diff --git a/verl_ascend_practice/patches/mtp_checkpoint_engine_reoffload.patch b/verl_ascend_practice/patches/mtp_checkpoint_engine_reoffload.patch
new file mode 100644
index 0000000..0311978
--- /dev/null
+++ b/verl_ascend_practice/patches/mtp_checkpoint_engine_reoffload.patch
@@ -0,0 +1,12 @@
+diff --git a/verl/workers/engine_workers.py b/verl/workers/engine_workers.py
+index 57b7da64e..97638bd51 100644
+--- a/verl/workers/engine_workers.py
++++ b/verl/workers/engine_workers.py
+@@ -755,4 +755,7 @@ class ActorRolloutRefWorker(Worker, DistProfilerExtension):
+                 return metrics or {}
+             per_tensor_param, _ = self.actor.engine.get_per_tensor_param()
+             metrics = await self.checkpoint_engine.send_weights(per_tensor_param, global_steps=global_steps)
++            if self.actor.engine.is_param_offload_enabled:
++                self.actor.engine.to("cpu", model=True, optimizer=False, grad=False)
++            aggressive_empty_cache(force_sync=True)
+             return metrics or {}
diff --git a/verl_ascend_practice/reports/mtp/README.md b/verl_ascend_practice/reports/mtp/README.md
new file mode 100644
index 0000000..4112970
--- /dev/null
+++ b/verl_ascend_practice/reports/mtp/README.md
@@ -0,0 +1,195 @@
+# MiMo-7B-RL MTP Ascend 整体交付报告
+
+对应任务：[verl-ascend-recipe #20](https://github.com/verl-project/verl-ascend-recipe/issues/20)
+
+## 1. 交付概览
+
+本交付提供 MiMo-7B-RL MTP 在 Ascend NPU 上的可复现训练 recipe。训练侧使用 Megatron
+与 MBridge 管理 actor 和 reference model，rollout 侧使用 SGLang-Ascend，训练入口为
+`verl.trainer.main_ppo`。
+
+| 项目 | 配置 |
+| --- | --- |
+| 模型 | XiaomiMiMo/MiMo-7B-RL |
+| 数据集 | MATH parquet |
+| 算法 | GRPO + MTP auxiliary loss |
+| 训练后端 | Megatron actor/reference |
+| Rollout 后端 | SGLang-Ascend |
+| 验证平台 | Atlas 800T A2，4 x Ascend 910B |
+| 训练规模 | 100 global steps |
+| 运行脚本 | `verl_ascend_practice/run_mimo_7b_mtp_megatron_npu.sh` |
+| 必要补丁 | `verl_ascend_practice/patches/mtp_checkpoint_engine_reoffload.patch` |
+
+## 2. 适配方案
+
+MTP 训练通过以下配置启用：
+
+```text
+actor_rollout_ref.model.mtp.enable=True
+actor_rollout_ref.model.mtp.enable_train=True
+actor_rollout_ref.model.mtp.enable_rollout=False
+actor_rollout_ref.model.mtp.mtp_loss_scaling_factor=0.1
+actor_rollout_ref.model.mtp.detach_encoder=True
+actor_rollout_ref.rollout.name=sglang
+actor_rollout_ref.rollout.checkpoint_engine.backend=nccl
+model_engine=megatron
+```
+
+MiMo-7B-RL 的 MTP head 参与 Megatron actor 训练，rollout 仍使用主模型生成。SGLang 在生成
+阶段计算 rollout log probability，`rollout_correction.bypass_mode` 将其复用为 PPO old-policy
+log probability。权重更新复用 verl 通用 checkpoint engine；必要补丁在同步完成后恢复已启用
+的 actor parameter offload。
+
+```text
+MATH prompts
+      |
+      v
+SGLang-Ascend TP1 rollout (4 replicas)
+      |
+      v
+math reward + GRPO advantage
+      |
+      v
+PPO objective + MTP auxiliary loss
+      |
+      v
+Megatron TP2 actor update
+      |
+      v
+generic checkpoint-engine weight synchronization
+      |
+      v
+actor parameter re-offload
+```
+
+### 2.1 关键训练配置
+
+| 配置项 | 值 |
+| --- | ---: |
+| train / validation batch size | 16 / 16 |
