Skip to content

[None][doc] update docs for EPLB #9166

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 2 commits into from
Nov 14, 2025
Merged

[None][doc] update docs for EPLB #9166

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
25b9d95
update docs for EPLB
dongxuy04 Nov 14, 2025
4868d05
Update examples/wide_ep/README.md
dongxuy04 Nov 14, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
13 changes: 13 additions & 0 deletions examples/wide_ep/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,19 @@ For GB200 NVL72, to make sure that Multi-Node NVLink (MNNVL) is correctly setup,

For more information on NVIDIA IMEX service for NVLink networks, refer to https://docs.nvidia.com/multi-node-nvlink-systems/imex-guide/overview.html.

#### Coherent Driver-Based Memory Management (CDMM)

Starting from R580 Driver, [Coherent Driver-Based Memory Management (CDMM)](https://docs.nvidia.com/datacenter/tesla/tesla-release-notes-580-65-06/index.html#hardware-software-support) for GB200 platforms is introduced. With CDMM, the driver manages GPU memory instead of the OS. CDMM avoids OS onlining of the GPU memory and the exposing of the GPU memory as a NUMA node to the OS. In Wide-EP, online EPLB need host threads be able to access the GPU memory to do the weights update.

When CDMM mode is off, GPU memory are exposed as NUMA nodes, so no additional prerequisites is required.

When CDMM mode is on, GPU memory doesn't exist in NUMA nodes, in that case, if online EPLB is needed, [GDRCopy](https://github.com/NVIDIA/gdrcopy?tab=readme-ov-file#build-and-installation) needs to be installed.

When GDRCopy is installed and the kernel module is loaded, you should be able to see the device file `/dev/gdrdrv` and kernel module `gdrdrv` by `lsmod`. The device file needs to be mapped into the container.

* For docker, this can be done by adding a device mapping like `--device=/dev/gdrdrv:/dev/gdrdrv`.
* For slurm with enroot, `--container-mounts="/dev/gdrdrv:/dev/gdrdrv"` needs to be added when starting containers and environment variable `export ENROOT_ALLOW_DEV=yes` needs to be set.

### Configurations

An example yaml file to enable wide EP:
Expand Down