Incorrect JSON format during Pytorch Execution Trace generation

[arjuntemura]

Bug Description

I am running a distributed Linear model (20 parameters) across 2 GPU Nodes, each node having 2 NVIDIA H100 NVL GPUs. The Model uses DDP parallelization strategy. I am generating the Pytorch ET (in json format) trace using the ExecutionTraceObserver() as mentioned in the instructions. I observe that the trace has many syntactical errors. Also, many nodes have incomplete data (images attached)
I tried this with the latest Pytorch version (2.5.0) as well but encountered the same problem.

Steps to Reproduce

Code used for distributed training: https://github.com/pytorch/examples/tree/main/distributed/ddp-tutorial-series Command to run across both nodes: torchrun --nproc_per_node=2 --nnodes=2 --node_rank=0 --rdzv_id=456 --rdzv-backend=c10d --rdzv_endpoint=<ip:port> <code>.py <no. of epochs> <epochs after which result will be saved> I am capturing the ET trace for one epoch.

Information for one GPU Node (Both nodes have the same configuration): Pytorch: 2.1.2 , 2.5.1 (tried both) OS: Linux Kernel version: 5.15.0-124-generic Ubuntu Version: Ubuntu 22.04.5 No. of CPUs : 64 CPU Architecture: x86_64 CPU op-mode(s): 32-bit, 64-bit CPU Address sizes: 52 bits physical, 57 bits virtual CPU Byte Order: Little Endian Memory: 503Gi No. of GPUs: 2 GPU Memory (each GPU): 95830MiB

I would be obliged if someone could help in this regard.

Screenshots

[wkaisertexas]

Hey, try collecting smaller traces from pytorch's execution trace observer. Try collecting one iteration and you get valid json.

This is a pretty annoying quirk of the execution trace observer serializer

[arjuntemura]

I did try this for a single epoch though.

[9LLPPLL6]

I did try this for a single epoch though.我确实尝试过一次这样的尝试。

hello, Have you solved this problem now

[arjuntemura]

I did. I generated seperate Execution traces for each GPU. Seems that there was some kind of race condition happening between both the GPUs writing to the same execution file.