Intro
Activation Beacon compresses the original KV into fewer yet more compact states (a.k.a. beacons) and hence enables the LLM to perceive longer context given its fixed context window. It is known for the following features:
- Effective
- there is little information loss given a compression ratio of 2, 4, and 8;
- Efficient
- it drastically reduces the GPU consumption of KV cache;
- Compatible
- it can work together with position extrapolation (e.g. YaRN) to further extends the context length; it can also work with grouped query attention to further reduce the KV cache size;
- Low-Cost
- it is light-weight and can be efficiently trained with roughly 1B tokens.
Environment
pip install transformers
pip install flash-attn --no-build-isolation
Usage
import json
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "namespace-Pt/beacon-qwen-2-7b-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
model_id,
trust_remote_code=True,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2"
)
model = model.cuda().eval()
with torch.no_grad():
# short context
messages = [{"role": "user", "content": "Tell me about yourself."}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda")
outputs = model.generate(**inputs, max_new_tokens=50)
print(f"Input Length: {inputs['input_ids'].shape[1]}")
print(f"Output: {repr(tokenizer.decode(outputs[0], skip_special_tokens=True))}")
# reset memory before new generation task
model.memory.reset()
# long context
with open("infbench.json", encoding="utf-8") as f:
example = json.load(f)
messages = [{"role": "user", "content": example["context"]}]
inputs = tokenizer.apply_chat_template(messages, tokenize=True, add_generation_prompt=True, return_tensors="pt", return_dict=True).to("cuda")
outputs = model.generate(**inputs, do_sample=False, top_p=1, temperature=1, max_new_tokens=20)[:, inputs["input_ids"].shape[1]:]
print("*"*20)
print(f"Input Length: {inputs['input_ids'].shape[1]}")
print(f"Answers: {example['answer']}")
print(f"Prediction: {tokenizer.decode(outputs[0], skip_special_tokens=True)}")
NOTE: It's okay to see warnings like This is a friendly reminder - the current text generation call will exceed the model's predefined maximum length (32768). Depending on the model, you may observe exceptions, performance degradation, or nothing at all.
Just ignore it.
Results
LongBench
Model | Single QA | Multi QA | Summarization | Few-Shot | Code | AVG |
---|---|---|---|---|---|---|
qwen-2-7b-instruct | 39.60 | 36.92 | 27.97 | 71.12 | 62.34 | 47.59 |
beacon-qwen-2-7b-instruct | 40.76 | 43.73 | 27.23 | 68.87 | 68.47 | 49.81 |
NIAH
- Downloads last month
- 276
This model does not have enough activity to be deployed to Inference API (serverless) yet. Increase its social
visibility and check back later, or deploy to Inference Endpoints (dedicated)
instead.