| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for CUDA kernels in cache_kernels.cu.""" |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | |
| 8 | try: |
| 9 | from vllm import _custom_ops as ops |
| 10 | except ImportError: |
| 11 | pytest.skip( |
| 12 | "Could not import vllm._custom_ops. (pip install -e .)", allow_module_level=True |
| 13 | ) |
| 14 | |
| 15 | |
| 16 | @pytest.mark.skipif(torch.accelerator.device_count() < 1, reason="Need CUDA device") |
| 17 | def test_gather_cache_oob(): |
| 18 | """Tests for OOB read in gather_and_maybe_dequant_cache (Issue #27909). |
| 19 | This test constructs a boundary case identified in the issue where |
| 20 | seq_starts causes the block_table offset to read out of bounds. |
| 21 | """ |
| 22 | block_size = 64 |
| 23 | # The kernel only supports the MLA entry sizes. |
| 24 | entry_size = 576 |
| 25 | |
| 26 | block_table = torch.tensor([[1, 2]], dtype=torch.int32, device="cuda") |
| 27 | |
| 28 | # This will result in offset = 128 / block_size = 128 / 64 = 2 |
| 29 | # This will cause the kernel to try to read from |
| 30 | # block_table[0, 2], but its size is only 2. |
| 31 | seq_starts = torch.tensor([128], dtype=torch.int32, device="cuda") |
| 32 | |
| 33 | seq_len = 65 |
| 34 | cu_seq_lens = torch.tensor([0, seq_len], dtype=torch.int32, device="cuda") |
| 35 | token_to_seq = torch.zeros(seq_len, dtype=torch.int32, device="cuda") |
| 36 | |
| 37 | # src_cache: [num_blocks, block_size, entry_size] |
| 38 | num_blocks = 5 |
| 39 | src_cache = torch.randn( |
| 40 | (num_blocks, block_size, entry_size), dtype=torch.float16, device="cuda" |
| 41 | ) |
| 42 | |
| 43 | dst = torch.empty((seq_len, entry_size), dtype=torch.float16, device="cuda") |
| 44 | |
| 45 | scale = torch.tensor([1.0], dtype=torch.float32, device="cuda") |
| 46 | |
| 47 | # Calling the C++ function gather_and_maybe_dequant_cache |
| 48 | ops.gather_and_maybe_dequant_cache( |
| 49 | src_cache, |
| 50 | dst, |
| 51 | block_table, |
| 52 | cu_seq_lens, |
| 53 | token_to_seq, |
| 54 | seq_len, |
| 55 | "auto", # kv_cache_dtype |
| 56 | scale, |
| 57 | seq_starts, |
| 58 | ) |
| 59 | |
| 60 | torch.accelerator.synchronize() |
| 61 | assert True |
| 62 | |
| 63 | |
| 64 | if __name__ == "__main__": |
| 65 | pytest.main([__file__]) |