facebookresearch/HolisticTraceAnalysis

Got NaN for communication time, looking for ncclKernel in the trace

Ouverte

#247 ouverte le 8 mai 2025

 (1 commentaire) (0 réaction) (0 personne assignée)Python (94 forks)auto 404
help wantedneeds triagequestion

Métriques du dépôt

Stars
 (535 étoiles)
Métriques de merge PR
 (Métriques PR en attente)

Description

What is your question?

I'm using HTA to analyze my tracem which is got from the pytorch profiler. However I got NaN for the communication time, and I dive into the code and I found in the communication_analysis.py, the communication kernel in my trace is 0 (which is not the case since I'm doing multi-gpu training). Later I found HTA is parsing the communication kernel using: NCCL_KERNEL_RE = re.compile(r"^nccl.*Kernel") in utils.py. However, I found my trace only have nccl::Allreduce as the name after I parse it. Is there a specific parameter I should set to let the pytorch profiler profile the kernel-level detail, or I made some mistake in writing the training script, so the nccl kernel is never being called?

Code

/home/yu/.local/lib/python3.8/site-packages/hta/analyzers/communication_analysis.py:73: RuntimeWarning: invalid value encountered in scalar divide return (shifted_overlap["time_y"] - shifted_overlap["time_x"]).sum() / (

What have you tried?

No response

Environment

Python: 3.10.16 Pytorch: 2.7 nccl/2.19.3-1.awsplugin cuda/11.8.0

Guide contributeur