| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Equivalence tests for the AITER MLA page-index expansion kernel. |
| 4 | |
| 5 | `_expand_page_indices_kernel` turns a block table into the per-token flat page |
| 6 | indices the AITER MLA decode kernel consumes (it always runs with page_size=1 |
| 7 | internally, so `kv_buffer` is flattened and each token needs its own index). |
| 8 | |
| 9 | The kernel was parallelised over token chunks as well as requests; at |
| 10 | concurrency 1 the previous `(num_reqs,)` grid put a single workgroup in a serial |
| 11 | loop over the whole sequence. These tests pin the result to a straightforward |
| 12 | PyTorch expansion so the parallelisation cannot change what is produced. |
| 13 | |
| 14 | Both block-table strides are passed by the caller, so a non-contiguous view is |
| 15 | covered explicitly rather than assumed away. |
| 16 | """ |
| 17 | |
| 18 | import pytest |
| 19 | import torch |
| 20 | |
| 21 | from vllm.v1.attention.backends.mla.rocm_aiter_mla import ( |
| 22 | _expand_page_indices_kernel, |
| 23 | ) |
| 24 | |
| 25 | |
| 26 | def _reference(block_table, cu_num_tokens, kernel_block_size, out_len, device): |
| 27 | """Per-token flat index, written the obvious way.""" |
| 28 | out = torch.full((out_len,), -1, dtype=torch.int64, device=device) |
| 29 | num_reqs = cu_num_tokens.numel() - 1 |
| 30 | for r in range(num_reqs): |
| 31 | start = int(cu_num_tokens[r]) |
| 32 | end = int(cu_num_tokens[r + 1]) |
| 33 | n = end - start |
| 34 | if n <= 0: |
| 35 | continue |
| 36 | tok = torch.arange(n, device=device) |
| 37 | blk = tok // kernel_block_size |
| 38 | off = tok % kernel_block_size |
| 39 | out[start:end] = block_table[r, blk].to(torch.int64) * kernel_block_size + off |
| 40 | return out |
| 41 | |
| 42 | |
| 43 | def _run(block_table, cu_num_tokens, kernel_block_size, out_len, device): |
| 44 | out = torch.full((out_len,), -1, dtype=torch.int64, device=device) |
| 45 | num_reqs = cu_num_tokens.numel() - 1 |
| 46 | max_tokens_per_req = block_table.shape[1] * kernel_block_size |
| 47 | num_chunks = max(1, -(-max_tokens_per_req // 1024)) |
| 48 | _expand_page_indices_kernel[(num_reqs, num_chunks)]( |
| 49 | out, |
| 50 | block_table, |
| 51 | block_table.stride(0), |
| 52 | block_table.stride(1), |
| 53 | cu_num_tokens, |
| 54 | KERNEL_BLOCK_SIZE=kernel_block_size, |
| 55 | BLOCK_SIZE=1024, |
| 56 | ) |
| 57 | return out |
| 58 | |
| 59 | |
| 60 | def _make(seq_lens, kernel_block_size, device, noncontig=False): |
| 61 | num_reqs = len(seq_lens) |
| 62 | max_blocks = ( |
| 63 | max(1, max((s + kernel_block_size - 1) // kernel_block_size for s in seq_lens)) |
| 64 | + 2 |
| 65 | ) |
| 66 | width = max_blocks * 2 if noncontig else max_blocks |
| 67 | torch.manual_seed(0) |
| 68 | bt = torch.randint(0, 10_000, (num_reqs, width), dtype=torch.int64, device=device) |
| 69 | if noncontig: |
| 70 | bt = bt[:, ::2] |
| 71 | assert not bt.is_contiguous() |
| 72 | cu = torch.zeros(num_reqs + 1, dtype=torch.int64, device=device) |
| 73 | cu[1:] = torch.tensor(seq_lens, dtype=torch.int64, device=device).cumsum(0) |
| 74 | return bt, cu, int(cu[-1]) |
| 75 | |
| 76 | |
| 77 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="needs a GPU") |
| 78 | @pytest.mark.parametrize( |
| 79 | "seq_lens", [[1], [128], [4096], [1, 1], [4096, 7], [37, 4096, 1], [1024] * 8] |
| 80 | ) |
| 81 | @pytest.mark.parametrize("kernel_block_size", [1, 16, 64]) |
| 82 | def test_matches_reference(seq_lens, kernel_block_size): |
| 83 | bt, cu, total = _make(seq_lens, kernel_block_size, "cuda") |
