RL/examples/configs/recipes/llm/performance/grpo-deepseek-v3-32n8g.yaml at 2520299036bbfa4e6635bebb3e6cb599330d9692 · NVIDIA-NeMo/RL · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
defaults: ../../../grpo_math_1B.yaml
grpo:
  num_prompts_per_step: 32
  num_generations_per_prompt: 16
  max_num_steps: 500
  val_batch_size: 5
  max_val_samples: 16
loss_fn:
  use_importance_sampling_correction: true
checkpointing:
  checkpoint_dir: results/grpo-deepseek-v3-32n8g
policy:
  model_name: "Follow instructions in https://github.com/NVIDIA-NeMo/RL/blob/main/docs/guides/deepseek.md"
  train_micro_batch_size: 1
  logprob_batch_size: 1
  max_total_sequence_length: 1536
  make_sequence_length_divisible_by: 1
  dtensor_cfg:
    enabled: false
  megatron_cfg:
    enabled: true
    empty_unused_memory_level: 1
    converter_type: LlamaForCausalLM
    pipeline_model_parallel_size: 16
    expert_model_parallel_size: 16
    activation_checkpointing: true
    moe_grouped_gemm: true
    num_layers_in_first_pipeline_stage: 3
    num_layers_in_last_pipeline_stage: 2
    apply_rope_fusion: false
    defer_fp32_logits: true
    # MTP — disabled
    mtp_num_layers: 0
    optimizer:
      lr: 5.0e-07
      min_lr: 5.0e-08
      weight_decay: 0.0
      use_precision_aware_optimizer: true
    scheduler:
      lr_warmup_iters: 2
      lr_warmup_init: 5.0e-08
    fp8_cfg:
      enabled: false
  generation:
    vllm_cfg:
      tensor_parallel_size: 32
      async_engine: true
logger:
  log_dir: logs/grpo-deepseek-v3-32n8g
  wandb_enabled: true
  tensorboard_enabled: true
  wandb:
    project: nemo-rl
    name: grpo-deepseek-v3-32n8g
cluster:
  gpus_per_node: 8
  num_nodes: 32