| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Unit tests for the fused vocab-parallel embedding kernel |
| 4 | (``vllm._custom_ops.vocab_parallel_embedding``). |
| 5 | |
| 6 | The guarantee: the kernel is a drop-in for the TP > 1 embedding path, i.e. it |
| 7 | returns bit-exactly what mask + shift + gather + ``masked_fill_`` returns on |
| 8 | each rank, and summing the per-rank outputs (what the all-reduce does) |
| 9 | reconstructs a full-table lookup. |
| 10 | """ |
| 11 | |
| 12 | import pytest |
| 13 | import torch |
| 14 | |
| 15 | from vllm import _custom_ops as ops |
| 16 | from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 17 | VocabParallelEmbedding, |
| 18 | get_masked_input_and_mask, |
| 19 | pad_vocab_size, |
| 20 | ) |
| 21 | from vllm.platforms import current_platform |
| 22 | from vllm.utils.torch_utils import set_random_seed |
| 23 | |
| 24 | requires_cuda = pytest.mark.skipif( |
| 25 | not current_platform.is_cuda(), |
| 26 | reason="the fused vocab-parallel embedding kernel is CUDA-only", |
| 27 | ) |
| 28 | |
| 29 | |
| 30 | def _shard_indices(vocab_size, org_vocab_size, tp_rank, tp_size, pad=64): |
| 31 | return VocabParallelEmbedding._get_indices( |
| 32 | pad_vocab_size(vocab_size, pad), |
| 33 | pad_vocab_size(org_vocab_size, pad), |
| 34 | vocab_size, |
| 35 | org_vocab_size, |
| 36 | tp_rank, |
| 37 | tp_size, |
| 38 | ) |
| 39 | |
| 40 | |
| 41 | def _reference(input_ids, weight, si): |
| 42 | """The path the kernel replaces.""" |
| 43 | masked_input, input_mask = get_masked_input_and_mask( |
| 44 | input_ids, |
| 45 | si.org_vocab_start_index, |
| 46 | si.org_vocab_end_index, |
| 47 | si.num_org_vocab_padding, |
| 48 | si.added_vocab_start_index, |
| 49 | si.added_vocab_end_index, |
| 50 | ) |
| 51 | out = torch.nn.functional.embedding(masked_input.long(), weight) |
| 52 | return out.masked_fill_(input_mask.unsqueeze(-1), 0) |
| 53 | |
| 54 | |
| 55 | def _fused(input_ids, weight, si): |
| 56 | return ops.vocab_parallel_embedding( |
| 57 | input_ids, |
| 58 | weight, |
| 59 | si.org_vocab_start_index, |
| 60 | si.org_vocab_end_index, |
| 61 | si.num_org_vocab_padding, |
| 62 | si.added_vocab_start_index, |
| 63 | si.added_vocab_end_index, |
| 64 | ) |
| 65 | |
| 66 | |
| 67 | def _shard(full_table, si, rows_per_rank, dtype): |
| 68 | """Lay a rank's rows out the way the weight loader does.""" |
| 69 | weight = torch.zeros(rows_per_rank, full_table.shape[1], dtype=dtype, device="cuda") |
| 70 | num_org = si.org_vocab_end_index - si.org_vocab_start_index |
| 71 | weight[:num_org] = full_table[si.org_vocab_start_index : si.org_vocab_end_index] |
| 72 | added_at = num_org + si.num_org_vocab_padding |
| 73 | num_added = si.added_vocab_end_index - si.added_vocab_start_index |
| 74 | weight[added_at : added_at + num_added] = full_table[ |
| 75 | si.added_vocab_start_index : si.added_vocab_end_index |
| 76 | ] |
| 77 | return weight |
| 78 | |
| 79 | |
| 80 | @requires_cuda |
| 81 | @torch.inference_mode() |
| 82 | @pytest.mark.parametrize("tp_size", [2, 4, 8]) |
| 83 | @pytest.mark.parametrize( |
| 84 | "vocab_size,org_vocab_size", [(128256, 128256), (32256, 32000)] |
| 85 | ) |
| 86 | @pytest.mark.parametrize("hidden", [4096, 3584]) |
| 87 | @pytest.mark.parametrize("id_dtype", [torch.int32, torch.int64]) |
| 88 | def test_matches_masked_gather(tp_size, vocab_size, org_vocab_size, hidden, id_dtype): |
| 89 | """Every rank's partial output is bit-exact with the eager path, including |
| 90 | the added-vocab (LoRA) rows and the ids owned by another rank.""" |
| 91 | set_random_seed(7) |
| 92 | dtype = torch.bfloat16 |
| 93 | rows = pad_vocab_size(vocab_size, 64) // tp_size |
| 94 | full_table = torch.randn(pad_vocab_size(vocab_size, 64), hidden, dtype=dtype).cuda() |
| 95 | ids = torch.randint(0, vocab_size, (257,), dtype=id_dtype, device="cuda") |
| 96 | |
| 97 | for tp_rank in range(tp_size): |
| 98 | si = _shard_indices(vocab_size, org_vocab_size, tp_rank, tp_size) |
| 99 | weight = _shard(full_table, si, rows, dtype) |
| 100 | torch.testing.assert_close( |
| 101 | _fused(ids, weight, si), _reference(ids, weight, si), atol=0.0, rtol=0.0 |
| 102 | ) |
| 103 | |
| 104 | |
| 105 | @requires_cuda |
| 106 | @torch.inference_mode() |
| 107 | @pytest.mark.parametrize("tp_size", [2, 8]) |
| 108 | def test_rank_sum_is_full_lookup(tp_size): |
| 109 | """What the all-reduce sees: the partials add up to a full-table lookup.""" |
| 110 | set_random_seed(11) |
| 111 | vocab_size, org_vocab_size, hidden = 32256, 32000, 2048 |
| 112 | dtype = torch.bfloat16 |
| 113 | rows = pad_vocab_size(vocab_size, 64) // tp_size |
| 114 | full_table = torch.randn(pad_vocab_size(vocab_size, 64), hidden, dtype=dtype).cuda() |
| 115 | # Ids from both the base and the added (LoRA) range. |
| 116 | ids = torch.randint(0, vocab_size, (129,), dtype=torch.int32, device="cuda") |
| 117 | ids[:16] = torch.randint( |
| 118 | org_vocab_size, vocab_size, (16,), dtype=torch.int32, device="cuda" |
| 119 | ) |
| 120 | |
| 121 | total = torch.zeros(ids.shape[0], hidden, dtype=dtype, device="cuda") |
| 122 | for tp_rank in range(tp_size): |
| 123 | si = _shard_indices(vocab_size, org_vocab_size, tp_rank, tp_size) |
| 124 | total += _fused(ids, _shard(full_table, si, rows, dtype), si) |
| 125 | |
| 126 | torch.testing.assert_close(total, full_table[ids.long()], atol=0.0, rtol=0.0) |