dask
Skill8 starsChanged 8 months agoMITView on GitHub
Provides Dask parallel computing framework for scaling Python workloads across cores and clusters. Covers DataFrames, Arrays, Bags, Delayed, distributed scheduler, GPU acceleration with cuDF/CuPy, and HPC deployment with Dask-Jobqueue.
16 other projects use a version of this.
Adopted by
Most recently taken up by
- Ivan-propio/claude-skillstheir own version, first seen Nov 16, 2026
- BrightGold70/skillstheir own version, first seen Nov 8, 2026
- MarieLynneBlock/arcanum-artifextheir own version, first seen Nov 1, 2026
- davidanthonyn/claude-skillstheir own version, first seen Oct 20, 2026
- patoangolano/bio-stacktheir own version, first seen Oct 15, 2026
- G1Joshi/Agent-Skillstheir own version, first seen Oct 13, 2026
Other versions16 projects use a version of this.
Word for word · 0
None found.
Edited · 0
None found.
Wrote their own version · 16
- wu-yc/LabClaw★ 1.1ksame name
- LeonChaoX/qinyan-academic-skills★ 910same name
- qinyan-ai/qinyan-academic-skills★ 80same name
- Barrios88/barrios-skills★ 63same name
- tondevrel/scientific-agent-skills★ 22same name
- G1Joshi/Agent-Skills★ 15same name
- MarieLynneBlock/arcanum-artifex★ 4same name
- JunMA98/Computer-science-claude-skills★ 3same name
- BrightGold70/skills★ 1same name
- garri333/Skills★ 1same name
6 more not listed
How matches are found. Word for word: the exact same text. Edited: shares passages with this file (“99% same” means 99% of its five-word phrases also appear here; 10% or more counts). Their own version: similar meaning, or, weakest, the same name. We can’t tell which project had it first. Projects are counted among the 200 most-starred candidates, and mentions are links to the repository we found, up to 50 per site, so both counts are a floor.
In this repository23 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- validate-skillsSkill
- alchemi-toolkit-opsSkill
- autoresearch-setupSkill
- bionemoSkill
- biosimspaceSkill
- cuda-quantumSkill
- cutileSkill
- flashinferSkill
- materials-projectSkill
- mpiSkill
- nemotronSkill
- omniverse-simreadySkill
- physicsnemoSkill
- polarsSkill
- prime-verifiersSkill
- qiskitSkill
- raySkill
- rdkitSkill
- slurmSkill
- tinkerSkill
- tritonSkill
The file
We split this skill into the rules it gives an agent
4 rules. 0 are common, 3 are distinctive, and 1 is very specific to this project.
0
Common
Most of the 16 edited versions we read keep these.
3
Distinctive
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 440
Persist intermediate results: Use .persist() for reused computations
- Distinctive: line 442
Write results to disk: Avoid returning large datasets to client
- Distinctive: line 444
Deploy on HPC: Use Dask-Jobqueue for SLURM/PBS clusters
- Distinctive: line 447
Profile before scaling: Ensure single-machine optimization first
“Common” is measured: kept by at least 6 in 10 of the 16 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Dask Agent Skill
Agent-optimized skill for Dask parallel computing framework.
