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 16, 03:15 UTC
[XPU] Make pp group initilized for pipeline-parallelism (#11648)
Signed-off-by: yisheng <yi.sheng@intel.com>
YiSheng5 committed
Jan 7, 2025 at 11:09 UTC
d93d2d74fd807a091add17c2065ee8869339f76a
1 file changed
+6
vllm/worker/xpu_worker.py
+6
@@ -11,6 +11,7 @@ import torch.distributed
11
from vllm.config import VllmConfig
12
from vllm.distributed import (ensure_model_parallel_initialized,
13
init_distributed_environment)
14
+from vllm.distributed.parallel_state import get_pp_group
15
from vllm.logger import init_logger
16
from vllm.model_executor import set_random_seed
17
from vllm.platforms import current_platform
@@ -176,3 +177,8 @@ class XPUWorker(LoraNotSupportedWorkerBase, Worker):
177
parallel_config.pipeline_parallel_size)
178
# global all_reduce needed for overall oneccl warm up
179
torch.distributed.all_reduce(torch.zeros(1).xpu())
180
+
181
+ if parallel_config.pipeline_parallel_size > 1:
182
+ # Add pp group init to avoid
183
+ # p2p communication as the first call
184
+ get_pp_group().all_reduce(torch.zeros(1).xpu())