|
| 1 | +# |
| 2 | +# Copyright (c) 2025 Huawei Technologies Co., Ltd. All Rights Reserved. |
| 3 | +# This file is a part of the vllm-ascend project. |
| 4 | +# |
| 5 | +# Licensed under the Apache License, Version 2.0 (the "License"); |
| 6 | +# you may not use this file except in compliance with the License. |
| 7 | +# You may obtain a copy of the License at |
| 8 | +# |
| 9 | +# http://www.apache.org/licenses/LICENSE-2.0 |
| 10 | +# |
| 11 | +# Unless required by applicable law or agreed to in writing, software |
| 12 | +# distributed under the License is distributed on an "AS IS" BASIS, |
| 13 | +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 14 | +# See the License for the specific language governing permissions and |
| 15 | +# limitations under the License. |
| 16 | +import pytest |
| 17 | + |
| 18 | +from tests.conftest import VllmRunner |
| 19 | +from vllm_ascend.ascend_config import clear_ascend_config, get_ascend_config |
| 20 | + |
| 21 | + |
| 22 | +def _clean_up_ascend_config(func): |
| 23 | + |
| 24 | + def wrapper(*args, **kwargs): |
| 25 | + clear_ascend_config() |
| 26 | + func(*args, **kwargs) |
| 27 | + clear_ascend_config() |
| 28 | + |
| 29 | + return wrapper |
| 30 | + |
| 31 | + |
| 32 | +@_clean_up_ascend_config |
| 33 | +def test_run_without_ascend_config(): |
| 34 | + with VllmRunner("facebook/opt-125m"): |
| 35 | + ascend_config = get_ascend_config() |
| 36 | + |
| 37 | + assert not ascend_config.torchair_graph_config.enabled |
| 38 | + assert not ascend_config.torchair_graph_config.use_cached_graph |
| 39 | + assert ascend_config.torchair_graph_config.graph_batch_sizes == [] |
| 40 | + assert not ascend_config.torchair_graph_config.graph_batch_sizes_init |
| 41 | + assert not ascend_config.ascend_scheduler_config.enabled |
| 42 | + assert ascend_config.expert_tensor_parallel_size == 1 |
| 43 | + |
| 44 | + |
| 45 | +@_clean_up_ascend_config |
| 46 | +def test_run_with_ascend_config(): |
| 47 | + input_additional_config = { |
| 48 | + "torchair_graph_config": { |
| 49 | + # torchair graph only works with deepseek. The e2e test should be added |
| 50 | + # in multicard test with deepseek models. |
| 51 | + "enabled": False, |
| 52 | + "use_cached_graph": True, |
| 53 | + "graph_batch_sizes": [1, 2, 4, 8], |
| 54 | + "graph_batch_sizes_init": False, |
| 55 | + }, |
| 56 | + "ascend_scheduler_config": { |
| 57 | + "enabled": True, |
| 58 | + "enable_chunked_prefill": True, |
| 59 | + }, |
| 60 | + "expert_tensor_parallel_size": 1 |
| 61 | + } |
| 62 | + with VllmRunner("facebook/opt-125m", |
| 63 | + additional_config=input_additional_config): |
| 64 | + ascend_config = get_ascend_config() |
| 65 | + |
| 66 | + assert ascend_config.is_initialized |
| 67 | + assert not ascend_config.torchair_graph_config.enabled |
| 68 | + assert ascend_config.torchair_graph_config.use_cached_graph |
| 69 | + assert ascend_config.torchair_graph_config.graph_batch_sizes == [ |
| 70 | + 1, 2, 4, 8 |
| 71 | + ] |
| 72 | + assert not ascend_config.torchair_graph_config.graph_batch_sizes_init |
| 73 | + assert ascend_config.ascend_scheduler_config.enabled |
| 74 | + assert ascend_config.ascend_scheduler_config.enable_chunked_prefill |
| 75 | + assert ascend_config.expert_tensor_parallel_size == 1 |
| 76 | + |
| 77 | + |
| 78 | +@_clean_up_ascend_config |
| 79 | +def test_ascend_config_init_error(): |
| 80 | + # ascend_config should be initialized first |
| 81 | + with pytest.raises(RuntimeError): |
| 82 | + _ = get_ascend_config() |
| 83 | + |
| 84 | + |
| 85 | +@_clean_up_ascend_config |
| 86 | +def test_ascend_config_load_error(): |
| 87 | + # graph_batch_sizes should be list. |
| 88 | + with pytest.raises(TypeError): |
| 89 | + input_additional_config_fake_1 = { |
| 90 | + "torchair_graph_config": { |
| 91 | + "graph_batch_sizes": "fake_size", |
| 92 | + }, |
| 93 | + } |
| 94 | + with VllmRunner("facebook/opt-125m", |
| 95 | + additional_config=input_additional_config_fake_1): |
| 96 | + pass |
| 97 | + |
| 98 | + # graph_batch_sizes_init should not be True when graph_batch_sizes is not empty. |
| 99 | + with pytest.raises(ValueError): |
| 100 | + input_additional_config_fake_2 = { |
| 101 | + "torchair_graph_config": { |
| 102 | + "graph_batch_sizes": [1, 2, 4, 8], |
| 103 | + "graph_batch_sizes_init": True, |
| 104 | + }, |
| 105 | + } |
| 106 | + with VllmRunner("facebook/opt-125m", |
| 107 | + additional_config=input_additional_config_fake_2): |
| 108 | + pass |
| 109 | + |
| 110 | + # torchair graph only works with deepseek. |
| 111 | + with pytest.raises(NotImplementedError): |
| 112 | + input_additional_config_fake_2 = { |
| 113 | + "torchair_graph_config": { |
| 114 | + "enabled": True, |
| 115 | + }, |
| 116 | + } |
| 117 | + with VllmRunner("facebook/opt-125m", |
| 118 | + additional_config=input_additional_config_fake_2): |
| 119 | + pass |
0 commit comments