABC):
         else:
             kv_size_GB = kv_size_bytes / GB
             logger.info(
-                f"{cache_name} is allocated. dtype: {self.dtype}, "
+                f"{cache_name} {'VA upper bound' if self.post_capture_active else 'is allocated'}. dtype: {self.dtype}, "
                 f"#tokens: {num_tokens}, KV size: {kv_size_GB:.2f} GB"
             )
             self.mem_usage = kv_size_GB

python/sglang/srt/mem_cache/swa_memory_pool.py

@@ -86,7 +86,7 @@ class SWAKVPool(BaseSWAKVPool):
         k_size, v_size = self.get_kv_size_bytes()
         self.mem_usage = (k_size + v_size) / GB
         logger.info(
-            f"SWAKVPool mem usage: {self.mem_usage:.2f} GB, swa size: {self.size_swa}, full size: {self.size}"
+            f"SWAKVPool {'VA upper bound' if self.post_capture_active else 'mem usage'}: {self.mem_usage:.2f} GB, swa size: {self.size_swa}, full size: {self.size}"
         )
 
     @property