Fix: fix the exception 'the memory capacity is unbalanced. Some GPUs … (#5426)

Co-authored-by: ocss884 <ocss.lin@gmail.com>
This commit is contained in:
mlmz
2025-04-18 10:51:39 +08:00
committed by GitHub
parent 211c7b31b8
commit 27e9538a7e

View File

@@ -73,6 +73,7 @@ from sglang.srt.utils import (
MultiprocessingSerializer,
enable_show_time_cost,
get_available_gpu_memory,
get_bool_env_var,
init_custom_process_group,
is_cuda,
is_fa3_default_architecture,
@@ -378,10 +379,16 @@ class ModelRunner:
local_gpu_memory = get_available_gpu_memory(self.device, self.gpu_id)
if self.tp_size > 1:
if min_per_gpu_memory < local_gpu_memory * 0.9:
raise ValueError(
"The memory capacity is unbalanced. Some GPUs may be occupied by other processes. "
f"{min_per_gpu_memory=}, {local_gpu_memory=}, {local_gpu_memory * 0.9=}"
)
if get_bool_env_var("SGL_DISABLE_TP_MEMORY_INBALANCE_CHECK"):
logger.warning(
"The memory capacity is unbalanced. Some GPUs may be occupied by other processes. "
f"{min_per_gpu_memory=}, {local_gpu_memory=}, {local_gpu_memory * 0.9=}"
)
else:
raise ValueError(
"The memory capacity is unbalanced. Some GPUs may be occupied by other processes. "
f"{min_per_gpu_memory=}, {local_gpu_memory=}, {local_gpu_memory * 0.9=}"
)
logger.info(
f"Init torch distributed ends. mem usage={(before_avail_memory - local_gpu_memory):.2f} GB"