| 1 | # Hidden State Extraction |
| 2 | |
| 3 | The Hidden State Extraction feature allows vLLM to save intermediate layer activations from a target model during inference. This is useful for training [EAGLE](eagle.md)-style draft models, knowledge distillation, or offline analysis of model internals. |
| 4 | |
| 5 | !!! note |
| 6 | It is possible to save the last-layer's output hidden states by passing `num_hidden_layers` as a layer id. Note that these are _not_ normalized using the output norm. |
| 7 | |
| 8 | ## Offline Example |
| 9 | |
| 10 | ```python |
| 11 | import tempfile |
| 12 | |
| 13 | from vllm import LLM, SamplingParams |
| 14 | from vllm.config.kv_transfer import KVTransferConfig |
| 15 | from vllm.distributed.kv_transfer.kv_connector.v1 import ( |
| 16 | example_hidden_states_connector, |
| 17 | ) |
| 18 | |
| 19 | with tempfile.TemporaryDirectory() as tmpdir: |
| 20 | llm = LLM( |
| 21 | model="Qwen/Qwen3-8B", |
| 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": [1, 2, 3, 4], |
| 28 | }, |
| 29 | }, |
| 30 | }, |
| 31 | kv_transfer_config=KVTransferConfig( |
| 32 | kv_connector="ExampleHiddenStatesConnector", |
| 33 | kv_role="kv_producer", |
| 34 | kv_connector_extra_config={ |
| 35 | "shared_storage_path": tmpdir, |
| 36 | }, |
| 37 | ), |
| 38 | ) |
| 39 | |
| 40 | outputs = llm.generate( |
| 41 | ["The future of AI is"], |
| 42 | SamplingParams(max_tokens=1), |
| 43 | ) |
| 44 | |
| 45 | for output in outputs: |
| 46 | path = output.kv_transfer_params["hidden_states_path"] |
| 47 | obj = example_hidden_states_connector.load_hidden_states(path) |
| 48 | print(f"token_ids: {obj['token_ids'].shape}") |
| 49 | print(f"hidden_states: {obj['hidden_states'].shape}") |
| 50 | ``` |
| 51 | |
| 52 | A complete example is available at [`examples/features/speculative_decoding/extract_hidden_states_offline.py`](../../../examples/features/speculative_decoding/extract_hidden_states_offline.py). |
| 53 | |
| 54 | ## Online Example |
| 55 | |
| 56 | For improved performance, it is recommended to use a RAM-mounted file system such as `/dev/shm/` for online usage in which the client cleans up the files soon after they are generated. |
| 57 | |
| 58 | ```bash |
| 59 | vllm serve Qwen/Qwen3-8B \ |
| 60 | --speculative_config '{"method": "extract_hidden_states", "num_speculative_tokens": 1, "draft_model_config": {"hf_config": {"eagle_aux_hidden_state_layer_ids": [1, 2, 3, 4]}}}' \ |
| 61 | --kv_transfer_config '{"kv_connector": "ExampleHiddenStatesConnector", "kv_role": "kv_producer", "kv_connector_extra_config": {"shared_storage_path": "/dev/shm/hidden_states"}}' |
| 62 | ``` |
| 63 | |
| 64 | ## Per-Request Options |
| 65 | |
| 66 | Both offline and online modes support per-request options via `kv_transfer_params`: |
| 67 | |
| 68 | | Parameter | Default | Description | |
| 69 | | --- | --- | --- | |
| 70 | | `hidden_states_path` | Auto-generated | Custom file path for saving hidden states. If not set, files are saved to `<shared_storage_path>/<request_id>.safetensors`. Requires `allow_custom_save_path` to be enabled in the server config. | |
| 71 | | `include_output_tokens` | `False` | When `True`, save hidden states for both prompt and generated output tokens. When `False`, only prompt token hidden states are saved. | |
| 72 | |
| 73 | ### Offline usage |
| 74 | |
| 75 | Pass per-request options via `extra_args` on `SamplingParams`: |
| 76 | |
| 77 | ```python |
| 78 | SamplingParams( |
| 79 | max_tokens=32, |
| 80 | extra_args={ |
| 81 | "kv_transfer_params": { |
| 82 | "hidden_states_path": "/tmp/my_output.safetensors", |
| 83 | "include_output_tokens": True, |
| 84 | } |
| 85 | }, |
| 86 | ) |
| 87 | ``` |
| 88 | |
| 89 | ### Online usage |
| 90 | |
| 91 | Pass `kv_transfer_params` as a top-level field in the API request: |
| 92 | |
| 93 | ```json |
| 94 | { |
| 95 | "model": "Qwen/Qwen3-8B", |
| 96 | "messages": [{"role": "user", "content": "Hello"}], |
| 97 | "max_tokens": 32, |
| 98 | "kv_transfer_params": { |
| 99 | "hidden_states_path": "/tmp/my_output.safetensors", |
| 100 | "include_output_tokens": true |
| 101 | } |
| 102 | } |
| 103 | ``` |
| 104 | |
| 105 | ## Configuration |
| 106 | |
| 107 | The `kv_connector_extra_config` dict accepts these server-level options: |
| 108 | |
| 109 | | Parameter | Default | Description | |
| 110 | | --- | --- | --- | |
| 111 | | `shared_storage_path` | `/tmp` | Directory where hidden state files are saved (used when `hidden_states_path` is not set per-request) | |
| 112 | | `allow_custom_save_path` | `False` | Allow API clients to specify custom file paths via `hidden_states_path`. When disabled, client-provided paths are ignored with a warning. Enable only with trusted clients — custom paths can write to arbitrary locations on the server. | |
| 113 | | `num_writer_threads` | `8` | Thread pool size for async disk writes | |
| 114 | | `use_synchronization_lock` | `True` | Use file locks so concurrent readers block until writes complete. Can be disabled for batch generation where synchronization is not needed. | |
| 115 | |
| 116 | ## Output Format |
| 117 | |
| 118 | Each request produces a `.safetensors` file containing: |
| 119 | |
| 120 | - **`hidden_states`** — shape `[num_tokens, num_extracted_layers, hidden_size]` |
| 121 | - **`token_ids`** — shape `[num_tokens]` |
| 122 | |
| 123 | The file path is returned in `output.kv_transfer_params["hidden_states_path"]`. Use `load_hidden_states()` from the connector module to read the file with proper synchronization. |
| 124 | |
| 125 | !!! note |
| 126 | Chunked prefill is not compatible with this feature and must be disabled. |