apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[BugFix] fix multinode deployment (#2977)

Browse Source
This commit is contained in:
ltd0924
2025-07-24 15:04:04 +08:00
committed by GitHub
parent 3792345c3a
commit f935d6f862
9 changed files with 71 additions and 81 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
fastdeploy/worker/gpu_worker.py
View File

@@ -100,13 +100,14 @@ class GpuWorker(WorkerBase):
# 1. Record memory state before profile run
start_time = time.perf_counter()
Gb = 1024**3
paddle.device.cuda.reset_max_memory_reserved(self.local_rank)
paddle.device.cuda.reset_max_memory_allocated(self.local_rank)
paddle_reserved_mem_before_run = paddle.device.cuda.max_memory_reserved(self.local_rank)
paddle_allocated_mem_before_run = paddle.device.cuda.max_memory_allocated(self.local_rank) # not reserved
local_rank = self.local_rank % self.max_chips_per_node
paddle.device.cuda.reset_max_memory_reserved(local_rank)
paddle.device.cuda.reset_max_memory_allocated(local_rank)
paddle_reserved_mem_before_run = paddle.device.cuda.max_memory_reserved(local_rank)
paddle_allocated_mem_before_run = paddle.device.cuda.max_memory_allocated(local_rank) # not reserved
pynvml.nvmlInit()
handle = pynvml.nvmlDeviceGetHandleByIndex(int(self.device_ids[self.local_rank]))
handle = pynvml.nvmlDeviceGetHandleByIndex(int(self.device_ids[local_rank]))
before_run_meminfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
logger.info(
@@ -124,8 +125,8 @@ class GpuWorker(WorkerBase):
self.model_runner.profile_run()
# 3. Statistical memory information
paddle_reserved_mem_after_run = paddle.device.cuda.max_memory_reserved(self.local_rank)
paddle_allocated_mem_after_run = paddle.device.cuda.max_memory_allocated(self.local_rank)
paddle_reserved_mem_after_run = paddle.device.cuda.max_memory_reserved(local_rank)
paddle_allocated_mem_after_run = paddle.device.cuda.max_memory_allocated(local_rank)
model_block_memory_used = self.cal_theortical_kvcache()
paddle_peak_increase = paddle_reserved_mem_after_run - paddle_allocated_mem_before_run