apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 17:17:14 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[LLM] fix multinode bugs (#2945)

Browse Source 
* [LLM] fix multinode bugs

* [LLM] fix multinode bugs

* [LLM] fix multinode bugs

* [LLM] fix ci bugs

* fix ci bugs

* fix ci bugs
This commit is contained in:
ltd0924
2025-07-22 20:23:37 +08:00
committed by GitHub
parent 69be77c8c0
commit b0f1e0eef4
9 changed files with 68 additions and 87 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
fastdeploy/worker/gpu_worker.py
View File

@@ -100,16 +100,17 @@ 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)
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(
self.local_rank)
local_rank)
paddle_allocated_mem_before_run = paddle.device.cuda.max_memory_allocated(
self.local_rank) # not reserved
local_rank) # not reserved
pynvml.nvmlInit()
handle = pynvml.nvmlDeviceGetHandleByIndex(
int(self.device_ids[self.local_rank]))
int(self.device_ids[local_rank]))
before_run_meminfo = pynvml.nvmlDeviceGetMemoryInfo(handle)
logger.info((
@@ -126,9 +127,9 @@ class GpuWorker(WorkerBase):
# 3. Statistical memory information
paddle_reserved_mem_after_run = paddle.device.cuda.max_memory_reserved(
self.local_rank)
local_rank)
paddle_allocated_mem_after_run = paddle.device.cuda.max_memory_allocated(
self.local_rank)
local_rank)
model_block_memory_used = self.cal_theortical_kvcache()
paddle_peak_increase = paddle_reserved_mem_after_run - paddle_allocated_mem_before_run