| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import os |
| 4 | import tempfile |
| 5 | |
| 6 | from vllm import LLM, SamplingParams |
| 7 | from vllm.config.kv_transfer import KVTransferConfig |
| 8 | from vllm.distributed.kv_transfer.kv_connector.v1 import ( |
| 9 | example_hidden_states_connector, |
| 10 | ) |
| 11 | |
| 12 | # NOTE: If changing the interface of the ExampleHiddenStatesConnector, please also |
| 13 | # update the benchmark in benchmarks/benchmark_hidden_state_extraction.py |
| 14 | # and the docs in docs/features/speculative_decoding/extract_hidden_states.md |
| 15 | |
| 16 | # Example: Using the custom "extract_hidden_states" speculator method and |
| 17 | # ExampleHiddenStatesConnector to extract and save hidden states from vllm |
| 18 | |
| 19 | with tempfile.TemporaryDirectory() as tmpdirname: |
| 20 | llm = LLM( |
| 21 | model="Qwen/Qwen3-8B", # Your target model |
| 22 | speculative_config={ |
| 23 | "method": "extract_hidden_states", |
| 24 | "num_speculative_tokens": 1, |
| 25 | "draft_model_config": { |
| 26 | "hf_config": { |
| 27 | "eagle_aux_hidden_state_layer_ids": [ # Target model layer indices |
| 28 | 1, |
| 29 | 2, |
| 30 | 3, |
| 31 | 4, |
| 32 | ], |
| 33 | }, |
| 34 | }, |
| 35 | }, |
| 36 | kv_transfer_config=KVTransferConfig( |
| 37 | kv_connector="ExampleHiddenStatesConnector", |
| 38 | kv_role="kv_producer", |
| 39 | kv_connector_extra_config={ |
| 40 | "shared_storage_path": tmpdirname, |
| 41 | "allow_custom_save_path": True, |
| 42 | }, |
| 43 | ), |
| 44 | ) |
| 45 | |
| 46 | prompts = ["Generate a sentence with hidden states", "Write a python function"] |
| 47 | |
| 48 | # One request uses defaults, the other uses a custom save path and |
| 49 | # includes output token hidden states via per-request kv_transfer_params. |
| 50 | sampling_params_list = [ |
| 51 | SamplingParams(max_tokens=1), |
| 52 | SamplingParams( |
| 53 | max_tokens=10, |
| 54 | extra_args={ |
| 55 | "kv_transfer_params": { |
| 56 | "hidden_states_path": os.path.join( |
| 57 | tmpdirname, "custom_output.safetensors" |
| 58 | ), |
| 59 | "include_output_tokens": True, |
| 60 | } |
| 61 | }, |
| 62 | ), |
| 63 | ] |
| 64 | outputs = llm.generate(prompts, sampling_params_list) |
| 65 | |
| 66 | for output in outputs: |
| 67 | print("\nPrompt:", output.prompt) |
| 68 | print("Prompt token ids:", output.prompt_token_ids) |
| 69 | |
| 70 | hidden_states_path = output.kv_transfer_params.get("hidden_states_path") |
| 71 | assert hidden_states_path is not None |
| 72 | print("Hidden states path:", hidden_states_path) |
| 73 | |
| 74 | obj = example_hidden_states_connector.load_hidden_states(hidden_states_path) |
| 75 | token_ids = obj["token_ids"] |
| 76 | hidden_states = obj["hidden_states"] |
| 77 | |
| 78 | print("Extracted token ids:", token_ids) |
| 79 | print( |
| 80 | "Extracted hidden states shape:", hidden_states.shape |
| 81 | ) # [num_tokens, num_extracted_layers, hidden_size] |
| 82 | print("Extracted hidden states:", hidden_states) |
| 83 | |
| 84 | example_hidden_states_connector.cleanup_hidden_states(hidden_states_path) |