Tools I Depend On
mamba
Use mamba instead of conda. It’s significantly faster.
Cloning envs
Currently I have ~60 mamba envs on my main dev machines. >300GB on disk, but tough early lessons that overloading envs would always lead to dll/version hell. One recently learned pro-tip. You can use a template env to speed this up: create -n [new-env] --clone [baseml-env]
On my local workstation that has a reasonably fast PCIe 4.0 NVMe SSD (5GB/s+ sequential writes) it takes about 28 s of wall time to create a clone (18GB env, 17GB of it is PyTorch):
🐟 ❯ time mamba create -n test --clone baseml
Source: /home/lhl/mambaforge/envs/baseml
Destination: /home/lhl/mambaforge/envs/test
Packages: 25
Files: 26407
Downloading and Extracting Packages:
Downloading and Extracting Packages:
Preparing transaction: done
Verifying transaction: done
Executing transaction: done
To activate this environment, use
$ mamba activate test
To deactivate an active environment, use
$ mamba deactivate
________________________________________________________
Executed in 27.91 secs fish external
usr time 4.18 secs 934.00 micros 4.18 secs
sys time 13.33 secs 126.00 micros 13.33 secsWait, that’s not so fast you might say, however, even with all libs cached and using uv it takes almost 80 s (of just runtime mind you, no time running commands) to load even the most basic packages you’ll use almost everywhere:
🐟 ❯ time mamba create -n test2 python=3.12 -y
Looking for: ['python=3.12']
conda-forge/linux-64 Using cache
conda-forge/noarch Using cache
pkgs/main/linux-64 Using cache
pkgs/main/noarch Using cache
pkgs/r/linux-64 Using cache
pkgs/r/noarch Using cache
Transaction
Prefix: /home/lhl/mambaforge/envs/test2
Updating specs:
- python=3.12
Package Version Build Channel Size
──────────────────────────────────────────────────────────────────────────────
Install:
──────────────────────────────────────────────────────────────────────────────
+ ld_impl_linux-64 2.43 h712a8e2_2 conda-forge Cached
+ _libgcc_mutex 0.1 conda_forge conda-forge Cached
+ ca-certificates 2024.12.14 hbcca054_0 conda-forge Cached
+ libgomp 14.2.0 h77fa898_1 conda-forge Cached
+ _openmp_mutex 4.5 2_gnu conda-forge Cached
+ libgcc 14.2.0 h77fa898_1 conda-forge Cached
+ libzlib 1.3.1 hb9d3cd8_2 conda-forge Cached
+ liblzma 5.6.3 hb9d3cd8_1 conda-forge Cached
+ libgcc-ng 14.2.0 h69a702a_1 conda-forge Cached
+ openssl 3.4.0 hb9d3cd8_0 conda-forge Cached
+ libexpat 2.6.4 h5888daf_0 conda-forge Cached
+ libsqlite 3.47.2 hee588c1_0 conda-forge Cached
+ tk 8.6.13 noxft_h4845f30_101 conda-forge Cached
+ ncurses 6.5 he02047a_1 conda-forge Cached
+ libxcrypt 4.4.36 hd590300_1 conda-forge Cached
+ libffi 3.4.2 h7f98852_5 conda-forge Cached
+ bzip2 1.0.8 h4bc722e_7 conda-forge Cached
+ libuuid 2.38.1 h0b41bf4_0 conda-forge Cached
+ libnsl 2.0.1 hd590300_0 conda-forge Cached
+ readline 8.2 h8228510_1 conda-forge Cached
+ tzdata 2024b hc8b5060_0 conda-forge Cached
+ python 3.12.8 h9e4cc4f_1_cpython conda-forge Cached
+ wheel 0.45.1 pyhd8ed1ab_1 conda-forge Cached
+ setuptools 75.6.0 pyhff2d567_1 conda-forge Cached
+ pip 24.3.1 pyh8b19718_2 conda-forge 1MB
Summary:
Install: 25 packages
Total download: 1MB
──────────────────────────────────────────────────────────────────────────────
pip 1.2MB @ 12.3MB/s 0.1s
Downloading and Extracting Packages:
Preparing transaction: done
Verifying transaction: done
Executing transaction: done
To activate this environment, use
$ mamba activate test2
To deactivate an active environment, use
$ mamba deactivate
________________________________________________________
Executed in 5.49 secs fish external
usr time 4.11 secs 253.00 micros 4.11 secs
sys time 0.85 secs 957.00 micros 0.84 secs
🐟 ❯ mamba activate test2
🐟 ❯ time pip install uv
Collecting uv
Using cached uv-0.5.11-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl.metadata (11 kB)
Using cached uv-0.5.11-py3-none-manylinux_2_17_x86_64.manylinux2014_x86_64.whl (15.0 MB)
Installing collected packages: uv
Successfully installed uv-0.5.11
________________________________________________________
Executed in 602.39 millis fish external
usr time 453.32 millis 316.00 micros 453.00 millis
sys time 92.58 millis 43.00 micros 92.54 millis
🐟 ❯ time uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.2
Using Python 3.12.8 environment at: mambaforge/envs/test2
Resolved 15 packages in 4.24s
Prepared 14 packages in 1m 02s
Installed 14 packages in 267ms
+ filelock==3.13.1
+ fsspec==2024.2.0
+ jinja2==3.1.3
+ markupsafe==2.1.5
+ mpmath==1.3.0
+ networkx==3.2.1
+ numpy==1.26.3
+ pillow==10.2.0
+ pytorch-triton-rocm==3.1.0
+ sympy==1.13.1
+ torch==2.5.1+rocm6.2
+ torchaudio==2.5.1+rocm6.2
+ torchvision==0.20.1+rocm6.2
+ typing-extensions==4.9.0
________________________________________________________
Executed in 66.65 secs fish external
usr time 32.43 secs 279.00 micros 32.43 secs
sys time 12.96 secs 32.00 micros 12.96 secs
🐟 ❯ time uv pip install transformers
Using Python 3.12.8 environment at: mambaforge/envs/test2
Resolved 17 packages in 13ms
Prepared 8 packages in 441ms
Installed 13 packages in 46ms
+ certifi==2024.12.14
+ charset-normalizer==3.4.0
+ huggingface-hub==0.27.0
+ idna==3.10
+ packaging==24.2
+ pyyaml==6.0.2
+ regex==2024.11.6
