| 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 | import torch |
| 7 | |
| 8 | |
| 9 | def test_nemotron_h_lm_head_receives_quant_config(): |
| 10 | from vllm.model_executor.models.nemotron_h import NemotronHForCausalLM |
| 11 | |
| 12 | mock_quant_config = Mock() |
| 13 | |
| 14 | mock_hf_config = Mock() |
| 15 | mock_hf_config.vocab_size = 128 |
| 16 | mock_hf_config.hidden_size = 64 |
| 17 | |
| 18 | mock_vllm_config = Mock() |
| 19 | mock_vllm_config.model_config.hf_config = mock_hf_config |
| 20 | mock_vllm_config.model_config.dtype = None |
| 21 | mock_vllm_config.scheduler_config = Mock() |
| 22 | mock_vllm_config.quant_config = mock_quant_config |
| 23 | |
| 24 | with ( |
| 25 | patch("vllm.model_executor.models.nemotron_h.NemotronHModel") as MockModel, |
| 26 | patch("vllm.model_executor.models.nemotron_h.ParallelLMHead") as MockLMHead, |
| 27 | patch("vllm.model_executor.models.nemotron_h.LogitsProcessor"), |
| 28 | ): |
| 29 | MockModel.return_value.make_empty_intermediate_tensors = Mock() |
| 30 | MockModel.return_value.has_moe = False |
| 31 | |
| 32 | NemotronHForCausalLM(vllm_config=mock_vllm_config) |
| 33 | |
| 34 | MockLMHead.assert_called_once() |
| 35 | call_kwargs = MockLMHead.call_args.kwargs |
| 36 | assert call_kwargs["quant_config"] is mock_quant_config |
| 37 | |
| 38 | |
| 39 | def test_relu2_fp8_fusion_uses_registry(): |
| 40 | from vllm.model_executor.models.nemotron_h import NemotronHMLP |
| 41 | |
| 42 | projected = torch.empty((1, 1), dtype=torch.bfloat16) |
| 43 | fused = Mock() |
| 44 | act_fn = Mock() |
| 45 | down_proj = Mock(side_effect=lambda x: (x, None)) |
| 46 | |
| 47 | mlp = NemotronHMLP.__new__(NemotronHMLP) |
| 48 | torch.nn.Module.__init__(mlp) |
| 49 | mlp.up_proj = Mock(return_value=(projected, None)) |
| 50 | mlp.down_proj = down_proj |
| 51 | mlp.act_fn = act_fn |
| 52 | |
| 53 | with patch( |
| 54 | "vllm.model_executor.models.nemotron_h.maybe_fused_act_quant", |
| 55 | return_value=fused, |
| 56 | ) as maybe_fused: |
| 57 | result = mlp(Mock()) |
| 58 | |
| 59 | maybe_fused.assert_called_once_with(act_fn, projected, down_proj) |
| 60 | act_fn.assert_not_called() |
| 61 | assert result is fused |