| 1 | # MLA prefill backend comparison |
| 2 | # |
| 3 | # Compares all available MLA prefill backends: |
| 4 | # FA backends: fa2, fa3, fa4 (FlashAttention versions) |
| 5 | # Non-FA: flashinfer, cudnn, trtllm (Blackwell-only, require flashinfer) |
| 6 | # CuTe DSL: tokenspeed (Blackwell + R1 dims, requires tokenspeed_mla) |
| 7 | # |
| 8 | # Uses cutlass_mla as the decode backend for impl construction |
| 9 | # (only the prefill path is exercised). |
| 10 | # |
| 11 | # Backends that aren't available on the current platform will report errors |
| 12 | # in the results table (e.g., fa3 on Blackwell, cudnn without artifactory). |
| 13 | # |
| 14 | # Usage: |
| 15 | # python benchmark.py --config configs/mla_prefill.yaml |
| 16 | |
| 17 | description: "MLA prefill backend comparison" |
| 18 | |
| 19 | model: |
| 20 | name: "deepseek-v3" |
| 21 | num_layers: 60 |
| 22 | num_q_heads: 128 |
| 23 | num_kv_heads: 1 |
| 24 | head_dim: 576 |
| 25 | kv_lora_rank: 512 |
| 26 | qk_nope_head_dim: 128 |
| 27 | qk_rope_head_dim: 64 |
| 28 | v_head_dim: 128 |
| 29 | block_size: 128 |
| 30 | |
| 31 | # model: |
| 32 | # name: "deepseek-v2-lite" |
| 33 | # num_layers: 27 |
| 34 | # num_q_heads: 16 |
| 35 | # num_kv_heads: 1 |
| 36 | # head_dim: 576 |
| 37 | # kv_lora_rank: 512 |
| 38 | # qk_nope_head_dim: 128 |
| 39 | # qk_rope_head_dim: 64 |
| 40 | # v_head_dim: 128 |
| 41 | # block_size: 128 |
| 42 | |
| 43 | batch_specs: |
| 44 | # Pure prefill |
| 45 | - "q512" |
| 46 | - "q1k" |
| 47 | - "q2k" |
| 48 | - "q4k" |
| 49 | - "q8k" |
| 50 | |
| 51 | # Batched pure prefill |
| 52 | - "2q512" |
| 53 | - "2q1k" |
| 54 | - "2q2k" |
| 55 | - "2q4k" |
| 56 | - "2q8k" |
| 57 | - "4q512" |
| 58 | - "4q1k" |
| 59 | - "4q2k" |
| 60 | - "4q4k" |
| 61 | - "4q8k" |
| 62 | - "8q512" |
| 63 | - "8q1k" |
| 64 | - "8q2k" |
| 65 | - "8q4k" |
| 66 | - "8q8k" |
| 67 | |
| 68 | # Chunked prefill / extend |
| 69 | # Short context |
| 70 | - "q128s1k" |
| 71 | - "q256s2k" |
| 72 | - "q512s4k" |
| 73 | - "q1ks4k" |
| 74 | - "q2ks8k" |
| 75 | - "2q128s1k" |
| 76 | - "2q256s2k" |
| 77 | - "2q512s4k" |
| 78 | - "2q1ks4k" |
| 79 | - "2q2ks8k" |
| 80 | - "4q128s1k" |
| 81 | - "4q256s2k" |
| 82 | - "4q512s4k" |
| 83 | - "4q1ks4k" |
| 84 | - "4q2ks8k" |
| 85 | - "8q128s1k" |
| 86 | - "8q256s2k" |
| 87 | - "8q512s4k" |
| 88 | - "8q1ks4k" |
| 89 | |
| 90 | # Medium context |
| 91 | - "q128s16k" |
| 92 | - "q512s16k" |
| 93 | - "q1ks16k" |
| 94 | - "q2ks16k" |
| 95 | - "2q128s16k" |
| 96 | - "2q512s16k" |
| 97 | - "2q1ks16k" |
| 98 | - "2q2ks16k" |
| 99 | - "4q128s16k" |
| 100 | - "4q512s16k" |
| 101 | - "4q1ks16k" |
| 102 | - "4q2ks16k" |
| 103 | |
| 104 | # Long context |
| 105 | - "q128s64k" |
| 106 | - "q512s64k" |
| 107 | - "q1ks64k" |
| 108 | - "q2ks64k" |
| 109 | - "2q128s64k" |
| 110 | - "2q512s64k" |
| 111 | - "2q1ks64k" |
| 112 | - "2q2ks64k" |
| 113 | |
| 114 | decode_backends: |
| 115 | - CUTLASS_MLA |
| 116 | |
| 117 | prefill_backends: |
| 118 | - fa2 |
| 119 | - fa3 |
| 120 | - fa4 |
| 121 | - flashinfer |
| 122 | - cudnn |
| 123 | - trtllm |
| 124 | - tokenspeed |
| 125 | |
| 126 | device: "cuda:0" |