verl-project · RordChang · Aug 16, 2026 · Aug 18, 2026
diff --git a/examples/cispo_trainer/README.md b/examples/cispo_trainer/README.md
@@ -8,10 +8,24 @@ Reference: [MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Att
 
 | Script                               | Infer | Train | Platform |
 |--------------------------------------|-------|-------|----------|
-| `run_qwen3_8b_fsdp.sh`          | vLLM  | FSDP  | NVIDIA   |
+| `run_qwen3_8b_fsdp.sh`               | vLLM        | FSDP     | NVIDIA    |
+| `run_qwen2_5_0_5b_megatron.sh`       | vLLM-Ascend | Megatron | Ascend NPU |
 
 ## Key Flags
 
 - `actor_rollout_ref.actor.policy_loss.loss_mode=cispo`
 - `actor_rollout_ref.actor.clip_ratio_low=10` (effectively unclamped on lower side)
 - `actor_rollout_ref.actor.clip_ratio_high=0.2`
+
+## Megatron + vLLM-Ascend
+
+```bash
+MODEL_PATH=/path/to/Qwen2.5-0.5B-Instruct \
+DATA_ROOT=/path/to/data \
+NPUS_PER_NODE=4 \
+bash examples/cispo_trainer/run_qwen2_5_0_5b_megatron.sh
+```
+
+Ensure the container provides enough `/dev/shm` capacity for the configured weight-transfer bucket.
+See [verl-ascend-recipe issue #17](https://github.com/verl-project/verl-ascend-recipe/issues/17)
+for the validated environment, training logs, and 100-step results.
diff --git a/examples/cispo_trainer/run_qwen2_5_0_5b_megatron.sh b/examples/cispo_trainer/run_qwen2_5_0_5b_megatron.sh
@@ -0,0 +1,206 @@
+#!/usr/bin/env bash
+# CISPO | Qwen2.5-0.5B-Instruct | Megatron training | vLLM-Ascend rollout | Ascend NPUs
+
+set -xeuo pipefail
+
+########################### environment ###########################
+
+export TOKENIZERS_PARALLELISM=${TOKENIZERS_PARALLELISM:-false}
+export HYDRA_FULL_ERROR=${HYDRA_FULL_ERROR:-1}
+export RAY_DEDUP_LOGS=${RAY_DEDUP_LOGS:-0}
+export VLLM_USE_V1=${VLLM_USE_V1:-1}
+export VLLM_ALLREDUCE_USE_SYMM_MEM=${VLLM_ALLREDUCE_USE_SYMM_MEM:-0}
+export VLLM_ASCEND_ENABLE_NZ=${VLLM_ASCEND_ENABLE_NZ:-0}
+export TASK_QUEUE_ENABLE=${TASK_QUEUE_ENABLE:-2}
+export CPU_AFFINITY_CONF=${CPU_AFFINITY_CONF:-1}
+export HCCL_OP_EXPANSION_MODE=${HCCL_OP_EXPANSION_MODE:-AIV}
+export HCCL_ASYNC_ERROR_HANDLING=${HCCL_ASYNC_ERROR_HANDLING:-0}
+export HCCL_EXEC_TIMEOUT=${HCCL_EXEC_TIMEOUT:-3600}
+export HCCL_CONNECT_TIMEOUT=${HCCL_CONNECT_TIMEOUT:-3600}
+
+
+########################### user-adjustable ###########################
+
+MODEL_PATH=${MODEL_PATH:-Qwen/Qwen2.5-0.5B-Instruct}
+DATA_ROOT=${DATA_ROOT:-${HOME}/data}
+NNODES=${NNODES:-1}
+NPUS_PER_NODE=${NPUS_PER_NODE:-4}
+
+train_batch_size=${TRAIN_BATCH_SIZE:-32}
+ppo_mini_batch_size=${PPO_MINI_BATCH_SIZE:-32}
+micro_batch_size=${MICRO_BATCH_SIZE:-4}
+max_prompt_length=${MAX_PROMPT_LENGTH:-512}
+max_response_length=${MAX_RESPONSE_LENGTH:-512}
+max_token_length=${MAX_TOKEN_LENGTH:-$((max_prompt_length + max_response_length))}
+dataloader_num_workers=${DATALOADER_NUM_WORKERS:-0}
+seed=${SEED:-42}
+
+actor_lr=${ACTOR_LR:-1e-6}
+kl_loss_coef=${KL_LOSS_COEF:-0.001}
+clip_ratio_low=${CLIP_RATIO_LOW:-10.0}
+clip_ratio_high=${CLIP_RATIO_HIGH:-0.2}
+
+actor_tp=${ACTOR_TP:-2}
+actor_pp=${ACTOR_PP:-1}
+rollout_tp=${ROLLOUT_TP:-2}
+rollout_n=${ROLLOUT_N:-4}
+rollout_gpu_memory_utilization=${ROLLOUT_GPU_MEMORY_UTILIZATION:-0.50}
+rollout_max_num_batched_tokens=${ROLLOUT_MAX_NUM_BATCHED_TOKENS:-8192}
+rollout_enforce_eager=${ROLLOUT_ENFORCE_EAGER:-True}
+weight_bucket_mb=${WEIGHT_BUCKET_MB:-512}
