From 5b8c64dc77394bbfb827bc1455d2054166b903fd Mon Sep 17 00:00:00 2001 From: "rshaw@neuralmagic.com" Date: Thu, 19 Jun 2025 13:12:43 +0000 Subject: [PATCH] updated Signed-off-by: rshaw@neuralmagic.com --- .../kv_connector/v1/nixl_connector.py | 18 +++++++++--------- 1 file changed, 9 insertions(+), 9 deletions(-) diff --git a/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py b/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py index d3da22546bfb5..053901c1cc42d 100644 --- a/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py +++ b/vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py @@ -921,14 +921,14 @@ class NixlConnectorWorker: # corresponding rank. With heterogeneous TP, fixing D>P, the D tp # workers will issue xfers to parts of the P worker remote kv caches. - # Sort block ids to ensure nixl can merge contiguous blocks. - start = time.perf_counter() - sorted_idx = sorted(range(len(local_block_ids)), - key=local_block_ids.__getitem__) - local_block_ids = [local_block_ids[i] for i in sorted_idx] - remote_block_ids = [remote_block_ids[i] for i in sorted_idx] - end = time.perf_counter() - print(f"REORDER took: {end - start}") + # # Sort block ids to ensure nixl can merge contiguous blocks. + # start = time.perf_counter() + # sorted_idx = sorted(range(len(local_block_ids)), + # key=local_block_ids.__getitem__) + # local_block_ids = [local_block_ids[i] for i in sorted_idx] + # remote_block_ids = [remote_block_ids[i] for i in sorted_idx] + # end = time.perf_counter() + # print(f"REORDER took: {end - start}") # Get descs ids. local_block_descs_ids: list[int] = [] @@ -974,7 +974,7 @@ class NixlConnectorWorker: remote_xfer_side_handle, remote_block_descs_ids, notif_msg=notif_id, - # skip_desc_merge=True, + skip_desc_merge=True, # this causes the issue to emerge immediately. ) # Begin async xfer.