---
license: llama3.1
---
Amazingly quick to inference on Ampere GPUs like 3090 Ti. in INT8. In VLLM I left it on a task for 10 minutes with prompt caching, average fixed input around 2000, variable input around 200 and output around 200.
Averaged over a second, that's 22.5k t/s prompt processing and 1.5k t/s generation.
Averaged over an hour that's 81M input tokens and 5.5M output tokens. Peak generation speed I see is around 2.6k/2.8k t/s.

Quantized on H100. On 3090 Ti I was OOMing.

Creation script:

```python
from transformers import AutoTokenizer
from datasets import Dataset
from llmcompressor.transformers import SparseAutoModelForCausalLM, oneshot
from llmcompressor.modifiers.quantization import GPTQModifier
import random

model_id = "NousResearch/Hermes-3-Llama-3.1-8B"

num_samples = 256
max_seq_len = 8192

tokenizer = AutoTokenizer.from_pretrained(model_id)

max_token_id = len(tokenizer.get_vocab()) - 1
input_ids = [[random.randint(0, max_token_id) for _ in range(max_seq_len)] for _ in range(num_samples)]
attention_mask = num_samples * [max_seq_len * [1]]
ds = Dataset.from_dict({"input_ids": input_ids, "attention_mask": attention_mask})

recipe = GPTQModifier(
  targets="Linear",
  scheme="W8A8",
  ignore=["lm_head"],
  dampening_frac=0.01,
)

model = SparseAutoModelForCausalLM.from_pretrained(
  model_id,
  device_map="auto",
)

oneshot(
  model=model,
  dataset=ds,
  recipe=recipe,
  max_seq_length=max_seq_len,
  num_calibration_samples=num_samples,
)

model.save_pretrained("NousResearch_Hermes-3-Llama-3.1-8B.w8a8")

```