run_qwen3_5_35b_a3b_veomni.sh

#!/usr/bin/env bash
# This script is a demo for GRPO training of Qwen3.5-35B-A3B using VeOmniEngine.
#
# Environment:
#   - transformers==5.3.0
#   - sglang==0.5.9
#   - flash-linear-attention==0.4.1
#   - veomni==0.1.9a1
# Tested configuration:
#   - Model: Qwen3.5-35B-A3B
#   - Sequence Parallel (SP): Tested with sp=1 and sp=2 (ulysses_parallel_size).

set -xeuo pipefail

# ---- user-adjustable ----
data_path=${data_path:-$HOME/data/geo3k}
model_path=${model_path:-$HOME/model/Qwen3.5-35B-A3B}
usp_size=${usp_size:-2}
nnodes=${nnodes:-2}

backend=fsdp2
model_engine=veomni
project_name='verl_grpo_qwen3_5_35b_a3b_geo3k'
exp_name='qwen3_5_35b_a3b_veomni_sp2'


# ===================================== Algorithm =====================================
adv_estimator=grpo
loss_mode=gspo

# reference policy
use_kl_in_reward=False
kl_coef=0.001
use_kl_loss=False
kl_loss_coef=0.001

clip_ratio_low=3e-4
clip_ratio_high=4e-4

actor_lr=1e-6
critic_lr=2e-6
gae_gamma=1.0
gae_lam=0.95
critic_warmup=0

# ===================================== Data/Model =====================================
train_files=$data_path/train.parquet
test_files=$data_path/test.parquet

actor_model_path=$model_path

max_prompt_length=$((1024 * 1))
max_response_length=$((1024 * 2))

train_batch_size=128
ppo_mini_batch_size=32
n_resp_per_prompt=8
n_resp_per_prompt_val=1

use_remove_padding=True
use_dynamic_bsz=False

# ===================================== Training =====================================
actor_max_token_len_per_gpu=$(((max_prompt_length + max_response_length) * 8))
ppo_micro_batch_size_per_gpu=1

# VeOmni config
ACTOR_VEOMNI_CONFIG="
    actor_rollout_ref.actor.veomni.param_offload=True \
    actor_rollout_ref.actor.veomni.optimizer_offload=True \
    actor_rollout_ref.actor.veomni.enable_full_shard=True \
    actor_rollout_ref.actor.veomni.ulysses_parallel_size=$usp_size \
    actor_rollout_ref.actor.veomni.expert_parallel_size=1 \
    actor_rollout_ref.actor.veomni.attn_implementation=flash_attention_2"
# ---- end user-adjustable ----

# ---- no user adjustment needed below ----
# Actor model config
ACTOR_CONFIG="
    actor_rollout_ref.actor.optim.lr=$actor_lr \
    actor_rollout_ref.model.path=$actor_model_path \
    actor_rollout_ref.model.use_remove_padding=$use_remove_padding \
    actor_rollout_ref.actor.use_kl_loss=$use_kl_loss \
    actor_rollout_ref.actor.kl_loss_coef=$kl_loss_coef \
    actor_rollout_ref.actor.clip_ratio_low=$clip_ratio_low \
    actor_rollout_ref.actor.clip_ratio_high=$clip_ratio_high \
    actor_rollout_ref.actor.clip_ratio_c=10.0 \
    actor_rollout_ref.actor.policy_loss.loss_mode=${loss_mode} \
    actor_rollout_ref.actor.use_dynamic_bsz=$use_dynamic_bsz \
    actor_rollout_ref.actor.ppo_mini_batch_size=$ppo_mini_batch_size \
    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=${actor_max_token_len_per_gpu} \
    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=$ppo_micro_batch_size_per_gpu"

CONFIG_NAME=ppo_trainer
ACTOR_CONFIG="$ACTOR_CONFIG $ACTOR_VEOMNI_CONFIG"

# ===================================== Inference =====================================
rollout_name=sglang
infer_tp=4
infer_dp=1
infer_ep=1
gpu_memory_utilization=0.6

ROLLOUT_CONFIG="
    actor_rollout_ref.rollout.name=$rollout_name \
    actor_rollout_ref.rollout.tensor_model_parallel_size=$infer_tp \
    actor_rollout_ref.rollout.data_parallel_size=$infer_dp \
    actor_rollout_ref.rollout.expert_parallel_size=$infer_ep \
    actor_rollout_ref.rollout.gpu_memory_utilization=$gpu_memory_utilization \
    actor_rollout_ref.rollout.n=$n_resp_per_prompt \
    actor_rollout_ref.rollout.val_kwargs.top_p=0.7 \
    actor_rollout_ref.rollout.val_kwargs.temperature=1.0 \
    actor_rollout_ref.rollout.val_kwargs.n=$n_resp_per_prompt_val \
    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=1 \
    actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=False"
########################### parameter arrays ###########################

CONFIG=(
    --config-path=./config
    --config-name=$CONFIG_NAME
)

DATA=(
    algorithm.adv_estimator=$adv_estimator
    algorithm.use_kl_in_reward=$use_kl_in_reward
    algorithm.kl_ctrl.kl_coef=$kl_coef
    algorithm.gamma=$gae_gamma
    algorithm.lam=$gae_lam
    data.train_files="$train_files"
    data.val_files="$test_files"
    data.return_raw_chat=True
    data.train_batch_size=$train_batch_size
    data.max_prompt_length=$max_prompt_length
    data.max_response_length=$max_response_length
    data.filter_overlong_prompts=True
    data.filter_overlong_prompts_workers=64
    data.truncation='error'
)

TRAINER=(
    trainer.critic_warmup=$critic_warmup
    trainer.logger=['console','wandb']
    trainer.project_name=$project_name
    trainer.experiment_name=$exp_name
    trainer.n_gpus_per_node=8
    trainer.nnodes=$nnodes
    trainer.val_before_train=False
    trainer.log_val_generations=100
    trainer.save_freq=-1
    trainer.test_freq=10
    trainer.total_epochs=10
    trainer.total_training_steps=500
)

EXTRA=(
    model_engine=$model_engine
    $ACTOR_CONFIG
    $ROLLOUT_CONFIG
)

########################### launch ###########################
python -m verl.trainer.main_ppo \
    "${CONFIG[@]}" \
    "${DATA[@]}" \
    "${TRAINER[@]}" \
    "${EXTRA[@]}" \
    "$@"