Generate Text Using Eagle Decoding#
Source NVIDIA/TensorRT-LLM.
1### Generate Text Using Eagle Decoding
2
3from tensorrt_llm import LLM, SamplingParams
4from tensorrt_llm.llmapi import (LLM, EagleDecodingConfig, KvCacheConfig,
5 SamplingParams)
6
7
8def main():
9 # Sample prompts.
10 prompts = [
11 "Hello, my name is",
12 "The president of the United States is",
13 "The capital of France is",
14 "The future of AI is",
15 ]
16 # The end user can customize the sampling configuration with the SamplingParams class
17 sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
18
19 # The end user can customize the kv cache configuration with the KVCache class
20 kv_cache_config = KvCacheConfig(enable_block_reuse=True)
21
22 llm_kwargs = {}
23
24 model = "lmsys/vicuna-7b-v1.3"
25
26 # The end user can customize the eagle decoding configuration by specifying the
27 # speculative_model, max_draft_len, num_eagle_layers, max_non_leaves_per_layer, eagle_choices
28 # greedy_sampling,posterior_threshold, use_dynamic_tree and dynamic_tree_max_topK
29 # with the EagleDecodingConfig class
30
31 speculative_config = EagleDecodingConfig(
32 speculative_model="yuhuili/EAGLE-Vicuna-7B-v1.3",
33 max_draft_len=63,
34 num_eagle_layers=4,
35 max_non_leaves_per_layer=10,
36 eagle_choices=[[0], [0, 0], [1], [0, 1], [2], [0, 0, 0], [1, 0], [0, 2], [3], [0, 3], [4], [0, 4], [2, 0], \
37 [0, 5], [0, 0, 1], [5], [0, 6], [6], [0, 7], [0, 1, 0], [1, 1], [7], [0, 8], [0, 0, 2], [3, 0], \
38 [0, 9], [8], [9], [1, 0, 0], [0, 2, 0], [1, 2], [0, 0, 3], [4, 0], [2, 1], [0, 0, 4], [0, 0, 5], \
39 [0, 0, 0, 0], [0, 1, 1], [0, 0, 6], [0, 3, 0], [5, 0], [1, 3], [0, 0, 7], [0, 0, 8], [0, 0, 9], \
40 [6, 0], [0, 4, 0], [1, 4], [7, 0], [0, 1, 2], [2, 0, 0], [3, 1], [2, 2], [8, 0], \
41 [0, 5, 0], [1, 5], [1, 0, 1], [0, 2, 1], [9, 0], [0, 6, 0], [0, 0, 0, 1], [1, 6], [0, 7, 0]]
42 )
43
44 llm = LLM(model=model,
45 kv_cache_config=kv_cache_config,
46 speculative_config=speculative_config,
47 max_batch_size=1,
48 max_seq_len=1024,
49 **llm_kwargs)
50
51 outputs = llm.generate(prompts, sampling_params)
52
53 # Print the outputs.
54 for output in outputs:
55 prompt = output.prompt
56 generated_text = output.outputs[0].text
57 print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
58
59
60if __name__ == '__main__':
61 main()