[UnifiedTree] gate load back pre-evict on full-attn availability only (#26302)

Signed-off-by: Vladislav Nosivskoy <vladnosiv@gmail.com>
This commit is contained in:
Vladislav Nosivskoy
2026-05-29 00:30:03 +08:00
committed by GitHub
parent e33bbbb467
commit 34ea682a07
2 changed files with 91 additions and 1 deletions
@@ -1419,7 +1419,10 @@ class UnifiedRadixCache(KVCacheEventMixin, BasePrefixCache):
self.dec_lock_ref(best_match_node, ancestor_lock_params)
return False
avail = self.token_to_kv_pool_allocator.available_size()
if self.supports_swa():
avail = self.token_to_kv_pool_allocator.full_available_size()
else:
avail = self.token_to_kv_pool_allocator.available_size()
if avail < kv_tokens:
needed = kv_tokens - avail
result = self.evict(EvictParams(num_tokens=needed))