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 8, 20:15 UTC
[Bugfix] Fix swapped engine_ids in NIXL Llama 4 local attention path (#33795)
Signed-off-by: Yoray Zack <yorayz@nvidia.com>
zackyoray committed
Feb 5, 2026 at 19:51 UTC
1ee95841bd251f9081c3a317984c4dcaa003b3c0
1 file changed
+3
-3
vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py
+3
-3
@@ -2323,15 +2323,15 @@ class NixlConnectorWorker:
2323
2324
# Get descs ids for the layer.
2325
layer_local_desc_ids = self._get_block_descs_ids(
2326
- dst_engine_id,
2326
+ self.engine_id,
2327
layer_local_block_ids,
2328
layer_idx,
2329
+ block_size_ratio=block_size_ratio,
2330
)
2331
layer_remote_desc_ids = self._get_block_descs_ids(
2331
- self.engine_id,
2332
+ dst_engine_id,
2333
layer_remote_block_ids,
2334
layer_idx,
2334
- block_size_ratio=block_size_ratio,
2335
)
2336
2337
local_descs_list.append(layer_local_desc_ids)