vllm/benchmarks/attention_benchmarks/configs/mla_mixed_batch.yaml

# MLA mixed batch benchmark (prefill + decode)
# Tests chunked prefill performance

model:
  name: "deepseek-v3"
  num_layers: 60
  num_q_heads: 128
  num_kv_heads: 1
  head_dim: 576
  kv_lora_rank: 512
  qk_nope_head_dim: 128
  qk_rope_head_dim: 64
  v_head_dim: 128
  block_size: 128

batch_specs:
  # Small prefill + decode
  - "1q1k_8q1s1k"           # 1 prefill + 8 decode
  - "2q2k_16q1s1k"          # 2 prefill + 16 decode
  - "4q1k_32q1s2k"          # 4 prefill + 32 decode

  # Medium prefill + decode
  - "2q4k_32q1s2k"          # 2 medium prefill + 32 decode
  - "4q4k_64q1s2k"          # 4 medium prefill + 64 decode
  - "8q2k_64q1s4k"          # 8 prefill + 64 decode

  # Large prefill + decode (chunked prefill stress test)
  - "2q8k_32q1s1k"          # 2 large prefill + 32 decode
  - "1q16k_16q1s2k"         # 1 very large prefill + 16 decode
  - "2q16k_32q1s4k"         # 2 very large prefill + 32 decode

  # Context extension + decode
  - "2q1kkv2k_16q1s1k"       # 2 extend + 16 decode
  - "4q2kkv4k_32q1s2k"       # 4 extend + 32 decode
  - "2q1kkv8k_32q1s2k"       # 2 large extend + 32 decode

  # Explicitly chunked prefill
  - "q8k"           # 8k prefill with chunking hint
  - "q16k"          # 16k prefill with chunking hint
  - "2q8k_32q1s2k"    # 2 chunked prefill + 32 decode

  # High decode ratio (realistic serving)
  - "1q2k_63q1s1k"          # 1 prefill + 63 decode
  - "2q2k_62q1s2k"          # 2 prefill + 62 decode
  - "4q4k_60q1s4k"          # 4 prefill + 60 decode

backends:
  - CUTLASS_MLA
  - FLASHINFER_MLA
  - FLASH_ATTN_MLA   # Hopper only
  - FLASHMLA         # Hopper only

device: "cuda:0"
repeats: 5
warmup_iters: 3
profile_memory: true

# Analyze chunked prefill workspace size impact
chunked_prefill:
  test_workspace_sizes: [4096, 8192, 16384, 32768, 65536]