Quick Reference
Core Collections
import dask.dataframe as dd
import dask.array as da
import dask.bag as db
from dask import delayed
# DataFrame - parallel pandas
df = dd.read_parquet('data/*.parquet')
result = df.groupby('key').value.mean().compute()
# Array - parallel NumPy
x = da.from_zarr('data.zarr')
y = (x + x.T).mean(axis=0).compute()
# Bag - parallel lists
b = db.read_text('logs/*.txt')
counts = b.map(str.split).flatten().frequencies().compute()
# Delayed - custom parallelism
@delayed
def process(x):
return x * 2
results = [process(i) for i in range(10)]
total = delayed(sum)(results).compute()Distributed Scheduler
from dask.distributed import Client, LocalCluster
# Local cluster (automatic)
client = Client()
# Local cluster (manual configuration)
cluster = LocalCluster(n_workers=4, threads_per_worker=2, memory_limit='4GB')
client = Client(cluster)
# Remote cluster
client = Client('scheduler-address:8786')
# Persist data in distributed memory
df = df.persist()
# Monitor progress
client.dashboard_link # http://localhost:8787/statusGPU Acceleration
import dask_cudf
import cupy as cp
# GPU DataFrame (cuDF)
df = dask_cudf.read_parquet('data/*.parquet')
result = df.groupby('key').value.mean().compute()
# GPU Array (CuPy)
x = da.from_array(cp.random.random((10000, 10000), dtype='float32'), chunks=(1000, 1000))
y = x @ x.T
result = y.compute()
# GPU cluster
from dask_cuda import LocalCUDACluster
cluster = LocalCUDACluster()
client = Client(cluster)HPC Deployment
from dask_jobqueue import SLURMCluster
# SLURM cluster
cluster = SLURMCluster(
cores=24,
processes=4,
memory="100GB",
walltime="02:00:00",
queue="gpu",
job_extra_directives=["--gres=gpu:2"]
)
cluster.scale(jobs=10) # Submit 10 jobs
client = Client(cluster)Common Patterns
DataFrame Operations
import dask.dataframe as dd
# Read multiple files
df = dd.read_csv('data/*.csv')
df = dd.read_parquet('data/*.parquet')
df = dd.read_json('data/*.json')
# Compute lazily, execute with .compute()
result = df.groupby('category').price.mean() # Lazy
result.compute() # Execute
# Persist in memory for reuse
df = df.persist()
# Write results
df.to_parquet('output/', compression='snappy')
df.to_csv('output/*.csv')
# Set index for efficient queries
df = df.set_index('timestamp')
subset = df.loc['2026-01-01':'2026-01-31'].compute()Array Operations
import dask.array as da
import numpy as np
# Create arrays
x = da.random.random((10000, 10000), chunks=(1000, 1000))
x = da.from_array(np.ones((5000, 5000)), chunks=(500, 500))
x = da.from_zarr('data.zarr')
# NumPy-like operations
y = x.mean(axis=0)
z = da.linalg.svd(x)
w = x @ x.T
# Rechunk for different access patterns
x = x.rechunk((2000, 500))
# Compute
result = y.compute()
# Store
da.to_zarr(x, 'output.zarr')Custom Parallelism with Delayed
from dask import delayed
import pandas as pd
# Wrap functions
@delayed
def load_data(filename):
return pd.read_csv(filename)
@delayed
def clean_data(df):
return df.dropna()
@delayed
def process_data(df):
return df.groupby('key').sum()
# Build task graph
files = ['data1.csv', 'data2.csv', 'data3.csv']
loaded = [load_data(f) for f in files]
cleaned = [clean_data(df) for df in loaded]
processed = [process_data(df) for df in cleaned]
# Combine and compute
result = delayed(pd.concat)(processed).compute()Distributed Futures
from dask.distributed import Client
client = Client()
# Submit tasks
def square(x):
return x ** 2
futures = client.map(square, range(1000))
# Gather results
results = client.gather(futures)
# Submit single task
future = client.submit(sum, range(1000))
result = future.result()
# Submit with dependencies
a = client.submit(sum, [1, 2, 3])
b = client.submit(lambda x: x * 2, a)
b.result()Resource Management
Cluster Configuration
from dask.distributed import Client, LocalCluster
# Configure workers
cluster = LocalCluster(
n_workers=8,
threads_per_worker=2,
processes=True,
memory_limit='8GB',
dashboard_address=':8787'
)
client = Client(cluster)
# Adaptive scaling
cluster.adapt(minimum=2, maximum=10)
# Scale manually
cluster.scale(5)GPU Clusters
from dask_cuda import LocalCUDACluster
from dask.distributed import Client
# Auto-detect GPUs
cluster = LocalCUDACluster()
client = Client(cluster)
# Manual GPU configuration
cluster = LocalCUDACluster(
