diff --git a/examples/diffusionnft_trainer/README.md b/examples/diffusionnft_trainer/README.md
index 9afe08073..45fb19ab1 100644
--- a/examples/diffusionnft_trainer/README.md
+++ b/examples/diffusionnft_trainer/README.md
@@ -77,6 +77,9 @@ Launch the example from the repository root:
 bash examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora.sh
 ```
 
+For CUDA V1 sync (TransferQueue + ReplayBuffer), use
+`examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_v1.sh`.
+
 ### NPU
 
 For Huawei Ascend NPUs, use the NPU-optimized script:
@@ -85,6 +88,9 @@ For Huawei Ascend NPUs, use the NPU-optimized script:
 bash examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_npu.sh
 ```
 
+For NPU V1 sync, use
+`examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_npu_v1.sh`.
+
 This script uses a 16-NPU global distribution strategy with:
 - `actor_rollout_ref.model.attn_backend='_native_npu'`
 - `actor_rollout_ref.rollout.tensor_model_parallel_size=2`
diff --git a/examples/diffusionnft_trainer/minimax_h3/README.md b/examples/diffusionnft_trainer/minimax_h3/README.md
index 26ed1665d..614227c48 100644
--- a/examples/diffusionnft_trainer/minimax_h3/README.md
+++ b/examples/diffusionnft_trainer/minimax_h3/README.md
@@ -192,6 +192,9 @@ export DATA_DIR=/path/to/h3_t2va_data
 bash examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_t2va_lora.sh
 ```
 
+For CUDA V1 sync (TransferQueue + ReplayBuffer), use
+`examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_t2va_lora_v1.sh`.
+
 MiniMax H3 t2va requires an explicit named `aspect_ratio` (one of
 `21:9/16:9/4:3/1:1/3:4/9:16`); the launch script sets `16:9` and explicit
 `height`/`width` control the actual canvas (must be multiples of 32).
@@ -259,6 +262,9 @@ DATA_DIR=/path/to/parquet \
 bash examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_fl2va_lora.sh
 ```
 
+For CUDA V1 sync, use
+`examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_fl2va_lora_v1.sh`.
+
 The latest vLLM-Omni contract requires 4–15 seconds at 24 FPS. The launcher's
 `NUM_FRAMES=96` is aligned by vLLM-Omni to the next valid `17n+5` boundary.
 Sampling edges must be multiples of 32 (the H3 pipeline silently floors
@@ -348,6 +354,9 @@ VAL_REF_IMAGE_SHORT_EDGE=1024 \
 bash examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_ref2va_lora.sh
 ```
 
+For CUDA V1 sync, use
+`examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_ref2va_lora_v1.sh`.
+
 The H3 Agent Loop keeps the user prompt token-ID-native while vLLM-Omni adds
 the reference presentation. The default `MAX_PROMPT_EMBEDS=12288` is both the
 prompt-embedding limit and the fixed transport limit for each video/audio
diff --git a/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_fl2va_lora_v1.sh b/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_fl2va_lora_v1.sh
new file mode 100644
index 000000000..540e24622
--- /dev/null
+++ b/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_fl2va_lora_v1.sh
@@ -0,0 +1,176 @@
+#!/usr/bin/env bash
+# MiniMax H3 FL2VA (image-conditioned) DiffusionNFT LoRA recipe
+# (V1 trainer: TransferQueue + ReplayBuffer + sync mode).
+#
+# This is the v1 counterpart of run_minimax_h3_fl2va_lora.sh. It uses
+# `verl_omni.trainer.main_diffusion_v1`, which selects
+# `PolicyGradientDiffusionTrainerV1Sync` via `trainer.v1.trainer_mode=sync`.
+# TransferQueue is force-enabled inside the runner, so it does not need to be
+# set on the CLI. GPU layout stays colocated (same as v0).
