File size: 1,558 Bytes
64bbe57 957dd91 64bbe57 2bd9edb 64bbe57 f485ece 20f6d69 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 |
---
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")
```
|