+ requests==2.32.3
+ safetensors==0.4.5
+ tokenizers==0.21.0
+ tqdm==4.67.1
+ transformers==4.47.1
+ urllib3==2.2.3
________________________________________________________
Executed in 527.05 millis fish external
usr time 235.67 millis 518.00 micros 235.16 millis
sys time 365.12 millis 0.00 micros 365.12 millis
🐟 ❯ time uv pip install xformers
Using Python 3.12.8 environment at: mambaforge/envs/test2
Resolved 13 packages in 108ms
Prepared 1 package in 1.34s
Installed 1 package in 15ms
+ xformers==0.0.28.post3
________________________________________________________
Executed in 1.48 secs fish external
usr time 231.86 millis 364.00 micros 231.49 millis
sys time 111.66 millis 42.00 micros 111.61 millis
🐟 ❯ time uv pip install triton
Using Python 3.12.8 environment at: mambaforge/envs/test2
Resolved 2 packages in 43ms
Prepared 1 package in 3.19s
Installed 1 package in 514ms
+ triton==3.1.0
________________________________________________________
Executed in 3.76 secs fish external
usr time 1.43 secs 0.00 micros 1.43 secs
sys time 0.75 secs 576.00 micros 0.75 secsNOTE: One important note is that when you --clone a new env, it doesn’t move over your environment variables. If you assign GPU specific stuff for example, you might want to either:
# export env to file, do what you want
conda env export -n baseml > environment.yml
# pipe template env direct to new_env
conda env export -n baseml | mamba env update -n new_env -f -uv
Implicit in that first example, but basically, I use mamba/conda since it will manage CUDA/GCC and other system libraries as well, but for Python libs, I basically stick completely to pip from within the mamba env. Or more specifically, nowadays I always first have uv as the first thing installed (pip install uv if it’s not in your --clone) and then run uv pip almost anytime I’d normally run pip.
I haven’t aliased that yet since it will probably still get me into trouble (I have upip though), but you can broadly think of it as on the order of how mamba improves conda performance.
byobu
I’m a long-time user of byobu, a wrapper against traditional tmux/screen terminal multiplexers. These are a must IMO for allowing easy attachment/detachment of long-running adhoc sessions.
Byobu has a bunch of extra quality of life additions, but I also add a few things to help. I have a custom function b that will also try to load up the mamba env if I’m trying to startup/reconnect to a named session. In ~/.bashrc:
# byobu
function b() {
if [ -z "$1" ]; then
byobu list-session
return 1
fi
byobu list-sessions | grep -q "$1" && byobu attach-session -t "$1" || byobu new-session -s "$1"
}
# Auto mamba env in tmux
if [ -n "$TMUX" ]; then
# Get the current tmux session name
tmux_session_name=$(tmux display-message -p '#S')
# Attempt to activate the Mamba environment with the same name as the tmux session
mamba activate $tmux_session_name 2>/dev/null
fior in ~/.config/fish/config.fish:
function b
if test -z "$argv[1]"
byobu list-session
return 1
end
if byobu list-sessions | grep -q "$argv[1]"
byobu attach-session -t "$argv[1]"
else
byobu new-session -s "$argv[1]"
end
end
# Try to start mamba env of same name if in tmux
# Check if inside a TMUX session
if test -n "$TMUX"
# Get the current tmux session name
set tmux_session_name (tmux display-message -p '#S')
# Attempt to activate the Mamba environment with the same name as the tmux session
# Redirect stderr to /dev/null to suppress any error messages
mamba activate $tmux_session_name[1]
endStarship
I use Starship, a simple/fast cross-shell prompt. For the past few years I’ve mostly used fish as my shell, but often need to hop into bash for POSIX compatibility (I will probably see if I can switch to zsh and get everything I’m used to in fish at some point).
I’m not really a power user or a ricer, but it has a lot of nice stuff built in. One customization I use in ~/.config/starship.toml to help me tell what shell I’m in:
[shell]
fish_indicator = "🐟"
zsh_indicator = "ℤ"
bash_indicator = "\\$"
Atuin
Atuin is a better history manager that I’ve been using for a few years now, and despite having to do a lot of configuration to get/keep it working how I want, I couldn’t imagine going back to anything that has less functionality. It really pains me when I am on a terminal system without this.
In my ~/.config/atuin/config.toml the major changes I make from defaults:
# I almost always want to find the last thing I typed *in each session*, not globally
filter_mode_shell_up_key_binding = "session"
# when I search I usually want to see my current context
inline_height = 16
# Maybe some people want to just execute, but I like to treat the picker like a picker and either execute or edit
enter_accept = false
I don’t use the cloud syncing nor have I ever bothered to setup a my own server - at some point maybe there will be a smart SSH-based lazy-sync or something but for now, I’ve been OK with not having the sync
Tools I Want to Integrate
llm
There are a bunch of command-line LLM tools but simonw’s llm is probably the most flexible/mature. I’ve yet to really integrate it (or any related tools into my workflow)
wut
This is still a little underbaked atm but still a great idea: https://github.com/shobrook/wut