+| PPO mini batch size | 8 |
+| micro batch size per NPU | 1 |
+| rollout responses per prompt | 2 |
+| prompt / response length | 512 / 1024 |
+| actor learning rate | `1e-6` |
+| actor TP / PP / CP | 2 / 1 / 1 |
+| rollout TP / replicas | 1 / 4 |
+| rollout max sequences per replica | 8 |
+| rollout max batched tokens | 8192 |
+| rollout memory utilization | 0.45 |
+| graph max batch size | 8 |
+| weight synchronization bucket | 1280 MB |
+| MTP loss scaling factor | 0.1 |
+| precision | BF16 |
+| checkpoint interval | 100 steps |
+
+actor 启用参数、梯度和 optimizer offload，并使用完整 activation recompute 与动态 token
+batch。SGLang-Ascend 启用 graph mode、chunked prefill 和 prefix caching。
+
+## 3. 环境与数据准备
+
+### 3.1 已验证软件环境
+
+| 组件 | 版本 |
+| --- | --- |
+| CANN | 9.0 |
+| Python | 3.11.15 |
+| torch / torch-npu | 2.8.0 / 2.8.0.post2 |
+| SGLang / sgl-kernel-npu | 0.5.10 / 2026.2.1 |
+| Megatron Core / MindSpeed | 0.16.0 / 0.16.0 |
+| MBridge | 0.15.1 |
+| Transformers | 5.3.0 |
+
+### 3.2 模型与数据准备
+
+下载 `XiaomiMiMo/MiMo-7B-RL`，并将模型 `config.json` 中的
+`max_position_embeddings` 设置为 `32768`。
+
+在 verl 根目录生成 MATH parquet：
+
+```bash
+python3 examples/data_preprocess/math_dataset.py \
+  --local_save_dir /path/to/math
+```
+
+数据目录应包含：
+
+```text
+/path/to/math/train.parquet
+/path/to/math/test.parquet
+```
+
+## 4. 运行与恢复
+
+在 verl 根目录应用必要补丁并启动训练：
+
+```bash
+git apply /path/to/verl-ascend-recipe/verl_ascend_practice/patches/mtp_checkpoint_engine_reoffload.patch
+
+DEVICE=npu \
+MODEL_PATH=/path/to/MiMo-7B-RL \
+DATA_ROOT=/path/to/math \
+NPUS_PER_NODE=4 \
+TOTAL_TRAINING_STEPS=100 \
+bash /path/to/verl-ascend-recipe/verl_ascend_practice/run_mimo_7b_mtp_megatron_npu.sh
+```
+
+模型、数据、设备数、并行度、batch、长度、MTP loss scaling、checkpoint 和日志目录均可
+通过环境变量覆盖；额外参数会作为 Hydra overrides 继续传递给 `verl.trainer.main_ppo`。
+
+默认在 step 100 保存 checkpoint。使用相同的 `OUTPUT_DIR` 重新启动时，
+`RESUME_MODE=auto` 会恢复最新 checkpoint。训练日志默认写入 `LOG_DIR`。
+
+## 5. 100-step 长跑结果
+
+本次训练在 Atlas 800T A2 的 4 x Ascend 910B 上连续完成 100/100 global steps，并完成
+step 100 checkpoint。100 个 MTP loss 和 reward 样本均为有限值。
+
+| 指标 | 结果 |
+| --- | ---: |
+| 连续训练步数 | 100 / 100 |
+| reward 首 10 步均值 | -0.808044 |
+| reward 末 10 步均值 | 0.089099 |
+| reward 首尾窗口增量 | +0.897144 |
+| MTP loss 首 10 步均值 | 0.597910 |
+| MTP loss 末 10 步均值 | 0.355875 |
+| MTP loss 首尾窗口变化 | -0.242035 |
+| 4 NPU 端到端吞吐 | 533.286 token/s |
+| 4 NPU 稳态吞吐 | 801.923 token/s |
+| 稳态吞吐 | 200.481 token/s/NPU |
+| actor 峰值分配 / reserved 显存 | 48.116 / 54.869 GiB/NPU |
+| step 时间中位数 | 40.184 s |
+
+端到端吞吐覆盖完整 100 steps 并包含 checkpoint 保存；稳态吞吐排除 step 1 和包含
+`timing_s/save_checkpoint` 的 step。
+
+### 5.1 长跑曲线
+
+下图覆盖完整 100 steps，依次展示 reward 和 MTP loss。两项指标均保留原始值和 10-step
+moving average；reward 使用 `critic/score/mean`，MTP loss 使用
+`actor/mtp_losses/mtp_1_loss`。
+
+![MiMo-7B-RL MTP 100-step reward and loss](training_curves.png)
+
+### 5.2 性能与稳定性
+
+- reward 的末 10 步均值比首 10 步提高 0.897144。
+- MTP loss 的末 10 步均值比首 10 步降低 0.242035。
+- 4 NPU 端到端和稳态吞吐均高于无 GPU 标杆时的 100 TPS 门槛。
+- 训练连续完成 100 steps，最终 checkpoint tracker 为 100。
+
+## 6. 复现证据与验收结论
+
+完整训练日志：
