Revert "[HiCache] Enable spec v2 + decode KV cache offloading compatibility" (#19613)

python/sglang/srt/disaggregation/decode_kvcache_offload_manager.py

@@ -20,7 +20,6 @@ from sglang.srt.mem_cache.memory_pool_host import (
     MLATokenToKVPoolHost,
 )
 from sglang.srt.server_args import ServerArgs
-from sglang.srt.utils.common import ceil_align
 
 if TYPE_CHECKING:
     from sglang.srt.managers.schedule_batch import Req
@@ -192,18 +191,6 @@ class DecodeKVCacheOffloadManager:
 
         # Free the incremental part of the request
         self.token_to_kv_pool_allocator.free(kv_indices)
-
-        # Free over-allocated KV cache slots (e.g. from speculative decoding v2).
-        # Without spec v2, start_p == end_p so this is a no-op.
-        start_p, end_p = req.pop_overallocated_kv_cache()
-        if self.page_size > 1:
-            start_p = ceil_align(start_p, self.page_size)
-        if start_p < end_p:
-            overalloc_indices = self.req_to_token_pool.req_to_token[
-                req.req_pool_idx, start_p:end_p
-            ]
-            self.token_to_kv_pool_allocator.free(overalloc_indices)
-
         self.req_to_token_pool.free(req)
         self.tree_cache.protected_size_ -= len(req.prefix_indices)
 

python/sglang/srt/server_args.py

@@ -2748,6 +2748,13 @@ class ServerArgs:
                 raise ValueError(
                     "The argument disaggregation-decode-enable-offload-kvcache is only supported for decode side."
                 )
+            if (
+                self.disaggregation_mode == "decode"
+                and envs.SGLANG_ENABLE_SPEC_V2.get()
+            ):
+                raise ValueError(
+                    "Spec v2 and decode offload kv cache are incompatible and cannot be enabled together."
+                )
         if not (0 < self.swa_full_tokens_ratio <= 1.0):
             raise ValueError("--swa-full-tokens-ratio should be in range (0, 1.0].")