| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import tracemalloc |
| 5 | |
| 6 | import pytest |
| 7 | import torch |
| 8 | |
| 9 | from vllm.platforms import current_platform |
| 10 | |
| 11 | if not current_platform.is_cpu(): |
| 12 | pytest.skip("skipping CPU-only tests", allow_module_level=True) |
| 13 | |
| 14 | import vllm._C # noqa: F401, E402 |
| 15 | |
| 16 | VOCAB_SIZES = [32000, 49152, 128256] |
| 17 | BATCH_SIZES = [1, 4, 16] |
| 18 | |
| 19 | |
| 20 | def reference_greedy_sample(logits: torch.Tensor) -> torch.Tensor: |
| 21 | return logits.argmax(dim=-1).view(-1) |
| 22 | |
| 23 | |
| 24 | class TestGreedyArgmax: |
| 25 | @pytest.mark.parametrize("vocab_size", VOCAB_SIZES) |
| 26 | @pytest.mark.parametrize("batch_size", BATCH_SIZES) |
| 27 | def test_exact_match(self, vocab_size: int, batch_size: int): |
| 28 | logits = torch.randn(batch_size, vocab_size, dtype=torch.float32) |
| 29 | expected = reference_greedy_sample(logits) |
| 30 | result = torch.ops._C.greedy_argmax(logits) |
| 31 | torch.testing.assert_close(result, expected) |
| 32 | |
| 33 | def test_single_dominant(self): |
| 34 | logits = torch.full((1, 50000), -1e9, dtype=torch.float32) |
| 35 | logits[0, 42] = 100.0 |
| 36 | assert torch.ops._C.greedy_argmax(logits).item() == 42 |
| 37 | |
| 38 | def test_negative_logits(self): |
| 39 | logits = torch.randn(8, 32000, dtype=torch.float32) - 10.0 |
| 40 | expected = reference_greedy_sample(logits) |
| 41 | result = torch.ops._C.greedy_argmax(logits) |
| 42 | torch.testing.assert_close(result, expected) |
| 43 | |
| 44 | |
| 45 | class TestFusedGumbelArgmax: |
| 46 | @pytest.mark.parametrize("vocab_size", VOCAB_SIZES) |
| 47 | def test_distribution_chi_squared(self, vocab_size: int): |
| 48 | """Verify sampling distribution via chi-squared goodness of fit.""" |
| 49 | small_vocab = min(vocab_size, 100) |
| 50 | logits = torch.randn(1, small_vocab, dtype=torch.float32) |
| 51 | probs = logits.softmax(dim=-1).squeeze(0) |
| 52 | |
| 53 | n_samples = 100_000 |
| 54 | counts = torch.zeros(small_vocab) |
| 55 | for trial in range(n_samples): |
| 56 | seed = torch.tensor([trial * 7 + 13], dtype=torch.long) |
| 57 | tile = logits.expand(1, -1).contiguous() |
| 58 | idx = torch.ops._C.fused_gumbel_argmax(tile, seed) |
| 59 | counts[idx.item()] += 1 |
| 60 | |
| 61 | expected = probs * n_samples |
| 62 | mask = expected > 5 |
| 63 | chi2 = ((counts[mask] - expected[mask]) ** 2 / expected[mask]).sum() |
| 64 | dof = mask.sum().item() - 1 |
| 65 | from scipy.stats import chi2 as chi2_dist |
| 66 | |
| 67 | p_value = 1.0 - chi2_dist.cdf(chi2.item(), dof) |
| 68 | assert p_value > 0.001, ( |
| 69 | f"Chi-squared test failed: chi2={chi2.item():.1f}, " |
| 70 | f"dof={dof}, p={p_value:.6f}" |
| 71 | ) |
| 72 | |
| 73 | def test_deterministic_same_seed(self): |
| 74 | """Same seed produces same result.""" |
| 75 | logits = torch.randn(4, 32000, dtype=torch.float32) |
| 76 | seeds = torch.tensor([42, 123, 456, 789], dtype=torch.long) |
| 77 | r1 = torch.ops._C.fused_gumbel_argmax(logits, seeds) |
| 78 | r2 = torch.ops._C.fused_gumbel_argmax(logits, seeds) |
| 79 | torch.testing.assert_close(r1, r2) |
| 80 | |
| 81 | def test_different_seeds_differ(self): |
| 82 | logits = torch.zeros(16, 50000, dtype=torch.float32) |
| 83 | seeds_a = torch.arange(16, dtype=torch.long) |
| 84 | seeds_b = torch.arange(16, dtype=torch.long) + 1_000_000 |
| 85 | r_a = torch.ops._C.fused_gumbel_argmax(logits, seeds_a) |
| 86 | r_b = torch.ops._C.fused_gumbel_argmax(logits, seeds_b) |
| 87 | assert not torch.equal(r_a, r_b) |
| 88 | |
| 89 | @pytest.mark.parametrize("vocab_size", VOCAB_SIZES) |
| 90 | @pytest.mark.parametrize("batch_size", BATCH_SIZES) |
| 91 | def test_output_in_range(self, vocab_size: int, batch_size: int): |
| 92 | logits = torch.randn(batch_size, vocab_size, dtype=torch.float32) |
| 93 | seeds = torch.arange(batch_size, dtype=torch.long) |
| 94 | result = torch.ops._C.fused_gumbel_argmax(logits, seeds) |
| 95 | assert result.min() >= 0 |
| 96 | assert result.max() < vocab_size |
| 97 | |
| 98 | |
| 99 | class TestMemory: |
| 100 | def test_fused_no_intermediate_allocs(self): |
| 101 | """Verify that the fused kernel does not allocate large intermediates.""" |
| 102 | logits = torch.randn(16, 128256, dtype=torch.float32) |
| 103 | seeds = torch.arange(16, dtype=torch.long) |
| 104 | |
| 105 | torch.ops._C.fused_gumbel_argmax(logits, seeds) |
| 106 | |
| 107 | tracemalloc.start() |
| 108 | snap_before = tracemalloc.take_snapshot() |
| 109 | for _ in range(50): |
| 110 | torch.ops._C.fused_gumbel_argmax(logits, seeds) |
| 111 | snap_after = tracemalloc.take_snapshot() |
| 112 | tracemalloc.stop() |
| 113 | |
| 114 | diff = snap_after.compare_to(snap_before, "lineno") |
| 115 | total_new_bytes = sum(s.size_diff for s in diff if s.size_diff > 0) |
| 116 | vocab_bytes = 16 * 128256 * 4 |
| 117 | assert total_new_bytes < vocab_bytes, ( |
| 118 | f"Fused kernel allocated {total_new_bytes} bytes, " |
| 119 | f"expected < {vocab_bytes} (one intermediate tensor)" |
| 120 | ) |