| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Layout plumbing for DeepSeek V4.1 mega attention. |
| 4 | |
| 5 | The kernel's Q and O layouts are produced by permuting ``wq_b`` rows and |
| 6 | ``wo_a`` columns once at load, so what has to hold is that the permuted GEMMs |
| 7 | agree with a reference that permutes the activation instead. |
| 8 | """ |
| 9 | |
| 10 | import torch |
| 11 | |
| 12 | from vllm.models.deepseek_v41.common.ops.fused_layout import ( |
| 13 | WV_GROUP_SIZE, |
| 14 | o_fused_chunk_permutation, |
| 15 | o_fused_permutation, |
| 16 | permute_q_to_fused, |
| 17 | permute_wo_a_, |
| 18 | permute_wq_b_, |
| 19 | q_fused_permutation, |
| 20 | ) |
| 21 | |
| 22 | HEAD_DIM = 512 |
| 23 | |
| 24 | |
| 25 | def test_permuted_wq_b_gemm_matches_permuted_activation(): |
| 26 | """Permuting wq_b's rows makes the Q GEMM emit the fused layout directly.""" |
| 27 | torch.manual_seed(0) |
| 28 | num_heads, q_lora_rank, num_tokens = 16, 64, 5 |
| 29 | weight = torch.randn(num_heads * HEAD_DIM, q_lora_rank) |
| 30 | # One scale row per weight row, as an MXFP8 wq_b shard carries. |
| 31 | scale = torch.randint( |
| 32 | 0, 256, (num_heads * HEAD_DIM, q_lora_rank // 32), dtype=torch.uint8 |
| 33 | ) |
| 34 | qr = torch.randn(num_tokens, q_lora_rank) |
| 35 | orig_weight, orig_scale = weight.clone(), scale.clone() |
| 36 | |
| 37 | standard = (qr @ weight.T).view(num_tokens, num_heads, HEAD_DIM) |
| 38 | expected = permute_q_to_fused(standard) |
| 39 | |
| 40 | permute_wq_b_(weight, scale, num_heads) |
| 41 | fused = (qr @ weight.T).view(num_tokens, num_heads, HEAD_DIM) |
| 42 | torch.testing.assert_close(fused, expected) |
| 43 | # The scale has to follow its row, or dequant reads another row's scale. |
| 44 | perm = q_fused_permutation(num_heads, HEAD_DIM) |
| 45 | torch.testing.assert_close(weight, orig_weight[perm]) |
| 46 | torch.testing.assert_close(scale, orig_scale[perm]) |
| 47 | |
| 48 | |
| 49 | def test_permuted_wo_a_consumes_fused_output(): |
| 50 | """Permuting wo_a's input columns lets it read the kernel's O layout.""" |
| 51 | torch.manual_seed(1) |
| 52 | out_features, num_tokens = 32, 4 |
| 53 | in_features = WV_GROUP_SIZE * HEAD_DIM |
| 54 | weight = torch.randn(out_features, in_features) |
| 55 | scale = torch.randint(0, 256, (out_features, in_features // 32), dtype=torch.uint8) |
| 56 | standard_o = torch.randn(num_tokens, in_features) |
| 57 | orig_weight, orig_scale = weight.clone(), scale.clone() |
| 58 | |
| 59 | expected = standard_o @ weight.T |
| 60 | # The kernel emits the same values in the fused chunk order. |
| 61 | perm = o_fused_permutation(WV_GROUP_SIZE, HEAD_DIM) |
| 62 | fused_o = standard_o[:, perm] |
| 63 | |
| 64 | permute_wo_a_(weight, scale, WV_GROUP_SIZE) |
| 65 | torch.testing.assert_close(weight, orig_weight[:, perm]) |
| 66 | torch.testing.assert_close( |
| 67 | scale, orig_scale[:, o_fused_chunk_permutation(WV_GROUP_SIZE, HEAD_DIM)] |
| 68 | ) |
| 69 | # Unlike wq_b, this permutation sits inside the 4096-term reduction, so the |
| 70 | # summation order changes and the result differs in the last few ulps. |
| 71 | torch.testing.assert_close(fused_o @ weight.T, expected, rtol=1e-3, atol=1e-3) |