This week All our GPUs are sold at cost price — zero margin on the server. See the cost sheets

This week Every GPU sold at cost price

Software

Pre-loading a model

How serving templates pull a model before first boot: Hugging Face and Ollama names, precision, the memory-fit rule the order form enforces, changing the model later.

4 min read · updated 2 September 2026

With vLLM, SGLang, TGI or Ollama, the order form asks for a model and a precision. The weights are pulled while the server is provisioned, so the API answers at first boot instead of an hour later.

Naming the model

  • Hugging Face repositories, as owner/name: meta-llama/Llama-3.3-70B-Instruct, Qwen/Qwen2.5-72B-Instruct. Gated repositories need your own token: give it after boot and restart the service.
  • Ollama tags for the Ollama template: llama3.1:70b, qwen2.5:72b-instruct-q4_K_M.

Precision

PrecisionWhenMemory
BF16 / FP16full quality · 2 bytes per parameter2 B/param
FP8near-lossless · 1 byte · Ada, Hopper, Blackwell, MI300X1 B/param
INT81 byte per parameter · any card1 B/param
INT4 (AWQ / GPTQ)0.5 byte per parameter · smallest footprint0.5 B/param

FP8 needs Ada, Hopper, Blackwell or MI300X silicon; the form refuses it on Ampere cards. INT8 and INT4 use quantised checkpoints (AWQ, GPTQ, GGUF); name one that exists for the model you want.

The fit rule

The form reads the parameter count from the name (70B, 8x22B, :72b) and checks that parameters × bytes per parameter × 1.2 is at most the server's total GPU memory. The 20 % covers the KV cache and runtime buffers at a normal context length. Example: 70 B parameters at BF16 / FP16 need about 168 GB; this server has 160 GB. Pick a smaller precision, more cards or a bigger card. At FP8: 70 B parameters at FP8 need about 84 GB with a working KV cache; this server has 160 GB.

A name without a parameter count is accepted without a check; make sure it fits. The VRAM guide has the full table.

Changing the model later

nano /opt/dedigpu/vllm/env        # MODEL=..., DTYPE=..., MAX_MODEL_LEN=...
systemctl restart vllm
journalctl -u vllm -f              # watch the new weights load

Weights are cached under /opt/dedigpu/models on the local NVMe; a 70B model in BF16 is about 140 GB, so keep an eye on df -h on the smaller cards.

Try it on a real server.

One month, at cost, root in under 5 minutes. Support opens with the first server.