| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import pytest |
| 4 | import torch |
| 5 | |
| 6 | from vllm.config import SpeculativeConfig |
| 7 | from vllm.model_executor.models.interfaces import supports_eagle3 |
| 8 | from vllm.platforms import current_platform |
| 9 | |
| 10 | |
| 11 | @pytest.mark.parametrize( |
| 12 | "model_path", |
| 13 | [ |
| 14 | pytest.param( |
| 15 | "nm-testing/SpeculatorLlama3-1-8B-Eagle3-converted-0717-quantized", |
| 16 | id="llama3-eagle3-speculator", |
| 17 | ), |
| 18 | pytest.param( |
| 19 | "nm-testing/Speculator-Qwen3-8B-Eagle3-converted-071-quantized", |
| 20 | id="qwen3-eagle3-speculator", |
| 21 | ), |
| 22 | pytest.param( |
| 23 | "nm-testing/Speculator-Qwen3-8B-Eagle3-converted-071-quantized-w4a16", |
| 24 | id="qwen3-eagle3-speculator-w4a16-verifier", |
| 25 | marks=pytest.mark.skipif( |
| 26 | current_platform.is_rocm(), |
| 27 | reason="The tests are skipped on rocm platform.", |
| 28 | ), |
| 29 | ), |
| 30 | ], |
| 31 | ) |
| 32 | def test_eagle3_speculators_model( |
| 33 | vllm_runner, example_prompts, model_path, monkeypatch |
| 34 | ): |
| 35 | """Test Eagle3 speculators models properly initialize speculative decoding. |
| 36 | |
| 37 | This test verifies: |
| 38 | 1. Eagle3 support is detected for the model |
| 39 | 2. Speculative config is automatically initialized from embedded config |
| 40 | 3. The draft model path is correctly set to the speculators model |
| 41 | 4. Speculative tokens count is valid |
| 42 | 5. Text generation works with speculative decoding enabled |
| 43 | """ |
| 44 | # Set environment variable for V1 engine serialization |
| 45 | monkeypatch.setenv("VLLM_ALLOW_INSECURE_SERIALIZATION", "1") |
| 46 | |
| 47 | with vllm_runner(model_path, dtype=torch.bfloat16) as vllm_model: |
| 48 | # Verify Eagle3 support is detected |
| 49 | eagle3_supported = vllm_model.apply_model(supports_eagle3) |
| 50 | assert eagle3_supported, f"Eagle3 should be supported for {model_path}" |
| 51 | |
| 52 | vllm_config = vllm_model.llm.llm_engine.vllm_config |
| 53 | |
| 54 | assert isinstance(vllm_config.speculative_config, SpeculativeConfig), ( |
| 55 | "Speculative config should be initialized for speculators model" |
| 56 | ) |
| 57 | |
| 58 | spec_config = vllm_config.speculative_config |
| 59 | assert spec_config.num_speculative_tokens > 0, ( |
| 60 | f"Expected positive speculative tokens, " |
| 61 | f"got {spec_config.num_speculative_tokens}" |
| 62 | ) |
| 63 | |
| 64 | assert spec_config.model == model_path, ( |
| 65 | f"Draft model should be {model_path}, got {spec_config.model}" |
| 66 | ) |
| 67 | |
| 68 | vllm_outputs = vllm_model.generate_greedy(example_prompts, max_tokens=20) |
| 69 | assert vllm_outputs, f"No outputs generated for speculators model {model_path}" |