No project description provided
Project description
DeltaFabric
A decentralized weight synchronization protocol for distributed ML training with built-in networking. Currently supporting PyTorch and Burn.
Core Concepts
No-Master Architecture
Eliminates master-server bottlenecks by letting every node participate equally in synchronization. It also makes life easier by cutting out the need to set up or manage a central server.
Built-in Networking and Automatic Peer Discovery
We include a built-in networking layer powered by the Zenoh P2P protocol. Nodes discover peers and exchange deltas automatically without external coordination.
Anchor-Active Weight Model
We maintain two weight buffers per node:
- W_active: weights currently being optimized by local gradients
- W_anchor: weights representing the last acknowledged state
The delta between these buffers captures the local updates discovered since the last synchronization round. This two-buffer design eliminates the need to store copies of every peer's model, reducing memory overhead compared to traditional decentralized approaches.
Compressed Delta
Sending an entire set of weights becomes impractically large as the model grows. To reduce bandwidth and increase speed, we only send the most significant changes (deltas) per step. This allows a single packet to hold not only one delta but also others relayed from additional peers.
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distributions
Built Distribution
Filter files by name, interpreter, ABI, and platform.
If you're not sure about the file name format, learn more about wheel file names.
Copy a direct link to the current filters
File details
Details for the file delta_fabric-0.6.0-cp313-cp313-manylinux_2_34_x86_64.whl.
File metadata
- Download URL: delta_fabric-0.6.0-cp313-cp313-manylinux_2_34_x86_64.whl
- Upload date:
- Size: 7.7 MB
- Tags: CPython 3.13, manylinux: glibc 2.34+ x86-64
- Uploaded using Trusted Publishing? No
- Uploaded via: uv/0.11.0 {"installer":{"name":"uv","version":"0.11.0","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"Fedora Linux","version":"43","id":"","libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}
File hashes
| Algorithm | Hash digest | |
|---|---|---|
| SHA256 |
671a7fb0aa2ed0c00a824329a44dacf6cfdd87b1d95dddef1a383b1ce194a448
|
|
| MD5 |
8cf6542f55334e893499ec0e34681627
|
|
| BLAKE2b-256 |
e11065bb0c449e8c04f35fa9a292b51e10b975fd95be1edb9a52d9943efa1050
|
