Skip to content

Request: SDK 2.2 release of the Frigate runtime bundle #14

Description

@JoshuaSeidel

Frigate installs this bundle into its container image by pinning tag v2.1.0 (see docker/main/install_memryx.sh). SDK 2.2 is now the current release, but there is no 2.2 tag here — main still ships memryx/x86/libmemx.so.2.1.0 and was last updated 2025-11-13.

What is already available for 2.2:

  • apt (https://developer.memryx.com/deb): memx-drivers 2.2.4-1, memx-accl 2.2.5-1, mxa-manager 2.2.5-1, for both amd64 and arm64
  • docs: SDK 2.2 is the unversioned developer.memryx.com root; 2.1 is archived under /2p1/

The memryx wheel from developer.memryx.com/pip is not a drop-in substitute for this bundle inside a container. Its package __init__ imports memryx.neural_compiler, so a plain --no-deps install fails with ModuleNotFoundError: No module named 'pyximport', and satisfying it fully would pull in torch/tensorflow/PySide6. Removing the neural_compiler and gui trees and substituting a runtime-only __init__ that imports mxa/mxapi alongside Dfp/*Accl does work (99M -> 31M, AsyncAccl connects to mxa-manager and a 2p0-era DFP parses), but that is vendor-package surgery that does not belong in Frigate's Dockerfile.

Request: tag a 2.2 bundle in the same shape as v2.1.0 — runtime-only memryx package plus libmemx.so/libmx_accl.so for cp311/cp312 on x86_64 and aarch64.

For context on urgency: this is not currently breaking anything. Per memx/accl/messages.h in memx-accl 2.2.5, SDK 2.0/2.1 clients speak PROTO_1, SDK 2.2 speaks PROTO_2, and a 2.2 mxa-manager accepts both [1,2] — so the existing 2.1 in-container client works against a 2.2 host. A 2.2 bundle is needed to move the container side forward, not to keep it working.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions