| 1 | # MLA prefill-only benchmark configuration for sparse backends |
| 2 | |
| 3 | model: |
| 4 | name: "deepseek-v3" |
| 5 | num_layers: 60 |
| 6 | num_q_heads: 128 |
| 7 | num_kv_heads: 1 |
| 8 | head_dim: 576 |
| 9 | kv_lora_rank: 512 |
| 10 | qk_nope_head_dim: 128 |
| 11 | qk_rope_head_dim: 64 |
| 12 | v_head_dim: 128 |
| 13 | block_size: 128 |
| 14 | |
| 15 | # Model parameter sweep: simulate tensor parallelism by varying num_q_heads |
| 16 | # TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads |
| 17 | model_parameter_sweep: |
| 18 | param_name: "num_q_heads" |
| 19 | values: [128, 64, 32, 16] |
| 20 | label_format: "{backend}_{value}h" |
| 21 | |
| 22 | batch_specs: |
| 23 | # Pure prefill |
| 24 | - "1q512" |
| 25 | - "1q1k" |
| 26 | - "1q2k" |
| 27 | - "1q4k" |
| 28 | - "1q8k" |
| 29 | |
| 30 | # Batched pure prefill |
| 31 | - "2q512" |
| 32 | - "2q1k" |
| 33 | - "2q2k" |
| 34 | - "2q4k" |
| 35 | - "2q8k" |
| 36 | - "4q512" |
| 37 | - "4q1k" |
| 38 | - "4q2k" |
| 39 | - "4q4k" |
| 40 | - "4q8k" |
| 41 | - "8q512" |
| 42 | - "8q1k" |
| 43 | - "8q2k" |
| 44 | - "8q4k" |
| 45 | - "8q8k" |
| 46 | |
| 47 | # Extend |
| 48 | - "1q512s4k" |
| 49 | - "1q512s8k" |
| 50 | - "1q1ks8k" |
| 51 | - "1q2ks8k" |
| 52 | - "1q2ks16k" |
| 53 | - "1q4ks16k" |
| 54 | |
| 55 | backends: |
| 56 | - FLASHMLA_SPARSE |
| 57 | - FLASHINFER_MLA_SPARSE |
| 58 | |
| 59 | device: "cuda:0" |
| 60 | profile_memory: true |