| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Demonstrates how to generate prompt embeddings using |
| 4 | Hugging Face Transformers and use them as input to vLLM |
| 5 | for both single and batch inference. |
| 6 | |
| 7 | Model: meta-llama/Llama-3.2-1B-Instruct |
| 8 | Note: This model is gated on Hugging Face Hub. |
| 9 | You must request access to use it: |
| 10 | https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct |
| 11 | |
| 12 | Requirements: |
| 13 | - vLLM |
| 14 | - transformers |
| 15 | |
| 16 | Run: |
| 17 | python examples/features/prompt_embed/prompt_embed_offline.py |
| 18 | """ |
| 19 | |
| 20 | import torch |
| 21 | from transformers import AutoModelForCausalLM, AutoTokenizer, PythonBackend |
| 22 | |
| 23 | from vllm import LLM |
| 24 | |
| 25 | |
| 26 | def init_tokenizer_and_llm(model_name: str): |
| 27 | tokenizer = AutoTokenizer.from_pretrained(model_name) |
| 28 | transformers_model = AutoModelForCausalLM.from_pretrained(model_name) |
| 29 | embedding_layer = transformers_model.get_input_embeddings() |
| 30 | llm = LLM(model=model_name, enable_prompt_embeds=True) |
| 31 | return tokenizer, embedding_layer, llm |
| 32 | |
| 33 | |
| 34 | def get_prompt_embeds( |
| 35 | chat: list[dict[str, str]], |
| 36 | tokenizer: PythonBackend, |
| 37 | embedding_layer: torch.nn.Module, |
| 38 | ): |
| 39 | token_ids = tokenizer.apply_chat_template( |
| 40 | chat, add_generation_prompt=True, return_tensors="pt", return_dict=True |
| 41 | ).input_ids |
| 42 | prompt_embeds = embedding_layer(token_ids).squeeze(0) |
| 43 | return prompt_embeds |
| 44 | |
| 45 | |
| 46 | def single_prompt_inference( |
| 47 | llm: LLM, tokenizer: PythonBackend, embedding_layer: torch.nn.Module |
| 48 | ): |
| 49 | chat = [{"role": "user", "content": "Please tell me about the capital of France."}] |
| 50 | prompt_embeds = get_prompt_embeds(chat, tokenizer, embedding_layer) |
| 51 | |
| 52 | outputs = llm.generate( |
| 53 | { |
| 54 | "prompt_embeds": prompt_embeds, |
| 55 | } |
| 56 | ) |
| 57 | |
| 58 | print("\n[Single Inference Output]") |
| 59 | print("-" * 30) |
| 60 | for o in outputs: |
| 61 | print(o.outputs[0].text) |
| 62 | print("-" * 30) |
| 63 | |
| 64 | |
| 65 | def batch_prompt_inference( |
| 66 | llm: LLM, tokenizer: PythonBackend, embedding_layer: torch.nn.Module |
| 67 | ): |
| 68 | chats = [ |
| 69 | [{"role": "user", "content": "Please tell me about the capital of France."}], |
| 70 | [{"role": "user", "content": "When is the day longest during the year?"}], |
| 71 | [{"role": "user", "content": "Where is bigger, the moon or the sun?"}], |
| 72 | ] |
| 73 | |
| 74 | prompt_embeds_list = [ |
| 75 | get_prompt_embeds(chat, tokenizer, embedding_layer) for chat in chats |
| 76 | ] |
| 77 | |
| 78 | outputs = llm.generate([{"prompt_embeds": embeds} for embeds in prompt_embeds_list]) |
| 79 | |
| 80 | print("\n[Batch Inference Outputs]") |
| 81 | print("-" * 30) |
| 82 | for i, o in enumerate(outputs): |
| 83 | print(f"Q{i + 1}: {chats[i][0]['content']}") |
| 84 | print(f"A{i + 1}: {o.outputs[0].text}\n") |
| 85 | print("-" * 30) |
| 86 | |
| 87 | |
| 88 | def main(): |
| 89 | model_name = "meta-llama/Llama-3.2-1B-Instruct" |
| 90 | tokenizer, embedding_layer, llm = init_tokenizer_and_llm(model_name) |
| 91 | single_prompt_inference(llm, tokenizer, embedding_layer) |
| 92 | batch_prompt_inference(llm, tokenizer, embedding_layer) |
| 93 | |
| 94 | |
| 95 | if __name__ == "__main__": |
| 96 | main() |