vllm/benchmarks/attention_benchmarks/configs/standard_attention.yaml

# Standard attention backend benchmark configuration

model:
  num_layers: 32
  num_q_heads: 32
  num_kv_heads: 8  # GQA with 4:1 ratio
  head_dim: 128
  block_size: 16

batch_specs:
  # Pure prefill
  - "q512"      # Small prefill (512 tokens)
  - "q2k"       # Medium prefill (2048 tokens)
  - "q4k"       # Large prefill (4096 tokens)
  - "q8k"       # Very large prefill (8192 tokens)

  # Pure decode
  - "8q1s1k"      # 8 requests, 1k KV cache each
  - "16q1s2k"     # 16 requests, 2k KV cache each
  - "32q1s1k"     # 32 requests, 1k KV cache each
  - "64q1s4k"     # 64 requests, 4k KV cache each

  # Mixed prefill/decode
  - "2q2k_8q1s1k"      # 2 prefill + 8 decode
  - "4q1k_16q1s2k"     # 4 prefill + 16 decode
  - "2q4k_32q1s1k"     # 2 large prefill + 32 decode

  # Speculative decode (q <= 8)
  - "16q2s1k"         # 16 requests, 2 spec tokens, 1k KV cache
  - "16q4s1k"         # 16 requests, 4 spec tokens, 1k KV cache
  - "16q8s1k"         # 16 requests, 8 spec tokens, 1k KV cache
  - "32q4s2k"         # 32 requests, 4 spec tokens, 2k KV cache
  - "8q8s4k"          # 8 requests, 8 spec tokens, 4k KV cache

  # Context extension (chunked prefill)
  - "q1ks2k"          # 1k query, 2k sequence
  - "2q1ks4k"         # 2 requests: 1k query, 4k sequence

# Available backends: flash, triton, flashinfer
backends:
  - flash
  - triton
  - flashinfer

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