MLServer installation fails on macOS due to tritonclient 2.61+ perf-analyzer dependency

[yesid-lopez]

Problem

MLServer 1.7.1 cannot be installed on macOS when poetry/pip resolves to tritonclient >= 2.61.0, because tritonclient 2.61+ introduced perf-analyzer as a hard dependency, which only provides Linux wheels (no macOS support).

Error

RuntimeError: Unable to find installation candidates for perf-analyzer (2.59.1)

Root Cause

• MLServer requires: tritonclient >= 2.42 with [http] extras
• tritonclient 2.61.0+ added perf-analyzer as a mandatory dependency
• perf-analyzer only distributes manylinux wheels (no macOS/Darwin wheels available)
• This blocks all macOS users from installing MLServer with default dependency resolution

Impact

• Prevents local development on macOS
• Forces developers to use Docker/Linux VMs for basic MLServer development
• Affects the developer experience for a significant portion of the ML community

Current Workaround

Pin tritonclient to pre-2.61 versions:

tritonclient = {version = ">=2.42,<2.61", extras = ["http"]}

Environment

• OS: macOS (Darwin)
• MLServer version: 1.7.1
• tritonclient version: 2.61.0
• Python: 3.11

[yesid-lopez]

Sharing in case someone else gets the same issue installing mlserver on mac

[yesid-lopez]

Considering mlserver is already being constrained due to issues installating triton on mac:

# add a min version to tritonclient due to https://github.com/triton-inference-server/server/issues/6246
tritonclient = {version = ">=2.42", extras = ["http"]}

does it make sense add the >=2.42,<2.61 constrain?

If it sounds good to you I can create a PR

[dtpryce]

@yesid-lopez we have also experienced this. If you can find a fix please do! We were installing mlserver 1.7.1 in python 3.12 using pdm. Same error - just testing if using a docker image (linux based) will fix for local development for now.

[dtpryce]

And can confirm that using docker to install very same requirements works fine from a Mac host machine.

[dtpryce]

And here is the discussion on getting that requirement removed or made extra in tritonclient https://github.com/triton-inference-server/client/issues/856

[dtpryce]

And an open PR to remove it entirely https://github.com/triton-inference-server/client/pull/857