+
+rollout_world_size=$((NNODES * NPUS_PER_NODE))
+if (( rollout_tp <= 0 || rollout_world_size % rollout_tp != 0 )); then
+    echo "ROLLOUT_TP must be a positive divisor of NNODES * NPUS_PER_NODE." >&2
+    exit 2
+fi
+if (( rollout_n < 2 )); then
+    echo "CISPO requires ROLLOUT_N >= 2." >&2
+    exit 2
+fi
+rollout_replicas=$((rollout_world_size / rollout_tp))
+default_rollout_max_num_seqs=$(((train_batch_size * rollout_n + rollout_replicas - 1) / rollout_replicas))
+rollout_max_num_seqs=${ROLLOUT_MAX_NUM_SEQS:-${default_rollout_max_num_seqs}}
+
+offload=${OFFLOAD:-False}
+total_training_steps=${TOTAL_TRAINING_STEPS:-100}
+total_epochs=${TOTAL_EPOCHS:-1}
+save_freq=${SAVE_FREQ:--1}
+test_freq=${TEST_FREQ:--1}
+resume_mode=${RESUME_MODE:-auto}
+max_actor_ckpt_to_keep=${MAX_ACTOR_CKPT_TO_KEEP:-1}
+
+project_name=${PROJECT_NAME:-verl_cispo_gsm8k}
+experiment_name=${EXPERIMENT_NAME:-qwen2_5_0_5b_cispo_megatron_vllm_ascend}
+output_dir=${OUTPUT_DIR:-${PWD}/checkpoints/${experiment_name}}
+log_dir=${LOG_DIR:-${PWD}/logs}
+mkdir -p "${output_dir}" "${log_dir}"
+
+train_files="['${DATA_ROOT}/gsm8k/train.parquet']"
+val_files="['${DATA_ROOT}/gsm8k/test.parquet']"
+
+########################### parameter arrays ###########################
+
+ALGORITHM=(
+    algorithm.adv_estimator=grpo
+    algorithm.use_kl_in_reward=False
+)
+
+DATA=(
+    "data.train_files=${train_files}"
+    "data.val_files=${val_files}"
+    data.train_batch_size=${train_batch_size}
+    data.max_prompt_length=${max_prompt_length}
+    data.max_response_length=${max_response_length}
+    data.dataloader_num_workers=${dataloader_num_workers}
+    data.seed=${seed}
+    data.filter_overlong_prompts=True
+    data.truncation=error
+)
+
+MODEL=(
+    "actor_rollout_ref.model.path=${MODEL_PATH}"
+    actor_rollout_ref.model.use_remove_padding=True
+    actor_rollout_ref.model.enable_gradient_checkpointing=False
+)
+
+ACTOR=(
+    actor_rollout_ref.actor.policy_loss.loss_mode=cispo
+    actor_rollout_ref.actor.loss_agg_mode=token-mean
+    actor_rollout_ref.actor.clip_ratio_low=${clip_ratio_low}
+    actor_rollout_ref.actor.clip_ratio_high=${clip_ratio_high}
+    actor_rollout_ref.actor.optim.lr=${actor_lr}
+    actor_rollout_ref.actor.ppo_mini_batch_size=${ppo_mini_batch_size}
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=${micro_batch_size}
+    actor_rollout_ref.actor.use_dynamic_bsz=True
+    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=${max_token_length}
+    actor_rollout_ref.actor.use_kl_loss=True
+    actor_rollout_ref.actor.kl_loss_coef=${kl_loss_coef}
+    actor_rollout_ref.actor.kl_loss_type=low_var_kl
+    actor_rollout_ref.actor.entropy_coeff=0
+    actor_rollout_ref.actor.megatron.tensor_model_parallel_size=${actor_tp}
+    actor_rollout_ref.actor.megatron.pipeline_model_parallel_size=${actor_pp}
+    actor_rollout_ref.actor.megatron.param_offload=${offload}
+    actor_rollout_ref.actor.megatron.grad_offload=${offload}
+    actor_rollout_ref.actor.megatron.optimizer_offload=${offload}
+    actor_rollout_ref.actor.megatron.use_mbridge=True
+    actor_rollout_ref.actor.megatron.vanilla_mbridge=False
+    actor_rollout_ref.actor.megatron.dtype=bfloat16
+    +actor_rollout_ref.actor.megatron.override_transformer_config.apply_rope_fusion=True
+    +actor_rollout_ref.actor.megatron.override_transformer_config.position_embedding_type=rope
