What it does
Core capabilities at a glance
- Cuda
- Deepep
- Expert Parallelism
- Hopper
- LLM Inference
- Mixture OF Experts
- Nccl
- Nvlink
Deep dive
The full breakdown - performance, comparisons, and setup
afdEP-commLink to this section
afdEP-comm is a local inference server - Expert-parallel communication for Attention–FFN Disaggregation (AFD), built on DeepEP V2.
OverviewLink to this section
Expert-parallel communication for Attention–FFN Disaggregation (AFD).
afdEP-comm connects N Attention GPUs, which route tokens, with M FFN GPUs, which hold the experts: it dispatches routed tokens to the experts and combines their outputs back at the source. It vendors DeepEP V2's kernels and specializes them for the two roles. Routing, quantization, expert compute and serving scheduling stay with the caller.
Every result here is measured against FastAFD's DeepEP M2N, the way FastAFD runs AFD over DeepEP V2 ('DeepEPM2NAdapter'). DeepEP sees one symmetric expert-parallel group: Attention ranks hold dummy expert slots that no token routes to, and every rank sends, receives and waits in both phases. The AFD plugin served the same code in vLLM through its 'deepep_m2n' connector.
afdEP-comm runs the same DeepEP V2 kernels and removes only what the two roles make unnecessary: Attention GPUs only send in dispatch and only receive in combine, and FFN GPUs the reverse. A change that would speed up FastAFD's DeepEP M2N as well is out of scope. The benchmarks run the baseline over the same vendored kernels and harness (benchmarks/fastafd.py). Every change keeps an environment control that reverts it for measurement, except the sparse directional hybrid, which is part of the cross-node protocol from 1.1.0.
afdEP-comm is open-source, written primarily in Python, with 94 GitHub stars under the Apache 2.0 license. The latest release is v1.2.0 (2026-11-16).
Key capabilitiesLink to this section
From the project's documentation:
- receive buffers sized per role, 50% of the union registration at 2A2F and about 75%
- each role launches only the warps it uses, and the FFN starts its copy epilogue at the
- with one FFN rank, a single-FFN reduction and a zero-copy collect view
- a directional protocol when every domain holds one role: no cross-domain barriers,
- in that protocol, signals only on the channels that carry tokens: sources start on
- domains that hold both roles (FFN GPUs sharing a node with Attention GPUs) run
InstallLink to this section
A quick way to get started (always check the official docs for the latest):
pip install .How it fits a local-AI stackLink to this section
afdEP-comm runs on your own hardware, so pair it with a model and a GPU sized to your needs. Use the VRAM calculator to pick a model that fits your card, and see what you can run for hardware guidance. Related local inference servers in the directory:
SourcesLink to this section
- Source code & docs: afdEP/afdEP-comm
- Official website: https://afdep.github.io/afdEP-comm/
Stats from GitHub, 2026-11-19.
Frequently asked
Quick answers to common questions
What is afdEP-comm?
afdEP-comm is a inference-server tool for local AI workloads. Expert-parallel communication for Attention–FFN Disaggregation (AFD), built on DeepEP V2
Is afdEP-comm free and open source?
Yes, afdEP-comm has 94 GitHub stars and is licensed under Apache 2.0. You can self-host it for free on .
What hardware do I need for afdEP-comm?
The hardware requirements depend on which models you run. Check our hardware directory for compatible GPUs and systems. afdEP-comm has 94 GitHub stars and an active community.
Does afdEP-comm support GPU acceleration?
afdEP-comm's GPU support depends on your specific setup. Check the documentation for details. For the best performance, pair it with an NVIDIA RTX 4090 or 5090.
What are the best alternatives to afdEP-comm?
Popular alternatives include other inference-server tools in our directory. Browse our full collection at /tool for comparisons, community reviews, and benchmark data to find the right fit for your workflow.
How much does afdEP-comm cost?
afdEP-comm is free-open-source. It is completely free and open source to self-host.
Pairs well with
Complementary tools, models, and hardware