# Generation config for Experiment 2 (Step 1)
# 基于 verl/trainer/config/generation.yaml，针对实验二调整参数

trainer:
  nnodes: 1
  n_gpus_per_node: 8

data:
  path: null  # 由 run_exp2.sh 覆盖
  prompt_key: prompt
  n_samples: 1
  output_path: null  # 由 run_exp2.sh 覆盖
  batch_size: 64

model:
  path: null  # 由 run_exp2.sh 覆盖
  external_lib: null

rollout:
  name: vllm
  temperature: 0.6
  top_k: -1
  top_p: 0.95
  prompt_length: 2048
  response_length: 4096
  dtype: bfloat16
  gpu_memory_utilization: 0.7
  ignore_eos: False
  enforce_eager: False
  free_cache_engine: False
  load_format: dummy_dtensor
  tensor_model_parallel_size: 1
  max_num_batched_tokens: 16384
  max_model_len: null
  max_num_seqs: 256
  log_prob_micro_batch_size: null
  log_prob_micro_batch_size_per_gpu: 8
  use_fire_sampling: False
  do_sample: True
  disable_log_stats: True
  enable_chunked_prefill: False
  n: 1

actor:
  strategy: fsdp
  ulysses_sequence_parallel_size: 1
  fsdp_config:
    fsdp_size: -1

ray_init:
  num_cpus: null