#vram

Where the memory goes when you fine-tune — weights, adapters, optimiser state, activations — and the concrete settings that fit a 1–3B model into 4–6

A local model can load and still run out of memory at longer context. Compare two controlled loads, inspect cache logs, and test the real workload.

Four VRAM numbers that looked wrong on an 8GB laptop, none of them a bug: what dxdiag...

HBM Architecture GPU memory is the most constrained resource in ML. This post explains HBM...

Qwen3.8-27B Q3_K_S loaded on an RTX 3070. VRAM used 7,435 of 8,192 MiB. Decode ran 2.07 tok/s. Fit on 8 GB is not a usable rate.

Most '16GB is enough' advice misses what breaks at 34B+. Full Q4-Q8 VRAM tiers + the budget mistake that costs you a year of upgrades.

AMD's RX 9050 4GB shows how low-end buyers may get modern GPU branding with VRAM limits that strain new games.

My RTX 5090 measurements show how I budget VRAM for a local model, a real workload, and the GPU processes that must stay resident beside it.

A local LLM needs about half a gigabyte of VRAM per billion parameters at Q4, then KV cache and context stack on top. Here is how to know a model fits before you download 40 GB.

Ollama Keeps Reloading the Model? Fix VRAM Unloading, Cold Starts, and Model Swapping (2026)

What hardware do you need for Llama 4 Maverick 400B? Multi-GPU requirements, cloud options, and whether it's worth self-hosting.

FP16 vs FP8 vs NF4 for Stable Diffusion and Flux — which quantization gives the best quality-to-VRAM tradeoff for image gen.