+[MiMo-7B-RL MTP 100-step training log](https://gist.githubusercontent.com/RordChang/c7730e4b733b544105b1efaa14dab32b/raw/c1026b5f0d6f7d67f35b2d5cb892344bd407481c/training_100step_sanitized.log)
+
+| Issue #20 验收项 | 本次结果 |
+| --- | --- |
+| 完成 100 steps 或运行 12 小时 | 完成连续 100/100 steps |
+| reward 上升 | 首 10 步均值 -0.808044，末 10 步均值 0.089099 |
+| MTP loss 有效下降 | 首 10 步均值 0.597910，末 10 步均值 0.355875 |
+| 无 GPU 标杆时 TPS > 100 | 4 NPU 端到端吞吐 533.286 token/s |
+| 提供可复现 recipe | 提供模型、数据、环境、补丁、启动、checkpoint/resume 和日志配置 |
+
+本次结果覆盖 Issue #20 的长跑、reward、MTP loss 和性能验收项，并提供了 MiMo-7B-RL
+MTP 在 Megatron + SGLang-Ascend 组合上的完整复现入口。
diff --git a/verl_ascend_practice/reports/mtp/training_curves.png b/verl_ascend_practice/reports/mtp/training_curves.png
new file mode 100644
index 0000000..cf0807b
Binary files /dev/null and b/verl_ascend_practice/reports/mtp/training_curves.png differ
diff --git a/verl_ascend_practice/run_mimo_7b_mtp_megatron_npu.sh b/verl_ascend_practice/run_mimo_7b_mtp_megatron_npu.sh
new file mode 100644
index 0000000..6dc14bc
--- /dev/null
+++ b/verl_ascend_practice/run_mimo_7b_mtp_megatron_npu.sh
@@ -0,0 +1,297 @@
+#!/usr/bin/env bash
+# MTP | MiMo-7B | Megatron training | SGLang-Ascend rollout | Ascend NPUs by default
+# With OFFLOAD=True, apply verl_ascend_practice/patches/mtp_checkpoint_engine_reoffload.patch to verl
+# until the equivalent generic checkpoint-engine fix is available upstream.
+
+set -xeuo pipefail
+
+########################### platform and defaults ###########################
+
+DEVICE=${DEVICE:-npu}
+MODEL_PATH=${MODEL_PATH:-XiaomiMiMo/MiMo-7B-RL}
+NNODES=${NNODES:-1}
+
+if [[ "$DEVICE" == "npu" ]]; then
+    devices_per_node=${NPUS_PER_NODE:-4}
+    export RAY_EXPERIMENTAL_NOSET_ASCEND_RT_VISIBLE_DEVICES=1
+    export TOKENIZERS_PARALLELISM=${TOKENIZERS_PARALLELISM:-false}
+    export HYDRA_FULL_ERROR=${HYDRA_FULL_ERROR:-1}
+    export RAY_DEDUP_LOGS=${RAY_DEDUP_LOGS:-0}
+    export TASK_QUEUE_ENABLE=${TASK_QUEUE_ENABLE:-1}
+    export CPU_AFFINITY_CONF=${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}
+    export HCCL_HOST_SOCKET_PORT_RANGE=${HCCL_HOST_SOCKET_PORT_RANGE:-60000-60050}
+    export HCCL_NPU_SOCKET_PORT_RANGE=${HCCL_NPU_SOCKET_PORT_RANGE:-61000-61050}
+
+    train_batch_size=${TRAIN_BATCH_SIZE:-16}
+    val_batch_size=${VAL_BATCH_SIZE:-16}
+    ppo_mini_batch_size=${PPO_MINI_BATCH_SIZE:-8}
+    ppo_micro_batch_size=${PPO_MICRO_BATCH_SIZE:-1}
+    max_prompt_length=${MAX_PROMPT_LENGTH:-512}
+    max_response_length=${MAX_RESPONSE_LENGTH:-1024}
+    ppo_max_token_len_per_gpu=${PPO_MAX_TOKEN_LEN_PER_GPU:-4096}
+    dataloader_num_workers=${DATALOADER_NUM_WORKERS:-0}
+
+    actor_tp=${ACTOR_TP:-2}
+    actor_pp=${ACTOR_PP:-1}
+    actor_cp=${ACTOR_CP:-1}
+    rollout_tp=${ROLLOUT_TP:-1}
+    rollout_n=${ROLLOUT_N:-2}
+    rollout_gpu_mem_util=${ROLLOUT_GPU_MEM_UTIL:-0.45}
+    rollout_max_num_seqs=${ROLLOUT_MAX_NUM_SEQS:-8}
+    rollout_max_num_batched_tokens=${ROLLOUT_MAX_NUM_BATCHED_TOKENS:-8192}
+    rollout_cuda_graph_max_bs=${ROLLOUT_CUDA_GRAPH_MAX_BS:-8}
+    checkpoint_bucket_mb=${CHECKPOINT_BUCKET_MB:-1280}
+
