| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import pytest |
| 5 | import torch |
| 6 | |
| 7 | from vllm.platforms import current_platform |
| 8 | |
| 9 | if current_platform.is_rocm(): |
| 10 | pytest.skip( |
| 11 | reason="FlashInfer GDN prefill is not supported on ROCm.", |
| 12 | allow_module_level=True, |
| 13 | ) |
| 14 | |
| 15 | import flashinfer.gdn_prefill # noqa: E402 |
| 16 | |
| 17 | from vllm.model_executor.layers.mamba.gdn.qwen_gdn_linear_attn import ( |
| 18 | fi_chunk_gated_delta_rule, |
| 19 | ) # noqa: E402 |
| 20 | |
| 21 | |
| 22 | def test_flashinfer_gdn_prefill_uses_int64_cu_seqlens(monkeypatch): |
| 23 | captured_cu_seqlens = None |
| 24 | |
| 25 | def fake_chunk_gated_delta_rule(**kwargs): |
| 26 | nonlocal captured_cu_seqlens |
| 27 | captured_cu_seqlens = kwargs["cu_seqlens"] |
| 28 | return kwargs["q"] |
| 29 | |
| 30 | monkeypatch.setattr( |
| 31 | flashinfer.gdn_prefill, |
| 32 | "chunk_gated_delta_rule", |
| 33 | fake_chunk_gated_delta_rule, |
| 34 | ) |
| 35 | q = torch.zeros(1, 2, 1, 2) |
| 36 | cu_seqlens = torch.tensor([0, 2], dtype=torch.int32) |
| 37 | |
| 38 | output, final_state = fi_chunk_gated_delta_rule( |
| 39 | q=q, |
| 40 | k=q, |
| 41 | v=q, |
| 42 | g=torch.zeros(1, 2, 1), |
| 43 | beta=torch.zeros(1, 2, 1), |
| 44 | initial_state=torch.zeros(1, 1, 2, 2), |
| 45 | output_final_state=False, |
| 46 | cu_seqlens=cu_seqlens, |
| 47 | use_qk_l2norm_in_kernel=False, |
| 48 | ) |
| 49 | |
| 50 | assert captured_cu_seqlens is not None |
| 51 | assert captured_cu_seqlens.dtype == torch.int64 |
| 52 | assert output.shape == q.shape |
| 53 | assert final_state is None |