QW
Qwen/Qwen2.5-72B-Instruct
text-generationtransformersqwen72.7B params · updated Jan 2025
Usage
usage.pytemplate from repo id
from osai import AutoModelForCausalLM, AutoTokenizer
repo = "Qwen/Qwen2.5-72B-Instruct"
tok = AutoTokenizer.from_pretrained(repo)
model = AutoModelForCausalLM.from_pretrained(
repo, dtype="bfloat16", device_map="auto",
)
msgs = [{"role": "user", "content": "Explain speculative decoding."}]
ids = tok.apply_chat_template(msgs, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(ids, max_new_tokens=256)[0]))Hardware
Computed from 72.71B parameters × bytes per parameter, plus 20% for activations and KV cache. Not measured.
| Precision | Weights | Estimated VRAM | Fits on |
|---|---|---|---|
| bf16 | 135 GB | 163 GB | Multi-GPU (2× A100 80 GB or more) |
| int8 | 67.7 GB | 82 GB | A100 80 GB |
| nf4 | 33.9 GB | 41 GB | A10G 24 GB |