| 84 | got = _run(bt, cu, kernel_block_size, total, "cuda") |
| 85 | want = _reference(bt, cu, kernel_block_size, total, "cuda") |
| 86 | torch.testing.assert_close(got, want, atol=0, rtol=0) |
| 87 | |
| 88 | |
| 89 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="needs a GPU") |
| 90 | @pytest.mark.parametrize("kernel_block_size", [1, 16]) |
| 91 | def test_non_contiguous_block_table(kernel_block_size): |
| 92 | """Column stride != 1 must be honoured, not assumed away.""" |
| 93 | bt, cu, total = _make([4096, 37], kernel_block_size, "cuda", noncontig=True) |
| 94 | got = _run(bt, cu, kernel_block_size, total, "cuda") |
| 95 | want = _reference(bt, cu, kernel_block_size, total, "cuda") |
| 96 | torch.testing.assert_close(got, want, atol=0, rtol=0) |
| 97 | |
| 98 | |
| 99 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="needs a GPU") |
| 100 | def test_ragged_batch_leaves_other_rows_untouched(): |
| 101 | """A short request must not have its tail written by a longer one's chunks.""" |
| 102 | kbs = 16 |
| 103 | bt, cu, total = _make([4096, 1, 512], kbs, "cuda") |
| 104 | out = torch.full((total + 8,), -999, dtype=torch.int64, device="cuda") |
| 105 | num_chunks = max(1, -(-(bt.shape[1] * kbs) // 1024)) |
| 106 | _expand_page_indices_kernel[(3, num_chunks)]( |
| 107 | out, |
| 108 | bt, |
| 109 | bt.stride(0), |
| 110 | bt.stride(1), |
| 111 | cu, |
| 112 | KERNEL_BLOCK_SIZE=kbs, |
| 113 | BLOCK_SIZE=1024, |
| 114 | ) |
| 115 | want = _reference(bt, cu, kbs, total, "cuda") |
| 116 | torch.testing.assert_close(out[:total], want, atol=0, rtol=0) |
| 117 | # The slack past the last request must be untouched. |
| 118 | assert torch.all(out[total:] == -999) |
| 119 | |
| 120 | |
| 121 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="needs a GPU") |
| 122 | def test_zero_length_request_writes_nothing(): |
| 123 | """A zero-token request in a padded batch must produce no stores.""" |
| 124 | kbs = 16 |
| 125 | bt, cu, total = _make([256, 0, 128], kbs, "cuda") |
| 126 | got = _run(bt, cu, kbs, total, "cuda") |
| 127 | want = _reference(bt, cu, kbs, total, "cuda") |
| 128 | torch.testing.assert_close(got, want, atol=0, rtol=0) |
| 129 | assert int(cu[2] - cu[1]) == 0 |
| 130 | |
| 131 | |
| 132 | @pytest.mark.skipif(not torch.cuda.is_available(), reason="needs a GPU") |
| 133 | def test_over_provisioned_chunks_do_not_change_result(): |
| 134 | """Extra chunks must be inert. |
| 135 | |
| 136 | Each program covers `chunk_idx * BLOCK_SIZE + [0, BLOCK_SIZE)`, so the grid |
| 137 | has to span the request: 8192 tokens need at least ceil(8192/1024) == 8 |
| 138 | chunks, and fewer would simply leave the tail unwritten. The caller sizes |
| 139 | the grid from `block_table.shape[1] * kernel_block_size`, an upper bound |
| 140 | over the batch, so in a ragged batch most requests get more chunks than |
| 141 | they need. That over-provisioning is the case worth pinning: the surplus |
| 142 | programs must return early and change nothing. |
| 143 | """ |
| 144 | kbs = 1 |
| 145 | bt, cu, total = _make([8192], kbs, "cuda") |
| 146 | ref = _reference(bt, cu, kbs, total, "cuda") |
| 147 | required = -(-total // 1024) |
| 148 | for num_chunks in (required, required * 2, required * 8): |
| 149 | out = torch.full((total,), -1, dtype=torch.int64, device="cuda") |
| 150 | _expand_page_indices_kernel[(1, num_chunks)]( |
| 151 | out, |
| 152 | bt, |
| 153 | bt.stride(0), |
| 154 | bt.stride(1), |
| 155 | cu, |
| 156 | KERNEL_BLOCK_SIZE=kbs, |
| 157 | BLOCK_SIZE=1024, |
| 158 | ) |
| 159 | torch.testing.assert_close(out, ref, atol=0, rtol=0) |