Should we also add a config, RL/examples/configs/grpo_math_8B_fp8_L3_F1_G_i.yaml
For example, below config can be a good candidate (with optionally set num_last_layers_in_bf16: 0 num_first_layers_in_bf16: 0):

GRPO Algorithm Configuration

defaults: "grpo_math_1B.yaml"

grpo:
num_prompts_per_step: 64
num_generations_per_prompt: 32

loss_fn:
use_importance_sampling_correction: true

policy:
model_name: "meta-llama/Llama-3.1-8B-Instruct"
tokenizer:
name: ${policy.model_name} ## specify if you'd like to use a tokenizer different from the model's default
train_global_batch_size: 512
train_micro_batch_size: 1
generation_batch_size: 32 # Only used when generating using HF backend
logprob_batch_size: 2
max_total_sequence_length: 4096
precision: "bfloat16"
fsdp_offload_enabled: false
activation_checkpointing_enabled: false

dtensor_cfg:
enabled: True

dynamic_batching:
train_mb_tokens: 4096
logprob_mb_tokens: 8192

optimizer:
name: "torch.optim.AdamW"
kwargs:
lr: 3.0e-7
weight_decay: 0.01
betas: [0.9, 0.999]
eps: 1e-8

scheduler:
- name: "torch.optim.lr_scheduler.LinearLR"
kwargs:
start_factor: 0.1
end_factor: 1.0
# The scheduler iteration is per GPRO step and is decoupled with the optimizer step (may be >=1 per GPRO step)
total_iters: 13
- name: "torch.optim.lr_scheduler.ConstantLR"
kwargs:
factor: 1.0
total_iters: 10000000000
- milestones: [13]

generation:
backend: "vllm"
max_new_tokens: ${policy.max_total_sequence_length}
temperature: 1.0
top_p: 1.0
top_k: null
stop_token_ids: null
stop_strings: null
vllm_cfg:
precision: 'fp8'
use_deep_gemm: true
num_last_layers_in_bf16: 3
num_first_layers_in_bf16: 1
tensor_parallel_size: 1
gpu_memory_utilization: 0.6
max_model_len: ${policy.max_total_sequence_length}

cluster:
gpus_per_node: 8
num_nodes: 1

Not super necessary immediately, but I think it'd be nice to include convergence plots for proof in the repo.