+    +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_rotary_pos_emb=True
+    +actor_rollout_ref.actor.megatron.override_transformer_config.normalization=RMSNorm
+    +actor_rollout_ref.actor.megatron.override_transformer_config.use_fused_rmsnorm=True
+    ++actor_rollout_ref.actor.megatron.override_transformer_config.attention_backend=flash
+    +actor_rollout_ref.actor.megatron.override_transformer_config.use_flash_attn=True
+)
+
+ROLLOUT=(
+    actor_rollout_ref.rollout.name=vllm
+    actor_rollout_ref.rollout.tensor_model_parallel_size=${rollout_tp}
+    actor_rollout_ref.rollout.gpu_memory_utilization=${rollout_gpu_memory_utilization}
+    actor_rollout_ref.rollout.max_model_len=${max_token_length}
+    actor_rollout_ref.rollout.max_num_seqs=${rollout_max_num_seqs}
+    actor_rollout_ref.rollout.max_num_batched_tokens=${rollout_max_num_batched_tokens}
+    actor_rollout_ref.rollout.checkpoint_engine.update_weights_bucket_megabytes=${weight_bucket_mb}
+    actor_rollout_ref.rollout.n=${rollout_n}
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=${micro_batch_size}
+    actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=True
+    actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=${max_token_length}
+    actor_rollout_ref.rollout.calculate_log_probs=False
+    actor_rollout_ref.rollout.enable_chunked_prefill=True
+    actor_rollout_ref.rollout.enable_prefix_caching=True
+    actor_rollout_ref.rollout.enforce_eager=${rollout_enforce_eager}
+    actor_rollout_ref.rollout.free_cache_engine=True
+    actor_rollout_ref.rollout.val_kwargs.n=1
+    actor_rollout_ref.rollout.val_kwargs.temperature=1.0
+    actor_rollout_ref.rollout.val_kwargs.top_p=0.7
+)
+
+REF=(
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=${micro_batch_size}
+    actor_rollout_ref.ref.log_prob_use_dynamic_bsz=True
+    actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=${max_token_length}
+    actor_rollout_ref.ref.megatron.tensor_model_parallel_size=${actor_tp}
+    actor_rollout_ref.ref.megatron.pipeline_model_parallel_size=${actor_pp}
+    actor_rollout_ref.ref.megatron.param_offload=${offload}
+    actor_rollout_ref.ref.megatron.use_mbridge=True
+    actor_rollout_ref.ref.megatron.vanilla_mbridge=False
+)
+
+TRAINER=(
+    trainer.balance_batch=True
+    trainer.critic_warmup=0
+    'trainer.logger=["console"]'
+    trainer.project_name=${project_name}
+    trainer.experiment_name=${experiment_name}
+    trainer.n_gpus_per_node=${NPUS_PER_NODE}
+    trainer.nnodes=${NNODES}
+    trainer.device=npu
+    trainer.val_before_train=False
+    trainer.save_freq=${save_freq}
+    trainer.test_freq=${test_freq}
+    trainer.resume_mode=${resume_mode}
+    trainer.max_actor_ckpt_to_keep=${max_actor_ckpt_to_keep}
+    trainer.total_epochs=${total_epochs}
+    trainer.total_training_steps=${total_training_steps}
+    "trainer.default_local_dir=${output_dir}"
+)
+
+EXTRA=(
+    model_engine=megatron
+)
+
+########################### launch ###########################
+
+log_file="${log_dir}/${experiment_name}_$(date +%Y%m%d_%H%M%S).log"
+PYTHONUNBUFFERED=1 python3 -m verl.trainer.main_ppo \
+    "${DATA[@]}" \
+    "${ALGORITHM[@]}" \
+    "${MODEL[@]}" \
+    "${ROLLOUT[@]}" \
+    "${ACTOR[@]}" \
+    "${REF[@]}" \
+    "${TRAINER[@]}" \
+    "${EXTRA[@]}" \
+    "$@" 2>&1 | tee "${log_file}"