n_workers=4, # 4 GPUs
threads_per_worker=1,
memory_limit='16GB',
device_memory_limit='16GB'
)SLURM Integration
from dask_jobqueue import SLURMCluster
cluster = SLURMCluster(
cores=24, # CPUs per job
processes=4, # Dask workers per job
memory="96GB", # Memory per job
walltime="04:00:00", # Job time limit
queue="gpu", # Partition
job_extra_directives=[
"--gres=gpu:2",
"--account=myproject"
]
)
# Submit worker jobs
cluster.scale(jobs=20) # 20 SLURM jobs
# Or adaptive
cluster.adapt(minimum_jobs=5, maximum_jobs=50)
client = Client(cluster)Performance Optimization
Chunking Strategy
# DataFrame partitions
df = dd.read_parquet('data.parquet', npartitions=100)
df = df.repartition(npartitions=50)
# Array chunks
x = da.from_array(data, chunks=(1000, 1000)) # 1000x1000 chunks
x = x.rechunk((2000, 500)) # Rechunk for different access
# Optimal chunk size: 100MB-1GB per chunkPersist vs Compute
# Compute: Execute and return to client
result = df.sum().compute()
# Persist: Execute and keep in distributed memory
df = df.persist() # Reuse without recomputation
# Explicit task graph
from dask import delayed, compute
x = delayed(load_data)('file1.csv')
y = delayed(load_data)('file2.csv')
z = delayed(merge)(x, y)
result = compute(z)[0] # Execute entire graphDashboard Monitoring
client = Client()
print(client.dashboard_link) # http://localhost:8787/status
# Monitor:
# - Task stream
# - Progress bars
# - Memory usage
# - Worker status
# - Task graph visualizationMachine Learning Integration
Dask-ML
import dask_ml.cluster as dask_cluster
from dask_ml.linear_model import LogisticRegression
# Scalable k-means
kmeans = dask_cluster.KMeans(n_clusters=10)
kmeans.fit(X)
# Incremental learning
lr = LogisticRegression()
lr.fit(X_train, y_train)XGBoost
import xgboost as xgb
import dask.dataframe as dd
df = dd.read_parquet('data.parquet')
X = df[['feature1', 'feature2', 'feature3']]
y = df['target']
dtrain = xgb.dask.DaskDMatrix(client, X, y)
params = {'max_depth': 5, 'objective': 'reg:squarederror'}
model = xgb.dask.train(client, params, dtrain, num_boost_round=100)GPU Machine Learning
import cuml.dask as cuml_dask
from cuml.dask.ensemble import RandomForestClassifier
# GPU random forest
rf = RandomForestClassifier(n_estimators=100)
rf.fit(X_train, y_train)
predictions = rf.predict(X_test).compute()Common Issues
Memory Management
# Avoid large results returned to client
result = df.compute() # Bad if result is huge
# Instead, write to disk
df.to_parquet('output/') # Good
# Or aggregate before computing
result = df.groupby('key').mean().compute() # Good if grouped result is smallScheduler Selection
# Threaded (default for arrays)
import dask
dask.config.set(scheduler='threads')
# Processes (default for dataframes)
dask.config.set(scheduler='processes')
# Distributed (for clusters)
from dask.distributed import Client
client = Client() # Uses distributed schedulerTask Graph Optimization
# Avoid large task graphs
for i in range(10000): # Bad: 10000 tasks
result = delayed(process)(i)
# Use collections instead
data = da.from_delayed([delayed(process)(i) for i in range(10000)],
shape=(10000,), dtype=float) # GoodReferences
Detailed documentation in references/:
- dataframe-array-bag.md: Core collections API and patterns
- distributed-scheduler.md: Cluster setup, client usage, task scheduling
- gpu-acceleration.md: cuDF, CuPy, LocalCUDACluster, best practices
- deployment.md: Local, cloud, HPC deployment options
Official Documentation
Best Practices
- Use appropriate collections: DataFrame for tabular, Array for numeric, Bag for semi-structured
- Chunk wisely: 100MB-1GB chunks for optimal performance
- Persist intermediate results: Use
.persist()for reused computations - Monitor with dashboard: Track task progress and resource usage
- Write results to disk: Avoid returning large datasets to client
- Use GPU acceleration: cuDF/CuPy for GPU-compatible operations
- Deploy on HPC: Use Dask-Jobqueue for SLURM/PBS clusters
- Adaptive scaling: Let cluster adjust to workload demands
- Optimize task graphs: Minimize overhead with collection-based parallelism
- Profile before scaling: Ensure single-machine optimization first
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

