|
14 | 14 | import example_mha_fwd_varlen |
15 | 15 |
|
16 | 16 |
|
17 | | -@tilelang.testing.requires_cuda |
18 | 17 | def bench_example_gqa_bwd_tma_reduce_varlen(): |
19 | 18 | tilelang.tools.bench.process_func(example_gqa_bwd_tma_reduce_varlen.main) |
20 | 19 |
|
21 | 20 |
|
22 | | -@tilelang.testing.requires_cuda |
23 | 21 | def bench_example_gqa_bwd(): |
24 | 22 | tilelang.tools.bench.process_func(example_gqa_bwd.main) |
25 | 23 |
|
26 | 24 |
|
27 | | -@tilelang.testing.requires_cuda |
28 | | -@tilelang.testing.requires_cuda_compute_version_ge(9, 0) |
29 | 25 | def bench_example_gqa_bwd_wgmma_pipelined(): |
30 | 26 | tilelang.tools.bench.process_func(example_gqa_bwd_wgmma_pipelined.main) |
31 | 27 |
|
32 | 28 |
|
33 | | -@tilelang.testing.requires_cuda |
34 | 29 | def bench_example_mha_bwd(): |
35 | 30 | tilelang.tools.bench.process_func(example_mha_bwd.main) |
36 | 31 |
|
37 | 32 |
|
38 | | -@tilelang.testing.requires_cuda |
39 | 33 | def bench_example_mha_bwd_bhsd(): |
40 | 34 | tilelang.tools.bench.process_func(example_mha_bwd_bhsd.main) |
41 | 35 |
|
42 | 36 |
|
43 | | -@tilelang.testing.requires_cuda |
44 | | -@tilelang.testing.requires_cuda_compute_version_ge(9, 0) |
45 | 37 | def bench_example_mha_bwd_wgmma_pipelined(): |
46 | 38 | tilelang.tools.bench.process_func(example_mha_bwd_wgmma_pipelined.main) |
47 | 39 |
|
48 | 40 |
|
49 | | -@tilelang.testing.requires_cuda |
50 | | -@tilelang.testing.requires_cuda_compute_version_ge(9, 0) |
51 | 41 | def bench_example_gqa_fwd_bshd_wgmma_pipelined(): |
52 | 42 | tilelang.tools.bench.process_func(example_gqa_fwd_bshd_wgmma_pipelined.main) |
53 | 43 |
|
54 | 44 |
|
55 | | -@tilelang.testing.requires_cuda |
56 | 45 | def bench_example_gqa_fwd_bshd(): |
57 | 46 | tilelang.tools.bench.process_func(example_gqa_fwd_bshd.main) |
58 | 47 |
|
59 | 48 |
|
60 | | -@tilelang.testing.requires_cuda |
61 | | -@tilelang.testing.requires_cuda_compute_version_ge(9, 0) |
62 | 49 | def bench_example_mha_fwd_bhsd_wgmma_pipelined(): |
63 | 50 | tilelang.tools.bench.process_func(example_mha_fwd_bhsd_wgmma_pipelined.main) |
64 | 51 |
|
65 | 52 |
|
66 | | -@tilelang.testing.requires_cuda |
67 | 53 | def bench_example_mha_fwd_bhsd(): |
68 | 54 | tilelang.tools.bench.process_func(example_mha_fwd_bhsd.main) |
69 | 55 |
|
70 | 56 |
|
71 | | -@tilelang.testing.requires_cuda |
72 | | -@tilelang.testing.requires_cuda_compute_version_ge(9, 0) |
73 | 57 | def bench_example_mha_fwd_bshd_wgmma_pipelined(): |
74 | 58 | tilelang.tools.bench.process_func(example_mha_fwd_bshd_wgmma_pipelined.main) |
75 | 59 |
|
76 | 60 |
|
77 | | -@tilelang.testing.requires_cuda |
78 | 61 | def bench_example_mha_fwd_bshd(): |
79 | 62 | tilelang.tools.bench.process_func(example_mha_fwd_bshd.main) |
80 | 63 |
|
81 | 64 |
|
82 | | -@tilelang.testing.requires_cuda |
83 | 65 | def bench_example_mha_fwd_varlen(): |
84 | 66 | tilelang.tools.bench.process_func(example_mha_fwd_varlen.main) |
85 | 67 |
|
|
0 commit comments