| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from unittest.mock import Mock, patch |
| 5 | |
| 6 | |
| 7 | def test_qwen3_5_lm_head_receives_quant_config(): |
| 8 | from vllm.model_executor.models.qwen3_5 import Qwen3_5ForCausalLMBase |
| 9 | |
| 10 | mock_quant_config = Mock() |
| 11 | |
| 12 | mock_hf_config = Mock() |
| 13 | mock_hf_config.tie_word_embeddings = False |
| 14 | mock_hf_config.vocab_size = 128 |
| 15 | mock_hf_config.hidden_size = 64 |
| 16 | |
| 17 | mock_vllm_config = Mock() |
| 18 | mock_vllm_config.model_config.hf_text_config = mock_hf_config |
| 19 | mock_vllm_config.cache_config.mamba_cache_mode = "align" |
| 20 | mock_vllm_config.scheduler_config = Mock() |
| 21 | mock_vllm_config.quant_config = mock_quant_config |
| 22 | mock_vllm_config.lora_config = None |
| 23 | |
| 24 | mock_pp_group = Mock() |
| 25 | mock_pp_group.is_last_rank = True |
| 26 | |
| 27 | with ( |
| 28 | patch("vllm.model_executor.models.qwen3_5.Qwen3_5Model") as MockModel, |
| 29 | patch("vllm.model_executor.models.qwen3_5.ParallelLMHead") as MockLMHead, |
| 30 | patch("vllm.model_executor.models.qwen3_5.LogitsProcessor"), |
| 31 | patch( |
| 32 | "vllm.model_executor.models.qwen3_5.get_pp_group", |
| 33 | return_value=mock_pp_group, |
| 34 | ), |
| 35 | ): |
| 36 | MockModel.return_value.make_empty_intermediate_tensors = Mock() |
| 37 | |
| 38 | Qwen3_5ForCausalLMBase(vllm_config=mock_vllm_config) |
| 39 | |
| 40 | MockLMHead.assert_called_once() |
| 41 | call_kwargs = MockLMHead.call_args.kwargs |
| 42 | assert call_kwargs["quant_config"] is mock_quant_config |
| 43 | |
| 44 | |
| 45 | def test_qwen3_5_mtp_lm_head_receives_quant_config(): |
| 46 | from vllm.config import CompilationMode |
| 47 | from vllm.model_executor.models.qwen3_5_mtp import Qwen3_5MTP |
| 48 | |
| 49 | mock_quant_config = Mock() |
| 50 | |
| 51 | mock_hf_config = Mock() |
| 52 | mock_hf_config.tie_word_embeddings = False |
| 53 | mock_hf_config.vocab_size = 128 |
| 54 | mock_hf_config.hidden_size = 64 |
| 55 | |
| 56 | mock_vllm_config = Mock() |
| 57 | mock_vllm_config.model_config.hf_text_config = mock_hf_config |
| 58 | mock_vllm_config.cache_config.mamba_cache_mode = "align" |
| 59 | mock_vllm_config.compilation_config.mode = CompilationMode.NONE |
| 60 | mock_vllm_config.quant_config = mock_quant_config |
| 61 | |
| 62 | mock_pp_group = Mock() |
| 63 | mock_pp_group.is_last_rank = True |
| 64 | |
| 65 | with ( |
| 66 | patch("vllm.model_executor.models.qwen3_5_mtp.Qwen3_5MultiTokenPredictor"), |
| 67 | patch("vllm.model_executor.models.qwen3_5_mtp.ParallelLMHead") as MockLMHead, |
| 68 | patch("vllm.model_executor.models.qwen3_5_mtp.LogitsProcessor"), |
| 69 | patch( |
| 70 | "vllm.model_executor.models.qwen3_5_mtp.get_pp_group", |
| 71 | return_value=mock_pp_group, |
| 72 | ), |
| 73 | ): |
| 74 | Qwen3_5MTP(vllm_config=mock_vllm_config) |
| 75 | |
| 76 | MockLMHead.assert_called_once() |
| 77 | call_kwargs = MockLMHead.call_args.kwargs |
| 78 | assert call_kwargs["quant_config"] is mock_quant_config |