|
| 1 | +from types import SimpleNamespace |
| 2 | +from unittest.mock import MagicMock, Mock, patch |
| 3 | + |
| 4 | +import pytest |
| 5 | +import torch |
| 6 | +from transformers import PretrainedConfig |
| 7 | +from vllm.config import CacheConfig, EPLBConfig, ParallelConfig |
| 8 | +from vllm.distributed.parallel_state import GroupCoordinator |
| 9 | + |
| 10 | + |
| 11 | +@pytest.fixture |
| 12 | +def base_config(): |
| 13 | + config = PretrainedConfig( |
| 14 | + hidden_size=128, |
| 15 | + num_attention_heads=8, |
| 16 | + num_hidden_layers=2, |
| 17 | + intermediate_size=256, |
| 18 | + hidden_act="silu", |
| 19 | + rms_norm_eps=1e-6, |
| 20 | + rope_theta=10000.0, |
| 21 | + max_position_embeddings=2048, |
| 22 | + n_routed_experts=4, |
| 23 | + n_shared_experts=1, |
| 24 | + moe_intermediate_size=256, |
| 25 | + num_experts_per_tok=2, |
| 26 | + routed_scaling_factor=1.0, |
| 27 | + first_k_dense_replace=0, |
| 28 | + moe_layer_freq=1, |
| 29 | + kv_lora_rank=16, |
| 30 | + qk_nope_head_dim=16, |
| 31 | + qk_rope_head_dim=16, |
| 32 | + v_head_dim=32, |
| 33 | + topk_method="noaux_tc", |
| 34 | + scoring_func="softmax", |
| 35 | + norm_topk_prob=True, |
| 36 | + n_group=1, |
| 37 | + topk_group=1, |
| 38 | + vocab_size=10000, |
| 39 | + ) |
| 40 | + return config |
| 41 | + |
| 42 | + |
| 43 | +@pytest.fixture |
| 44 | +def vllm_config(base_config): |
| 45 | + model_config = SimpleNamespace( |
| 46 | + hf_config=base_config, |
| 47 | + tensor_parallel_size=1, |
| 48 | + dtype=torch.float32, |
| 49 | + use_mla=True, |
| 50 | + quant_config=None, |
| 51 | + max_model_len=2048, |
| 52 | + ) |
| 53 | + parallel_config = MagicMock(spec=ParallelConfig) |
| 54 | + eplb_config = MagicMock(spec=EPLBConfig) |
| 55 | + eplb_config.num_redundant_experts = 0 |
| 56 | + parallel_config.eplb_config = eplb_config |
| 57 | + |
| 58 | + cache_config = CacheConfig() |
| 59 | + vllm_config = Mock() |
| 60 | + vllm_config.model_config = model_config |
| 61 | + vllm_config.cache_config = cache_config |
| 62 | + vllm_config.quant_config = None |
| 63 | + vllm_config.parallel_config = parallel_config |
| 64 | + return vllm_config |
| 65 | + |
| 66 | + |
| 67 | +@pytest.fixture |
| 68 | +def mock_distributed(): |
| 69 | + tp_group = Mock(spec=GroupCoordinator) |
| 70 | + tp_group.rank_in_group = 0 |
| 71 | + tp_group.world_size = 1 |
| 72 | + tp_group.device_group = Mock() |
| 73 | + |
| 74 | + dp_group = Mock(spec=GroupCoordinator) |
| 75 | + dp_group.rank_in_group = 0 |
| 76 | + dp_group.world_size = 1 |
| 77 | + |
| 78 | + ep_group = Mock(spec=GroupCoordinator) |
| 79 | + ep_group.rank_in_group = 0 |
| 80 | + ep_group.world_size = 1 |
| 81 | + ep_group.device_group = Mock() |
| 82 | + ep_group.device_group.rank.return_value = 0 |
| 83 | + ep_group.device_group.size.return_value = 1 |
| 84 | + |
| 85 | + pp_group = Mock(spec=GroupCoordinator) |
| 86 | + pp_group.rank_in_group = 0 |
| 87 | + pp_group.world_size = 1 |
| 88 | + |
| 89 | + mock_vllm_config = Mock() |
| 90 | + mock_vllm_config.scheduler_config = Mock(max_num_seqs=256) |
| 91 | + mock_vllm_config.model_config = Mock(max_model_len=2048, quant_config=None) |
| 92 | + |
| 93 | + with patch("vllm_ascend.models.deepseek_v2.get_tensor_model_parallel_rank", return_value=0), \ |
| 94 | + patch("vllm_ascend.models.deepseek_v2.get_tensor_model_parallel_world_size", return_value=1), \ |
| 95 | + patch("vllm_ascend.models.deepseek_v2.get_tp_group", return_value=tp_group), \ |
| 96 | + patch("vllm_ascend.models.deepseek_v2.get_pp_group", return_value=pp_group), \ |
| 97 | + patch("vllm_ascend.models.deepseek_v2.get_pp_group", |
| 98 | + return_value=Mock(is_first_rank=False, is_last_rank=False)), \ |
| 99 | + patch("vllm_ascend.ops.fused_moe.get_current_vllm_config", return_value=mock_vllm_config), \ |
| 100 | + patch("vllm_ascend.ops.moe.token_dispatcher.torch.distributed.get_rank", return_value=0), \ |
| 101 | + patch("vllm_ascend.ops.moe.token_dispatcher.get_ascend_soc_version", return_value=None), \ |
| 102 | + patch.dict("vllm.distributed.parallel_state.__dict__", _TP=tp_group, _EP=ep_group, _DP=dp_group, |
| 103 | + _PP=pp_group), \ |
| 104 | + patch.dict("vllm_ascend.distributed.parallel_state.__dict__", _MC2=ep_group), \ |
| 105 | + patch("torch.npu.current_device", return_value=0): |
| 106 | + yield |
| 107 | + |
| 108 | + |
| 109 | +@pytest.fixture |
| 110 | +def mock_forward_context(): |
| 111 | + forward_context = Mock(in_profile_run=False, with_prefill=False) |
| 112 | + with patch("vllm_ascend.models.deepseek_v2.get_forward_context", |
| 113 | + return_value=forward_context): |
| 114 | + yield |
0 commit comments