Tag: python

Crawl a docs site to clean markdown with Crawl4AI, embed it with Ollama, store it in Qdrant, and query it. A working local RAG ingest pipeline, with politeness.

A 5.8M-parameter GPT scores 1.626 bits per byte on Holmes. Gemma 4 E4B scores 0.953 untrained, and a 34-minute QLoRA run on an 8 GB GPU gets it to 0.865.

Merge Chrome, Firefox, Pocket, Raindrop and Pinboard exports into one deduplicated file with tags, triage dead links, and import it all into linkding.

The KV cache sped a 5.8M-parameter GPT up 4.4x on a laptop CPU and did almost nothing for the RTX 3070 at batch 1. Then sampling settings decide how it reads.

Train a 5.8M-parameter GPT on Sherlock Holmes: 59 seconds on an RTX 3070 vs 48 minutes on its CPU, and the overfitting that starts near step 1,250 of training.

Build causal self-attention and RoPE in plain PyTorch, then measure naive vs fused attention on an 8GB GPU: 6400 MiB vs 8.3 MiB of VRAM at 16k tokens.

Build a byte-level BPE tokenizer in about 100 lines of Python, then compare it to tiktoken's gpt2, cl100k_base and o200k_base on prose, code, YAML and German.

The Python libraries actually worth your time in 2026: data tools, ML frameworks, and CLI kit, plus which ones earn a permanent spot in your home lab.

FastAPI is async and type-driven. Flask is the micro-framework. Django is batteries-included. Here's how to pick for homelab apps and small SaaS.

LangChain does everything and LlamaIndex does one thing brilliantly. Here's how to pick the right RAG framework without regretting it at 2 AM.

LiteLLM proxies every LLM, local or cloud, behind one OpenAI-compatible endpoint. Pair it with vLLM for GPU-backed serving and ditch the SDK sprawl.

Text Generation Web UI vs KoboldCpp: setup, model formats, samplers, APIs, and performance compared so you can pick the right local LLM frontend fast.