+    total_epochs=${TOTAL_EPOCHS:-10}
+    total_training_steps=${TOTAL_TRAINING_STEPS:-100}
+    save_freq=${SAVE_FREQ:-100}
+    test_freq=${TEST_FREQ:--1}
+    lr_warmup_steps=${LR_WARMUP_STEPS:-10}
+
+    data_root=${DATA_ROOT:-$HOME/data/math}
+    train_file=${TRAIN_FILE:-$data_root/train.parquet}
+    val_file=${VAL_FILE:-$data_root/test.parquet}
+    project_name=${PROJECT_NAME:-verl_mtp}
+    experiment_name=${EXPERIMENT_NAME:-mimo_7b_mtp_sglang_megatron_ascend}
+    logger='["console"]'
+else
+    devices_per_node=${NGPUS_PER_NODE:-8}
+    export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+    train_batch_size=${TRAIN_BATCH_SIZE:-128}
+    ppo_mini_batch_size=${PPO_MINI_BATCH_SIZE:-32}
+    max_prompt_length=${MAX_PROMPT_LENGTH:-2048}
+    max_response_length=${MAX_RESPONSE_LENGTH:-8192}
+    ppo_max_token_len_per_gpu=${PPO_MAX_TOKEN_LEN_PER_GPU:-20480}
+
+    actor_tp=${ACTOR_TP:-2}
+    actor_pp=${ACTOR_PP:-2}
+    actor_cp=${ACTOR_CP:-2}
+    rollout_tp=${ROLLOUT_TP:-4}
+    rollout_n=${ROLLOUT_N:-16}
+    rollout_gpu_mem_util=${ROLLOUT_GPU_MEM_UTIL:-0.8}
+
+    total_epochs=${TOTAL_EPOCHS:-10}
+    total_training_steps=${TOTAL_TRAINING_STEPS:-400}
+    save_freq=${SAVE_FREQ:--1}
+    test_freq=${TEST_FREQ:-10}
+    lr_warmup_steps=${LR_WARMUP_STEPS:-10}
+
+    train_file=${TRAIN_FILE:-$HOME/data/dapo-math-17k/train.parquet}
+    val_file=${VAL_FILE:-$HOME/data/aime-2024/test.parquet}
+    project_name=${PROJECT_NAME:-verl_mtp}
+    experiment_name=${EXPERIMENT_NAME:-mimo_7b_mtp_sglang_megatron}
+    logger='["console","wandb"]'
+fi
+
+actor_lr=${ACTOR_LR:-1e-6}
+entropy_coeff=${ENTROPY_COEFF:-0}
+clip_ratio_low=${CLIP_RATIO_LOW:-0.2}
+clip_ratio_high=${CLIP_RATIO_HIGH:-0.28}
+mtp_loss_scaling_factor=${MTP_LOSS_SCALING_FACTOR:-0.1}
+offload=${OFFLOAD:-True}
+
+
+########################### parameter arrays ###########################
+
+ALGORITHM=(
+    algorithm.adv_estimator=grpo
+    algorithm.use_kl_in_reward=False
+)
+
+DATA=(
+    "data.train_files=['$train_file']"
+    "data.val_files=['$val_file']"
+    data.train_batch_size=${train_batch_size}
+    data.max_prompt_length=${max_prompt_length}
+    data.max_response_length=${max_response_length}
+    data.trust_remote_code=True
+)
+
+MODEL=(
+    actor_rollout_ref.model.path="$MODEL_PATH"
+    actor_rollout_ref.model.use_remove_padding=True
+    actor_rollout_ref.model.trust_remote_code=True
+    actor_rollout_ref.model.mtp.enable=True
+    actor_rollout_ref.model.mtp.enable_train=True
+    actor_rollout_ref.model.mtp.enable_rollout=False
+    actor_rollout_ref.model.mtp.mtp_loss_scaling_factor=${mtp_loss_scaling_factor}
+    actor_rollout_ref.model.mtp.detach_encoder=True
+)
+
+ACTOR=(
+    actor_rollout_ref.actor.optim.lr=${actor_lr}
+    actor_rollout_ref.actor.optim.lr_warmup_steps=${lr_warmup_steps}
+    actor_rollout_ref.actor.optim.weight_decay=0.1
+    actor_rollout_ref.actor.optim.clip_grad=1.0
+    actor_rollout_ref.actor.ppo_mini_batch_size=${ppo_mini_batch_size}
+    actor_rollout_ref.actor.use_dynamic_bsz=True
