# MLA prefill-only benchmark configuration for sparse backends

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

# Model parameter sweep: simulate tensor parallelism by varying num_q_heads
# TP=1: 128 heads, TP=2: 64 heads, TP=4: 32 heads, TP=8: 16 heads
model_parameter_sweep:
  param_name: "num_q_heads"
  values: [128, 64, 32, 16]
  label_format: "{backend}_{value}h"

batch_specs:
  # Pure prefill
  - "1q512"
  - "1q1k"
  - "1q2k"
  - "1q4k"
  - "1q8k"

  # Batched pure prefill
  - "2q512"
  - "2q1k"
  - "2q2k"
  - "2q4k"
  - "2q8k"
  - "4q512"
  - "4q1k"
  - "4q2k"
  - "4q4k"
  - "4q8k"
  - "8q512"
  - "8q1k"
  - "8q2k"
  - "8q4k"
  - "8q8k"

  # Extend
  - "1q512s4k"
  - "1q512s8k"
  - "1q1ks8k"
  - "1q2ks8k"
  - "1q2ks16k"
  - "1q4ks16k"

backends:
  - FLASHMLA_SPARSE
  - FLASHINFER_MLA_SPARSE

device: "cuda:0"
repeats: 10
warmup_iters: 3
profile_memory: true