| 1 | # MLA decode-only benchmark configuration |
| 2 | |
| 3 | model: |
| 4 | name: "deepseek-v3" |
| 5 | num_layers: 60 |
| 6 | num_q_heads: 128 # Base value, can be swept for TP simulation |
| 7 | num_kv_heads: 1 # MLA uses single latent KV |
| 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 # CUTLASS MLA and FlashAttn MLA use 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 | # Small batches, varying sequence lengths |
| 24 | - "16q1s512" # 16 requests, 512 KV cache |
| 25 | - "16q1s1k" # 16 requests, 1k KV cache |
| 26 | - "16q1s2k" # 16 requests, 2k KV cache |
| 27 | - "16q1s4k" # 16 requests, 4k KV cache |
| 28 | |
| 29 | # Medium batches |
| 30 | - "32q1s1k" # 32 requests, 1k KV cache |
| 31 | - "32q1s2k" # 32 requests, 2k KV cache |
| 32 | - "32q1s4k" # 32 requests, 4k KV cache |
| 33 | - "32q1s8k" # 32 requests, 8k KV cache |
| 34 | |
| 35 | # Large batches |
| 36 | - "64q1s1k" # 64 requests, 1k KV cache |
| 37 | - "64q1s2k" # 64 requests, 2k KV cache |
| 38 | - "64q1s4k" # 64 requests, 4k KV cache |
| 39 | - "64q1s8k" # 64 requests, 8k KV cache |
| 40 | |
| 41 | # Very large batches |
| 42 | - "128q1s1k" # 128 requests, 1k KV cache |
| 43 | - "128q1s2k" # 128 requests, 2k KV cache |
| 44 | - "128q1s4k" # 128 requests, 4k KV cache |
| 45 | - "128q1s8k" # 128 requests, 8k KV cache |
| 46 | |
| 47 | # Long context |
| 48 | - "32q1s16k" # 32 requests, 16k KV cache |
| 49 | - "32q1s32k" # 32 requests, 32k KV cache |
| 50 | |
| 51 | backends: |
| 52 | - FLASHMLA_SPARSE |
| 53 | - FLASHINFER_MLA_SPARSE |
| 54 | |
| 55 | device: "cuda:0" |
| 56 | profile_memory: true |