+    actor_rollout_ref.actor.ppo_max_token_len_per_gpu=${ppo_max_token_len_per_gpu}
+    actor_rollout_ref.actor.use_kl_loss=False
+    actor_rollout_ref.actor.entropy_coeff=${entropy_coeff}
+    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.clip_ratio_c=10.0
+    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.context_parallel_size=${actor_cp}
+    actor_rollout_ref.actor.megatron.param_offload=${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
+)
+
+ROLLOUT=(
+    actor_rollout_ref.rollout.name=sglang
+    actor_rollout_ref.rollout.tensor_model_parallel_size=${rollout_tp}
+    actor_rollout_ref.rollout.gpu_memory_utilization=${rollout_gpu_mem_util}
+    actor_rollout_ref.rollout.n=${rollout_n}
+    actor_rollout_ref.rollout.log_prob_use_dynamic_bsz=True
+    actor_rollout_ref.rollout.log_prob_max_token_len_per_gpu=${ppo_max_token_len_per_gpu}
+)
+
+REF=(
+    actor_rollout_ref.ref.log_prob_use_dynamic_bsz=True
+    actor_rollout_ref.ref.log_prob_max_token_len_per_gpu=${ppo_max_token_len_per_gpu}
+    actor_rollout_ref.ref.megatron.tensor_model_parallel_size=${actor_tp}
+    actor_rollout_ref.ref.megatron.pipeline_model_parallel_size=${actor_pp}
+    actor_rollout_ref.ref.megatron.context_parallel_size=${actor_cp}
+    actor_rollout_ref.ref.megatron.param_offload=${offload}
+)
+
+REWARD=(
+    reward.reward_manager.name=dapo
+    +reward.reward_kwargs.overlong_buffer_cfg.enable=True
+    +reward.reward_kwargs.overlong_buffer_cfg.penalty_factor=1.0
+    +reward.reward_kwargs.max_resp_len=${max_response_length}
+)
+
+TRAINER=(
+    trainer.balance_batch=True
+    trainer.critic_warmup=0
+    "trainer.logger=$logger"
+    trainer.project_name=${project_name}
+    trainer.experiment_name=${experiment_name}
+    trainer.n_gpus_per_node=${devices_per_node}
+    trainer.nnodes=${NNODES}
+    trainer.val_before_train=False
+    trainer.save_freq=${save_freq}
+    trainer.test_freq=${test_freq}
+    trainer.total_epochs=${total_epochs}
+    trainer.total_training_steps=${total_training_steps}
+)
+
+if [[ "$DEVICE" == "npu" ]]; then
+    output_dir=${OUTPUT_DIR:-$PWD/checkpoints/$experiment_name}
+    log_dir=${LOG_DIR:-$PWD/logs}
+    resume_mode=${RESUME_MODE:-auto}
+    max_actor_ckpt_to_keep=${MAX_ACTOR_CKPT_TO_KEEP:-1}
+    mkdir -p "$output_dir" "$log_dir"
+
+    ALGORITHM+=(
+        algorithm.rollout_correction.bypass_mode=True
+        algorithm.rollout_correction.loss_type=ppo_clip
+    )
+    DATA+=(
+        data.val_batch_size=${val_batch_size}
+        data.dataloader_num_workers=${dataloader_num_workers}
+        data.filter_overlong_prompts=True
+        data.truncation=error
+    )
+    ACTOR+=(
+        actor_rollout_ref.actor.optim.use_checkpoint_opt_param_scheduler=True
+        actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=${ppo_micro_batch_size}
+        actor_rollout_ref.actor.megatron.vanilla_mbridge=False
+        +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_method=uniform
+        +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_granularity=full
+        +actor_rollout_ref.actor.megatron.override_transformer_config.recompute_num_layers=1
