huggingface

name	huggingface
description	Browse, download, and deploy models, datasets, and spaces on HuggingFace Hub
version	1.0.0
author	kai-agent
metadata	{"kai":{"tags":["kai","compute","huggingface","models","datasets","inference","ml"]}}

HuggingFace Hub

Skill for browsing, downloading, uploading, and running inference on models, datasets, and spaces from HuggingFace.

Authentication

This skill uses the HF_TOKEN environment variable. If it is not set, ask your admin to add it in Agent Settings. You can generate a token at https://huggingface.co/settings/tokens.

Some operations (public model browsing, public dataset loading) work without a token, but gated models, private repos, and uploads always require one.

Install the Hub library

pip install huggingface-hub transformers datasets

Browsing and searching models

Search by task

from huggingface_hub import HfApi

api = HfApi()

# Find top text-generation models sorted by downloads
models = api.list_models(
    task="text-generation",
    sort="downloads",
    direction=-1,
    limit=10,
)
for m in models:
    print(f"{m.id}  downloads={m.downloads}")

Search by name or keyword

models = api.list_models(search="llama", sort="downloads", direction=-1, limit=5)

Filter by library (vLLM, TGI, ONNX, etc.)

models = api.list_models(
    task="text-generation",
    library="vllm",
    sort="downloads",
    direction=-1,
    limit=10,
)

This is particularly useful for inference optimization: filter by vllm or text-generation-inference to find models already packaged for fast serving.

Reading model cards

from huggingface_hub import ModelCard

card = ModelCard.load("meta-llama/Llama-3.1-8B-Instruct")
print(card.text)          # Full markdown
print(card.data.tags)     # Metadata tags
print(card.data.license)  # License info

Downloading models

Download a single file

from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="meta-llama/Llama-3.1-8B-Instruct",
    filename="config.json",
)
print(path)

Download an entire model snapshot

from huggingface_hub import snapshot_download

local_dir = snapshot_download(
    repo_id="meta-llama/Llama-3.1-8B-Instruct",
    local_dir="./llama-3.1-8b",
    ignore_patterns=["*.md", "*.txt"],  # skip non-essential files
)

Download quantized / optimized variants

For inference optimization, look for GGUF, AWQ, or GPTQ variants:

models = api.list_models(search="Llama-3.1-8B AWQ", sort="downloads", direction=-1, limit=5)

Uploading models

api.upload_folder(
    folder_path="./my-fine-tuned-model",
    repo_id="your-org/my-model",
    repo_type="model",
    commit_message="Upload fine-tuned checkpoint",
)

Datasets

Search datasets

datasets = api.list_datasets(search="code", sort="downloads", direction=-1, limit=10)

Load a dataset

from datasets import load_dataset

ds = load_dataset("openai/gsm8k", split="train")
print(ds[0])

Stream large datasets without downloading

ds = load_dataset("allenai/c4", split="train", streaming=True)
for example in ds:
    print(example["text"][:200])
    break

Inference API

Serverless inference (no GPU required)

from huggingface_hub import InferenceClient

client = InferenceClient(model="meta-llama/Llama-3.1-8B-Instruct")

response = client.text_generation(
    "Explain quantization in 3 sentences:",
    max_new_tokens=200,
    temperature=0.7,
)
print(response)

Chat-style inference

response = client.chat_completion(
    messages=[{"role": "user", "content": "What is KV-cache quantization?"}],
    max_tokens=500,
)
print(response.choices[0].message.content)

Embeddings

embeddings = client.feature_extraction("Inference optimization for LLMs")
print(len(embeddings[0]))  # embedding dimension

Spaces

Create a Space

api.create_repo(
    repo_id="your-org/my-demo",
    repo_type="space",
    space_sdk="gradio",  # or "streamlit", "docker"
)

Upload files to a Space

api.upload_folder(
    folder_path="./my-space-app",
    repo_id="your-org/my-demo",
    repo_type="space",
)

Duplicate an existing Space

api.duplicate_space("stabilityai/stable-diffusion", to_id="your-org/sd-copy")

Inference optimization tips

1. Check quantized variants first: Search for AWQ/GPTQ/GGUF versions before downloading full-precision weights.
2. Use library filter: Filter by vllm or text-generation-inference for deployment-ready models.
3. Check model size vs. your GPU memory: Read the model card for memory requirements.
4. Use streaming for large datasets: Avoid downloading multi-TB datasets to disk.
5. Cache management: Models are cached in ~/.cache/huggingface/. Use huggingface-cli scan-cache to inspect and huggingface-cli delete-cache to free space.

Rules

1. ALWAYS check if a gated model requires access before attempting download (e.g., Llama models require accepting a license)
2. Prefer quantized models when the use case is inference, not training
3. Use ignore_patterns to skip downloading unnecessary files (READMEs, training configs)
4. Set HF_HOME if you want to customize the cache directory
5. Never commit tokens to code: rely on HF_TOKEN environment variable