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 6, 19:15 UTC
[BUG] Fix rank calculation in NCCLWeightTransferEngine (#36940)
Signed-off-by: hao-aaron <ahao@anyscale.com>
Aaron Hao committed
Mar 12, 2026 at 18:56 UTC
5e1a373d2e62c04ba464c88303600839d6973365
1 file changed
+1
-1
vllm/distributed/weight_transfer/nccl_engine.py
+1
-1
@@ -132,7 +132,7 @@ class NCCLWeightTransferEngine(
132
133
# Calculate the global rank in the trainer-worker process group
134
# Must account for data parallel to get unique ranks across all workers
135
- dp_rank = self.parallel_config.data_parallel_rank
135
+ dp_rank = self.parallel_config.data_parallel_index
136
world_size_per_dp = self.parallel_config.world_size # TP * PP
137
rank_within_dp = self.parallel_config.rank
138