Comparison of performance between source‑built and pip‑installed MuJoCo 3.3.4

[kankanzheli]

Intro

Hi MuJoCo team and community,

I’m investigating performance differences between building MuJoCo from source and installing the official wheel (via pip), both at version 3.3.4, under Conda environments. I’m seeing significantly different CPU usage and performance in RL training scenarios.

My setup

mujoco 3.3.4
python 3.10
ubuntu 22.04
torch 2.5.1

My question

✅ Test 1 – specs_test.py

• pip-installed MuJoCo: Ran 32 tests in 0.031s
• source-built MuJoCo: Ran 32 tests in 0.047s

This suggests pip-installed version is ~50% faster on basic tests.

✅ Test 2 – My custom multi-process RL code (2 environments)

Using htop, I observed:

• pip version: each CPU core < 20% usage
• source build: each core > 60%, multiple over 70%

🔍 Questions

1. What could cause this performance gap?
2. Is it possible to achieve comparable performance by re-building from source with appropriate flags?
3. Could you please share the full set of compiler flags used in the official wheel build?

I want to confirm if rebuilding with identical build configuration will allow my source-built version to match the performance of the official wheel.

Thanks for any advice or pointers!

Minimal model and/or code that explain my question

No response

Confirmations

• I searched the latest documentation thoroughly before posting.
• I searched previous Issues and Discussions, I am certain this has not been raised before.