FlashML is a lightweight library (with minimal import overhead) dedicated to ML/RL/NLP engineers working with PyTorch. Not published yet on PyPI, but you can install it by opening this project in VSCode, and with conda env call:
pip install -e .
and add this line to VSCode's settings.json:
python.analysis.extraPath":["path\\to\\flashml"],from flashml import resource_monitor
resource_monitor()
from flashml import inspect_model, inspect_tokenizer
tokenizer = AutoTokenizer.from_pretrained("<any-huggingface-tokenizer>")
model = AutoModelForCausalLM.from_pretrained("<any-pytorch-model>")
inspect_model(model)
inspect_tokenizer(tokenizer)
from flashml import plot_tensor
plot_tensor(torch.randn(10, 8, device: 'cuda:0'))
from flashml.schedulers import LRScheduler
scheduler = LRScheduler(
optimizer=optim,
max_steps=1000,
warmup_steps=100, # linear warmup steps
constant_steps=200, # constant lr steps
min_lr=1e-8,
annealing_type='cosine', # ['linear', 'exponential', 'logarithmic']
num_cycles=3,
max_lr_decay_factor=0.6, # decay of max lr after each cycle
)
from flashml import plot_graph
plot_graph(([10, 15, 17, 18, 18.3], [8, 14, 17.1, 18.9, 19.02]))from flashml import plot_rt_graph
for i in range(T):
plot_rt_graph((<priceBTC>, <priceETH>), <timestep>, x_label="Time", y_label="Price", color=["yellow", "purple])from flashml import parallel_for, parallel_foreach
def my_func(x):
...
result = parallel_for(0, 1e6, my_func, num_workers=16)from flashml import log_metrics, BatchIterator,log_checkpoint, load_checkpoint
from flashml.classification import plot_confusion_matrix
for batch in BatchIterator(df=train_df, num_epochs=10, batch_size=32, mode="train"):
# batch.value, batch.index, batch.step, batch.ids
# Compute loss and perform update
# Compute validation
log_metrics(
metrics={ "loss": loss , "acc" : acc},
step=idx)
plot_confusion_matrix(yHat, y)
log_checkpoint({"model":model.state_dict(), "optim":optim.state_dict()}) # this will be logged in MLFlow
# later load with load_checkpoint(run_name="charming-seal-738", version=1)Output (rt + MLFlow session log):
75%|████████████████████████████████████████████████████████████████████████ | 225/300 [00:00<00:00, 2059.44it/s, loss=2.25, acc=5.17]
from flashml.classification import make_dummy_classification_dataset, run_dummy_classifiers, run_linear_classifier
x, y, a, b = make_dummy_classification_dataset(num_samples=2048, num_features=10, num_classes=2, weights=[0.25, 0.75])
run_dummy_classifiers(x, y, a, b) # dummy classifications (uniform, prior, full zeros/ones, stratified)
run_linear_classifier(x, y, a, b, regularization="l1")from flashml.rl import log_episodes
while step < max_steps:
for i in range(buffer_size):
# Step env, collect info
if episode is done:
log_episode(
cumulative_reward=reward,
episode_length=episode_len,
step = (step, max_steps))Output (rt):
Cumulated Reward [max: 3.236] [µ: 2.599] [σ: 0.30z]
Episode Length [max: 685] [µ: 87.310] [σ: 73.73z]
[Episode 350]: 37%|███████████████████████████▎ | 37409/100000 [00:00<00:00, 112028.23it/s]
from flashml.nlp import *
df = pl.read_csv(...)
df = lowercase(df, "text_col")
df = lemmatize(df, "text_col")
df = replace_emojis(df, "text_col", "E")
df = remove_double_spacing(df, "text_col")
... and other operationsfrom flashml import plot_tsne
import numpy as np
x1 = np.random.normal(loc=34, scale=3, size=(120, 99))
x2 = np.random.normal(loc=22, scale=5, size=(120, 99))
x3 = np.random.normal(loc=-32, scale=2, size=(200, 99))
x = np.concatenate([x1, x2, x3], axis=0)
plot_tsne(x, labels=["A"] * 120 + ["B"] * 120 + ["C"] * 200)
Enumerate GQA, SwiGLU, FFN, MinGRU ...
- Color picker
- Image converter, resizer, processor etc.
- File converter
- Local chat (working with ollama)
#!/bin/bash
set -e
# Define variables for clarity
INSTALLER_PATH="/workspace/Anaconda3-2025.06-0-Linux-x86_64.sh"
INSTALL_DIR="/workspace/anaconda3"
BASHRC="$HOME/.bashrc"
ENV_NAME="ml"
echo "==== 1. Installing essential system packages ===="
apt-get update
apt-get install -y build-essential curl git bzip2 wget nvtop
echo "==== 2. Checking if Anaconda is installed ===="
# Logic Fix: Check if the DIRECTORY exists to determine if we need to install
if [ -d "$INSTALL_DIR" ]; then
echo "Anaconda directory ($INSTALL_DIR) already exists. Skipping installation."
else
echo "Anaconda not found. Checking for installer..."
# Check if installer file exists, download if not
if [ ! -f "$INSTALLER_PATH" ]; then
echo "Downloading Anaconda..."
wget https://repo.anaconda.com/archive/Anaconda3-2025.06-0-Linux-x86_64.sh -P /workspace
fi
echo "Running Anaconda Installer..."
bash "$INSTALLER_PATH" -b -p "$INSTALL_DIR"
fi
# Initialize conda logic
export PATH="$INSTALL_DIR/bin:$PATH"
source "$INSTALL_DIR/etc/profile.d/conda.sh"
conda init bash
# Accept TOS
echo "==== Accepting Anaconda Terms of Service ===="
# Note: This command is specific to newer Anaconda versions/paid tiers.
# If this fails, you may simply not need it on the free tier.
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main || true
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r || true
# Set solver
conda config --set solver libmamba
# Add HF_HOME to bashrc
if ! grep -q 'HF_HOME' "$BASHRC"; then
echo 'export HF_HOME=/workspace/hf_cache' >> "$BASHRC"
fi
echo "==== 3. Checking if '$ENV_NAME' environment exists ===="
if conda env list | grep -qE "(^|\s)$ENV_NAME($|\s)"; then
echo "'$ENV_NAME' environment already exists."
else
echo "Creating '$ENV_NAME' environment..."
conda create -y -n $ENV_NAME python=3.12.9
fi
echo "==== 4. Installing flashml package ===="
if [ ! -d "/workspace/flashml" ]; then
cd /workspace
git clone https://github.com/smtmRadu/flashml flashml
cd flashml
"$INSTALL_DIR/envs/$ENV_NAME/bin/pip" install --upgrade pip setuptools wheel
"$INSTALL_DIR/envs/$ENV_NAME/bin/pip" install -e .
#"$INSTALL_DIR/envs/$ENV_NAME/bin/pip" install vllm
# i comment the vllm install because unsloth doesn't really match with the newest version, and installs an outdated pytorch version. So install it separately, manually
else
echo "flashml already exists, skipping..."
fi
# Make 'ml' auto-activate on new terminals
if ! grep -q "conda activate $ENV_NAME" "$BASHRC"; then
echo "conda activate $ENV_NAME" >> "$BASHRC"
fi
echo "==== 5. Setup completed! ===="
echo "Please restart your terminal or run: source ~/.bashrc"