Merge branch 'v0.9.2-dev-main+mtp-zero' into 'v0.9.2-dev'

fix: 只有当kv block中不含有MTP的假数据时才会被cached，以修复cache_full_blocks同一个kv block保存两次的bug

See merge request dcutoolkit/deeplearing/vllm!331

vllm/v1/core/single_type_kv_cache_manager.py

@@ -10,7 +10,7 @@ from vllm.v1.core.kv_cache_utils import BlockHash, KVCacheBlock
 from vllm.v1.kv_cache_interface import (FullAttentionSpec, KVCacheSpec,
                                         MambaSpec, SlidingWindowSpec)
 from vllm.v1.request import Request
-
+from vllm import envs
 
 class SingleTypeKVCacheManager(ABC):
     """
@@ -141,7 +141,9 @@ class SingleTypeKVCacheManager(ABC):
         """
         num_cached_blocks = self.num_cached_block[request.request_id]
         num_full_blocks = num_tokens // self.block_size
-
+        if envs.VLLM_ZERO_OVERHEAD_ENHANCE:
+            if num_full_blocks > num_cached_blocks and num_tokens % self.block_size < len(request.spec_token_ids):
+                num_full_blocks -= 1
         self.block_pool.cache_full_blocks(
             request=request,
             blocks=self.req_to_blocks[request.request_id],