Comments (2)
My case is 4 gpu devices: 2 x tp2 inference engines in one process
from nccl.
I found same question here #195 (comment) (my case is launch the allreduce between GPU0 and GPU2 and GPU1 and GPU3 concurrently
)
Is it safe to launch concurrent allreduce on communicators with different GPUs? For example, lets say we launch the allreduce on all 4 GPUs and wait for it to complete. Then we launch the allreduce between GPU0 and GPU2 and GPU1 and GPU3 concurrently. Would this be safe since the GPUs used are distinct?
And answer is #195 (comment)
Yes, that is safe.
So if I don’t have dependencies between groups (0,2 GPU and 1,3 GPU), then I don’t need any synchronization at all, right?
from nccl.
Related Issues (20)
- 2 node 16 H20 GPU allreduce performance is not as expected with NVL sharp HOT 3
- Question about the use of tree algorithm in DGX A100 HOT 2
- AllReduce info.proto returns 0 (LL protocol) despite both setting env variable to SIMPLE / Tuning.cc protoEnable array disabling LL/LL128 HOT 5
- torch.distributed.DistBackendError: NCCL error in: /root/pytorch/torch/csrc/distributed/c10d/ProcessGroupNCCL.cpp:1333, internal error HOT 2
- Question: "NCCL needs all GPUs of a host to be part of a collective in other to reliably use NVLinks"?
- reduceCopy minSrcs and maxSrcs HOT 2
- all-reduce latency is lower with 2 nodes, 2 GPUs each vs. 2 nodes, 1 GPU each
- Leak in FIFO queue HOT 7
- Is this description of NCCL_BUFFSIZE's role appropriate? HOT 4
- nccl error for driver 535.161.08 HOT 2
- Question on Multi-Machine Topology Connection in NCCL HOT 1
- 'NCCLCommunicator' object has no attribute 'comm' HOT 4
- NCCL communication stuck with CPU usage 100% HOT 1
- Is intra-node latency supposed to increase with the number of GPUs when using NVLink SHARP? HOT 4
- When NCCL timeout happens, can the timeout rank print which ranks are not joining the collective? HOT 1
- LL128 communication protocol HOT 6
- Can two AllReduce operations run in parallel using different non-blocking cudaStreams?
- some question about log HOT 3
- Only ~783GByte/s out of theoretical 900GB/s HGX H100 SXM Nvlink4 HOT 2
- Program stuck when destroying NCCL
Recommend Projects
-
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
-
Vue.js
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
-
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
-
TensorFlow
An Open Source Machine Learning Framework for Everyone
-
Django
The Web framework for perfectionists with deadlines.
-
Laravel
A PHP framework for web artisans
-
D3
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
-
Recommend Topics
-
javascript
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
-
web
Some thing interesting about web. New door for the world.
-
server
A server is a program made to process requests and deliver data to clients.
-
Machine learning
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
-
Visualization
Some thing interesting about visualization, use data art
-
Game
Some thing interesting about game, make everyone happy.
Recommend Org
-
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
-
Microsoft
Open source projects and samples from Microsoft.
-
Google
Google ❤️ Open Source for everyone.
-
Alibaba
Alibaba Open Source for everyone
-
D3
Data-Driven Documents codes.
-
Tencent
China tencent open source team.
from nccl.