+        +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.max_model_len=$((max_prompt_length + max_response_length))
+        actor_rollout_ref.rollout.max_num_seqs=${rollout_max_num_seqs}
+        actor_rollout_ref.rollout.max_num_batched_tokens=${rollout_max_num_batched_tokens}
+        actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=${ppo_micro_batch_size}
+        actor_rollout_ref.rollout.calculate_log_probs=True
+        actor_rollout_ref.rollout.enable_chunked_prefill=True
+        actor_rollout_ref.rollout.enable_prefix_caching=True
+        actor_rollout_ref.rollout.enforce_eager=False
+        actor_rollout_ref.rollout.free_cache_engine=True
+        actor_rollout_ref.rollout.checkpoint_engine.backend=nccl
+        actor_rollout_ref.rollout.checkpoint_engine.update_weights_bucket_megabytes=${checkpoint_bucket_mb}
+        +actor_rollout_ref.rollout.engine_kwargs.sglang.attention_backend=ascend
+        +actor_rollout_ref.rollout.engine_kwargs.sglang.cuda_graph_max_bs=${rollout_cuda_graph_max_bs}
+        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
+        actor_rollout_ref.rollout.val_kwargs.top_k=-1
+        actor_rollout_ref.rollout.val_kwargs.do_sample=True
+    )
+    REF+=(
+        actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=${ppo_micro_batch_size}
+        actor_rollout_ref.ref.megatron.vanilla_mbridge=False
+    )
+    REWARD+=(
+        +reward.reward_kwargs.overlong_buffer_cfg.len=256
+        +reward.reward_kwargs.overlong_buffer_cfg.log=False
+    )
+    TRAINER+=(
+        trainer.device=npu
+        trainer.resume_mode=${resume_mode}
+        trainer.max_actor_ckpt_to_keep=${max_actor_ckpt_to_keep}
+        "trainer.default_local_dir=$output_dir"
+    )
+else
+    DATA+=(data.truncation=left)
+    REWARD+=(+reward.reward_kwargs.overlong_buffer_cfg.len=4096)
+fi
+
+########################### launch ###########################
+
+EXTRA=(
+    model_engine=megatron
+)
+
+# uv (set VERL_USE_UV=0 for system python): on GPU, the driver and every Ray worker
+# (runtime_env.py_executable) run through `uv run` on the sglang × megatron extras of the committed uv.lock;
+# NPU falls back to ambient python. Run from the verl repo root.
+LAUNCH=(python3)
+RAY=(ray_kwargs.ray_init.runtime_env.py_executable=null)
+if [[ "${VERL_USE_UV:-1}" != 0 && "$DEVICE" != "npu" ]]; then
+    LAUNCH=(uv run --frozen --all-packages --extra sglang --extra megatron python3)
+    RAY=(ray_kwargs.ray_init.runtime_env.py_executable="uv -v run --frozen --all-packages --extra sglang --extra megatron")
+fi
+
+COMMAND=(
+    "${LAUNCH[@]}"
+    -m verl.trainer.main_ppo
+    "${DATA[@]}"
+    "${ALGORITHM[@]}"
+    "${MODEL[@]}"
+    "${ACTOR[@]}"
+    "${ROLLOUT[@]}"
+    "${REF[@]}"
+    "${REWARD[@]}"
+    "${TRAINER[@]}"
+    "${EXTRA[@]}"
+    "${RAY[@]}"
+    "$@"
+)
+
+if [[ "$DEVICE" == "npu" ]]; then
+    log_file="$log_dir/${experiment_name}_$(date +%Y%m%d_%H%M%S).log"
+    PYTHONUNBUFFERED=1 "${COMMAND[@]}" 2>&1 | tee "$log_file"
+else
+    "${COMMAND[@]}"
+fi