$HOME/miniconda3 unless you explicitly need multi-user./opt) unless you manage permissions + modules.Standardize on:
conda-forge (primary)defaultsEnforce priority:
conda config --set channel_priority strict
basenanobot, llama-hf)Export environments:
conda env export > env.yml
Prefer --from-history for minimal spec:
conda env export --from-history
Install mamba (faster solver):
conda install -n base -c conda-forge mamba
conda update --all in productionUse:
conda-pack (binary relocation)env.yml recreation (preferred for clean rebuild)Miniconda is a minimal Python distribution providing:
conda package & environment managerSupported platforms:
curl -O https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
(Use aarch64 if ARM)
sha256sum Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh
$HOME/miniconda3conda initsource ~/.bashrc
conda update -n base -c defaults conda
conda config --add channels conda-forge
conda config --set channel_priority strict
rm -rf ~/miniconda3
rm -rf ~/.conda ~/.condarc ~/.continuum
Remove from shell config:
# edit ~/.bashrc and remove conda init block
conda create -n <env-name> python=3.11
conda activate <env-name>
conda create -n nanobot python=3.11
conda activate nanobot
# example dependencies (adjust to your stack)
conda install -c conda-forge requests fastapi uvicorn
conda create -n llama-hf python=3.11
conda activate llama-hf
Install core packages:
conda install -c conda-forge llama-cpp
conda install -c conda-forge llama-cpp-python
conda install -c conda-forge huggingface_hub
Verify:
llama-cli
hf
mkdir -p ~/models/qwen
cd ~/models/qwen
hf download HauhauCS/Qwen3.6-35B-A3B-Uncensored-HauhauCS-Aggressive \
--local-dir ./model
conda activate nanobot
conda update --all
conda activate llama-hf
conda update --all
conda update <package-name>
conda activate llama-hf
conda env export --from-history > llama-hf.yml
conda activate nanobot
conda env export --from-history > nanobot.yml
conda env create -f llama-hf.yml
conda env create -f nanobot.yml
conda install -c conda-forge conda-pack
conda activate llama-hf
conda-pack -o llama-hf.tar.gz
mkdir -p ~/envs/llama-hf
tar -xzf llama-hf.tar.gz -C ~/envs/llama-hf
~/envs/llama-hf/bin/conda-unpack
Never pollute base
Use mamba for speed:
mamba install <pkg>
Keep environments small and purpose-specific
Store models outside env (e.g. /data/models)
Use .condarc for consistent team setups
Mamba is a high-performance drop-in replacement for the conda package manager. It is designed to solve one of conda’s biggest operational pain points: slow dependency resolution.
dnf/zypper)In practice:
You keep your existing Miniconda setup — just replace conda install with mamba install.
Dependency solving is often 10–100× faster
Especially noticeable with:
conda-forge heavy usageSame syntax as conda:
mamba install numpy
mamba create -n test python=3.11
Install into base environment:
conda install -n base -c conda-forge mamba
mamba create -n llama-hf python=3.11
mamba install -c conda-forge llama-cpp huggingface_hub
mamba update --all
| Feature | conda | mamba |
|---|---|---|
| Language | Python | C++ |
| Solver speed | Slow | Very fast |
| Compatibility | Native | Fully compatible |
| Default in Miniconda | Yes | No |
Mamba does not replace conda entirely
conda still handles:
.condarc)Best practice:
Even lighter, standalone binary
No base environment required
Ideal for:
If you're working with:
llama.cpp, huggingface)👉 Mamba is effectively mandatory for efficiency on Ubuntu/RHEL systems.
If you want, I can refactor your refcard to use mamba-first workflows (recommended for your use case).