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, 00:45 UTC
Make vllm compatible with verl (#12824)
Co-authored-by: zhangshulai <zhangshulai@bytedance.com>
ZSL98 committed
Feb 7, 2025 at 11:54 UTC
433c4a49230a470f13657f06e7612cde86e4fb40
2 files changed
+1
-8
vllm/distributed/parallel_state.py
-7
@@ -1024,13 +1024,6 @@ def initialize_model_parallel(
1024
backend = backend or torch.distributed.get_backend(
1025
get_world_group().device_group)
1026
1027
- if (world_size
1028
- != tensor_model_parallel_size * pipeline_model_parallel_size):
1029
- raise RuntimeError(
1030
- f"world_size ({world_size}) is not equal to "
1031
- f"tensor_model_parallel_size ({tensor_model_parallel_size}) x "
1032
- f"pipeline_model_parallel_size ({pipeline_model_parallel_size})")
1033
-
1027
# Build the tensor model-parallel groups.
1028
num_tensor_model_parallel_groups: int = (world_size //
1029
tensor_model_parallel_size)
vllm/executor/uniproc_executor.py
+1
-1
@@ -101,7 +101,7 @@ class ExecutorWithExternalLauncher(UniProcExecutor):
101
# - MASTER_PORT
102
distributed_init_method = "env://"
103
rank = int(os.environ["RANK"])
104
- local_rank = rank
104
+ local_rank = int(os.environ["LOCAL_RANK"])
105
is_driver_worker = True
106
kwargs = dict(
107
vllm_config=self.vllm_config,