| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import numpy as np |
| 5 | |
| 6 | from vllm.distributed.aux_output_connector.connector import ( |
| 7 | AuxRequestOutput, |
| 8 | ) |
| 9 | from vllm.v1.executor.ray_utils import detach_zero_copy_from_model_runner_output |
| 10 | from vllm.v1.outputs import ( |
| 11 | LogprobsLists, |
| 12 | LogprobsTensors, |
| 13 | ModelRunnerOutput, |
| 14 | ) |
| 15 | |
| 16 | |
| 17 | def _make_readonly(arr: np.ndarray) -> np.ndarray: |
| 18 | arr.setflags(write=False) |
| 19 | return arr |
| 20 | |
| 21 | |
| 22 | def test_detach_zero_copy_from_model_runner_output_copies_only_numpy_views(): |
| 23 | cu_num_generated_tokens = [0, 2] |
| 24 | prompt_logprobs = LogprobsTensors.empty_cpu(1, 2) |
| 25 | output = ModelRunnerOutput( |
| 26 | req_ids=["req-0"], |
| 27 | req_id_to_index={"req-0": 0}, |
| 28 | logprobs=LogprobsLists( |
| 29 | logprob_token_ids=_make_readonly( |
| 30 | np.array([[1, 2], [3, 4]], dtype=np.int32) |
| 31 | ), |
| 32 | logprobs=_make_readonly( |
| 33 | np.array([[0.1, 0.2], [0.3, 0.4]], dtype=np.float32) |
| 34 | ), |
| 35 | sampled_token_ranks=_make_readonly(np.array([1, 2], dtype=np.int32)), |
| 36 | cu_num_generated_tokens=cu_num_generated_tokens, |
| 37 | ), |
| 38 | prompt_logprobs_dict={"req-0": prompt_logprobs}, |
| 39 | ) |
| 40 | |
| 41 | original_logprobs = output.logprobs |
| 42 | assert original_logprobs is not None |
| 43 | |
| 44 | detach_zero_copy_from_model_runner_output(output) |
| 45 | |
| 46 | detached_logprobs = output.logprobs |
| 47 | assert detached_logprobs is not None |
| 48 | assert detached_logprobs is not original_logprobs |
| 49 | assert ( |
| 50 | detached_logprobs.logprob_token_ids is not original_logprobs.logprob_token_ids |
| 51 | ) |
| 52 | assert detached_logprobs.logprobs is not original_logprobs.logprobs |
| 53 | assert ( |
| 54 | detached_logprobs.sampled_token_ranks |
| 55 | is not original_logprobs.sampled_token_ranks |
| 56 | ) |
| 57 | assert detached_logprobs.logprob_token_ids.flags.writeable |
| 58 | assert detached_logprobs.logprobs.flags.writeable |
| 59 | assert detached_logprobs.sampled_token_ranks.flags.writeable |
| 60 | assert detached_logprobs.cu_num_generated_tokens is cu_num_generated_tokens |
| 61 | assert output.prompt_logprobs_dict["req-0"] is prompt_logprobs |
| 62 | |
| 63 | |
| 64 | def test_detach_zero_copy_aux_output_without_logprobs(): |
| 65 | rows = _make_readonly(np.arange(12, dtype=np.uint8).reshape(2, 3, 2)) |
| 66 | output = ModelRunnerOutput( |
| 67 | req_ids=["req-0"], |
| 68 | req_id_to_index={"req-0": 0}, |
| 69 | aux_output_connector_output={"req-0": AuxRequestOutput(0, rows)}, |
| 70 | ) |
| 71 | |
| 72 | detach_zero_copy_from_model_runner_output(output) |
| 73 | |
| 74 | aux_output = output.aux_output_connector_output |
| 75 | assert aux_output is not None |
| 76 | detached = aux_output["req-0"].rows |
| 77 | assert detached is not rows |
| 78 | assert detached.flags.writeable |
| 79 | np.testing.assert_array_equal(detached, rows) |