File size: 2,409 Bytes
dfbcd52 | 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 | # Tuning Examples
Examples that focus on *tuning-related* features rather than on a specific RL algorithm. Everything in here trains with `verl.trainer.main_ppo` and the current Hydra API.
## Subdirectories
### `lora/` — LoRA fine-tuning
Canonical LoRA GRPO scripts (training only adapters, rollout still serves the adapter via `load_format=safetensors + layered_summon`).
| Script | Model | Infer | Train | Notes |
|------------------------------------------------|-------------------|-------|----------|-----------------------------|
| `run_qwen3_8b_fsdp.sh` | Qwen3-8B | vLLM | FSDP | text, GSM8K |
| `run_qwen3_8b_from_adapter_fsdp.sh` | Qwen3-8B | vLLM | FSDP | start from existing adapter |
| `run_qwen3_8b_merge_fsdp.sh` | Qwen3-8B | vLLM | FSDP | merge adapter into base |
| `run_qwen2_5_vl_7b_fsdp.sh` | Qwen2.5-VL-7B | vLLM | FSDP | vision, Geo3K |
| `run_qwen3_30b_a3b_megatron.sh` | Qwen3-30B-A3B | vLLM | Megatron | MoE |
Key flags:
- `actor_rollout_ref.model.lora_rank`, `actor_rollout_ref.model.lora_alpha`
- `actor_rollout_ref.rollout.load_format=safetensors`
- `actor_rollout_ref.rollout.layered_summon=True`
### `scaling/` — Large-model scale demos
Single/multi-node tuning recipes for large dense models; geared to practitioners trying to fit and run these models out of the box with GRPO + GSM8K/MATH.
| Script | Model | Infer | Train | Hardware |
|-------------------------------------------|-----------------|-------|----------|--------------------------|
| `run_qwen2_5_32b_megatron.sh` | Qwen2.5-32B | vLLM | Megatron | 1×8 GPUs (TP=8) |
| `run_qwen2_5_72b_fsdp.sh` | Qwen2.5-72B | vLLM | FSDP | 4×8 GPUs (TP=16, offload)|
## Conventions
- All scripts expose `MODEL_PATH`, `NNODES`, `NGPUS_PER_NODE`, batch sizes, learning rates, `ROLLOUT_TP`, `ROLLOUT_N`, etc. via `VAR=${VAR:-default}`.
- Dynamic batch size and `trainer.balance_batch=True` are enabled by default.
- No deprecated knobs (`ppo_micro_batch_size`, `data.val_batch_size`, top-level `reward_model.*`, `actor.ulysses_sequence_parallel_size`, `ppo_megatron_trainer.yaml`).
|