A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 23:45 UTC
[BugFix] Fix shared storage connector load kv only load attention layer (#21428)
Signed-off-by: David Chen <530634352@qq.com>
WeiQing Chen committed
Jul 26, 2025 at 22:07 UTC
97d6c30cc965e70579bfdad27e7514592752096e
1 file changed
+10
-2
vllm/distributed/kv_transfer/kv_connector/v1/shared_storage_connector.py
+10
-2
@@ -156,8 +156,16 @@ class SharedStorageConnector(KVConnectorBase_V1):
156
logger.info("Inject KV cache of %d tokens to the paged memory",
157
len(request.slot_mapping))
158
for layer_name in forward_context.no_compile_layers:
159
- attn_layer = forward_context.no_compile_layers[layer_name]
160
- kv_cache_layer = attn_layer.kv_cache[\
159
+ layer = forward_context.no_compile_layers[layer_name]
160
+
161
+ # Only process layers that have kv_cache
162
+ # attribute (attention layers) Skip non-attention
163
+ # layers like FusedMoE/MLP etc.
164
+ kv_cache_attr = getattr(layer, 'kv_cache', None)
165
+ if kv_cache_attr is None:
166
+ continue
167
+
168
+ kv_cache_layer = kv_cache_attr[ \
169
forward_context.virtual_engine]
170
171
filename = self._generate_filename_debug(