| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import torch |
| 4 | |
| 5 | from vllm.multimodal.inputs import ( |
| 6 | MultiModalBatchedField, |
| 7 | MultiModalFeatureSpec, |
| 8 | MultiModalFieldElem, |
| 9 | MultiModalKwargsItem, |
| 10 | PlaceholderRange, |
| 11 | ) |
| 12 | from vllm.multimodal.utils import strip_covered_mm_data |
| 13 | from vllm.v1.core.sched.output import NewRequestData |
| 14 | |
| 15 | |
| 16 | def _create_new_requests_data(prompt_embeds: torch.Tensor | None) -> NewRequestData: |
| 17 | return NewRequestData( |
| 18 | req_id="test_req", |
| 19 | prompt_token_ids=None, |
| 20 | mm_features=[], |
| 21 | sampling_params=None, |
| 22 | pooling_params=None, |
| 23 | block_ids=([],), |
| 24 | num_computed_tokens=0, |
| 25 | lora_request=None, |
| 26 | prompt_embeds=prompt_embeds, |
| 27 | ) |
| 28 | |
| 29 | |
| 30 | def test_repr_with_none() -> None: |
| 31 | """Test repr when prompt_embeds is None.""" |
| 32 | new_requests_data = _create_new_requests_data(None) |
| 33 | |
| 34 | assert "prompt_embeds_shape=None" in repr(new_requests_data) |
| 35 | assert "prompt_embeds_shape=None" in new_requests_data.anon_repr() |
| 36 | |
| 37 | |
| 38 | def test_repr_with_multi_element_tensor() -> None: |
| 39 | """Test repr when prompt_embeds is a multi-element tensor.""" |
| 40 | prompt_embeds = torch.randn(10, 768) |
| 41 | new_requests_data = _create_new_requests_data(prompt_embeds) |
| 42 | |
| 43 | assert "prompt_embeds_shape=torch.Size([10, 768])" in repr(new_requests_data) |
| 44 | assert "prompt_embeds_shape=torch.Size([10, 768])" in new_requests_data.anon_repr() |
| 45 | |
| 46 | |
| 47 | def _mm_feature(offset: int, length: int) -> MultiModalFeatureSpec: |
| 48 | return MultiModalFeatureSpec( |
| 49 | data=MultiModalKwargsItem.dummy(), |
| 50 | mm_position=PlaceholderRange(offset=offset, length=length), |
| 51 | identifier=f"hash_{offset}", |
| 52 | modality="image", |
| 53 | ) |
| 54 | |
| 55 | |
| 56 | def test_strip_covered_mm_data() -> None: |
| 57 | """Items fully inside the computed prefix lose their data; items touching |
| 58 | the uncomputed region keep it; already-None data stays None.""" |
| 59 | from dataclasses import replace |
| 60 | |
| 61 | covered = _mm_feature(offset=0, length=100) |
| 62 | boundary = _mm_feature(offset=150, length=100) # ends exactly at 250 |
| 63 | uncovered = _mm_feature(offset=300, length=100) |
| 64 | already_none = replace(_mm_feature(offset=100, length=50), data=None) |
| 65 | |
| 66 | stripped = strip_covered_mm_data( |
| 67 | [covered, boundary, uncovered, already_none], num_computed_tokens=250 |
| 68 | ) |
| 69 | |
| 70 | assert stripped[0].data is None # fully covered -> stripped |
| 71 | assert stripped[1].data is None # span end == computed -> covered -> stripped |
| 72 | assert stripped[2].data is not None # extends past prefix -> kept |
| 73 | assert stripped[3].data is None # was already None |
| 74 | # non-data fields are preserved |
| 75 | assert stripped[0].identifier == covered.identifier |
| 76 | assert stripped[0].mm_position == covered.mm_position |
| 77 | # original list is not mutated |
| 78 | assert covered.data is not None |
| 79 | |
| 80 | |
| 81 | def test_strip_covered_mm_data_zero_computed() -> None: |
| 82 | """With no prefix hit nothing is stripped.""" |
| 83 | features = [_mm_feature(offset=0, length=100)] |
| 84 | stripped = strip_covered_mm_data(features, num_computed_tokens=0) |
| 85 | assert stripped[0].data is not None |
| 86 | |
| 87 | |
| 88 | def _mm_feature_mixed(offset: int, length: int) -> MultiModalFeatureSpec: |
| 89 | data = MultiModalKwargsItem( |
| 90 | { |
| 91 | "pixel_values": MultiModalFieldElem( |
| 92 | data=torch.empty(4), field=MultiModalBatchedField() |
| 93 | ), |
| 94 | "image_grid_thw": MultiModalFieldElem( |
| 95 | data=torch.ones(1, 3, dtype=torch.long), |
| 96 | field=MultiModalBatchedField(keep_on_cpu=True), |
| 97 | ), |
| 98 | } |
| 99 | ) |
| 100 | return MultiModalFeatureSpec( |
| 101 | data=data, |
| 102 | mm_position=PlaceholderRange(offset=offset, length=length), |
| 103 | identifier=f"hash_{offset}", |
| 104 | modality="image", |
| 105 | ) |
| 106 | |
| 107 | |
| 108 | def test_strip_covered_mm_data_mrope() -> None: |
| 109 | """For M-RoPE models, covered items keep their keep_on_cpu metadata fields |
| 110 | (the worker needs them to compute positions); payload fields are dropped.""" |
| 111 | covered = _mm_feature_mixed(offset=0, length=100) |
| 112 | uncovered = _mm_feature_mixed(offset=300, length=100) |
| 113 | |
| 114 | stripped = strip_covered_mm_data( |
| 115 | [covered, uncovered], num_computed_tokens=250, uses_mrope=True |
| 116 | ) |
| 117 | |
| 118 | assert stripped[0].data is not None |
| 119 | assert list(stripped[0].data.keys()) == ["image_grid_thw"] |
| 120 | assert stripped[1].data is not None |
| 121 | assert set(stripped[1].data.keys()) == {"pixel_values", "image_grid_thw"} |
| 122 | # original list is not mutated |
| 123 | assert set(covered.data.keys()) == {"pixel_values", "image_grid_thw"} |
| 124 | |
| 125 | |
| 126 | def test_strip_covered_mm_data_shm_address_item() -> None: |
| 127 | """SHM address descriptors must survive stripping even when covered: the |
| 128 | worker needs the address to resolve the payload and to acknowledge the |
| 129 | sender's writer reference. Stripping it crashed the SHM receiver cache on |
| 130 | the second identical request (vllm-project/vllm#54994).""" |
| 131 | address_item = MultiModalKwargsItem( |
| 132 | { |
| 133 | "address": MultiModalFieldElem(data=4096, field=MultiModalBatchedField()), |
| 134 | "monotonic_id": MultiModalFieldElem(data=7, field=MultiModalBatchedField()), |
| 135 | } |
| 136 | ) |
| 137 | feature = MultiModalFeatureSpec( |
| 138 | data=address_item, |
| 139 | mm_position=PlaceholderRange(offset=0, length=100), |
| 140 | identifier="shm_item", |
| 141 | modality="image", |
| 142 | ) |
| 143 | |
| 144 | stripped = strip_covered_mm_data([feature], num_computed_tokens=250) |
| 145 | |
| 146 | assert stripped[0].data is address_item |