+#
+# Reference (legacy v0 script):
+# verl-omni/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_fl2va_lora.sh
+set -euo pipefail
+
+export WANDB_MODE=${WANDB_MODE:-offline}
+export WANDB_RUN_ID=${WANDB_RUN_ID:-minimax_h3_fl2va_lora_v1}
+export WANDB_RESUME=${WANDB_RESUME:-allow}
+export RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+export PYTORCH_CUDA_ALLOC_CONF=${PYTORCH_CUDA_ALLOC_CONF:-expandable_segments:True}
+
+: "${DATA_DIR:?Set DATA_DIR to the parquet directory produced by prepare_data.py}"
+if [[ -z "${MODEL_PATH:-}" || ! -d "$MODEL_PATH/FL2VA" || ! -d "$MODEL_PATH/transformer" ]]; then
+    echo "MODEL_PATH must point to a MiniMax-H3 repo root containing FL2VA/ (fused rollout) and transformer/ (diffusers actor) (got: '${MODEL_PATH:-<unset>}')" >&2
+    exit 1
+fi
+
+N_GPUS=${N_GPUS:-8}
+ROLLOUT_TP=${ROLLOUT_TP:-4}
+ROLLOUT_N=${ROLLOUT_N:-16}
+HEIGHT=${HEIGHT:-288}
+WIDTH=${WIDTH:-448}
+VAL_HEIGHT=${VAL_HEIGHT:-576}
+VAL_WIDTH=${VAL_WIDTH:-928}
+NUM_FRAMES=${NUM_FRAMES:-96}
+INFER_STEPS=${INFER_STEPS:-10}
+MAX_PROMPT_EMBEDS=${MAX_PROMPT_EMBEDS:-1024}
+FRAME_INDICES=${FRAME_INDICES:-'[0]'}
+ACTOR_ATTN_BACKEND=${ACTOR_ATTN_BACKEND:-_flash_3_varlen_hub}
+ROLLOUT_ATTN_BACKEND=${ROLLOUT_ATTN_BACKEND:-FLASH_ATTN_3_HUB}
+TOTAL_TRAINING_STEPS=${TOTAL_TRAINING_STEPS:-1000}
+
+script_path=$(readlink -f "$0")
+script_name=$(basename "$script_path" .sh)
+repo_root=$(dirname "$script_path")
+while [[ "$repo_root" != "/" && ! -f "$repo_root/LICENSE" ]]; do
+    repo_root=$(dirname "$repo_root")
+done
+output_dir=${OUTPUT_DIR:-$(dirname "$script_path")/outputs/$script_name}
+checkpoint_dir=$output_dir/checkpoints
+run_timestamp=$(date +"%Y%m%d_%H%M")
+log_file=$output_dir/logs/$run_timestamp/${NODE_RANK:-0}.log
+mkdir -p "$checkpoint_dir" "$(dirname "$log_file")"
+
+lora_warmstart_arg=()
+if [[ -n "${LORA_WARMSTART_PATH:-}" ]]; then
+    lora_warmstart_arg=(actor_rollout_ref.model.lora_adapter_path=$LORA_WARMSTART_PATH)
+fi
+
+python3 -m verl_omni.trainer.main_diffusion_v1 \
+  algorithm.trainer_type=direct_preference \
+  algorithm.sample_source=online \
+  algorithm.adv_mode=continuous \
+  algorithm.timestep_fraction=1.0 \
+  algorithm.old_policy_decay_schedule=delayed_linear_to_0_999 \
+  algorithm.old_policy_update_interval=2 \
+  data.train_files="$DATA_DIR/train.parquet" \
+  data.val_files="$DATA_DIR/test.parquet" \
+  data.train_batch_size=32 \
+  data.val_max_samples=128 \
+  data.max_prompt_length=4096 \
+  data.truncation=error \
+  data.seed=42 \
+  actor_rollout_ref.model.path="$MODEL_PATH/FL2VA" \
+  actor_rollout_ref.model.tokenizer_path="$MODEL_PATH/FL2VA/tokenizer" \
+  actor_rollout_ref.model.config_path="$MODEL_PATH/transformer" \
+  +actor_rollout_ref.model.architecture=MiniMaxH3Pipeline \
+  actor_rollout_ref.model.external_lib=verl_omni.pipelines.minimax_h3_diffusion_nft \
+  actor_rollout_ref.model.algorithm=diffusion_nft \
+  actor_rollout_ref.model.model_type=diffusion_nft_model \
+  actor_rollout_ref.model.attn_backend="$ACTOR_ATTN_BACKEND" \
+  actor_rollout_ref.model.enable_gradient_checkpointing=True \
+  actor_rollout_ref.model.lora_rank=64 \
+  actor_rollout_ref.model.lora_alpha=128 \
+  "${lora_warmstart_arg[@]}" \
+  actor_rollout_ref.model.policy_state_adapters='["default","old"]' \
+  actor_rollout_ref.model.target_modules='["to_q","to_k","to_v","to_out.0","ff.net.0.proj","ff.net.2"]' \
+  actor_rollout_ref.model.fsdp_layer_prefixes="['transformer_blocks.','token_refiner.refiner_blocks.']" \
+  actor_rollout_ref.actor.strategy=fsdp2 \
+  actor_rollout_ref.actor.optim.lr=3e-4 \
+  actor_rollout_ref.actor.optim.weight_decay=1e-4 \
+  actor_rollout_ref.actor.optim.betas="[0.9,0.999]" \
+  actor_rollout_ref.actor.optim.override_optimizer_config="{eps: 1e-8}" \
+  actor_rollout_ref.actor.optim.clip_grad=1.0 \
+  actor_rollout_ref.actor.ppo_mini_batch_size=16 \
+  actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=16 \
+  actor_rollout_ref.actor.diffusion_loss.loss_mode=diffusion_nft \
+  actor_rollout_ref.actor.diffusion_loss.clip_ratio=1e-5 \
+  actor_rollout_ref.actor.diffusion_loss.mix_beta=0.1 \
+  actor_rollout_ref.actor.diffusion_loss.ref_kl_coef=0.0001 \
+  actor_rollout_ref.actor.diffusion_loss.adv_clip_max=5.0 \
+  actor_rollout_ref.actor.use_kl_loss=False \
+  actor_rollout_ref.actor.fsdp_config.model_dtype=bfloat16 \
+  actor_rollout_ref.actor.fsdp_config.param_offload=True \
+  actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+  actor_rollout_ref.actor.fsdp_config.ulysses_sequence_parallel_size=1 \
+  actor_rollout_ref.rollout.name=vllm_omni \
+  actor_rollout_ref.rollout.max_num_seqs=1 \
+  actor_rollout_ref.rollout.rollout_attn_backend="$ROLLOUT_ATTN_BACKEND" \
+  actor_rollout_ref.rollout.rollout_adapter=old \
+  actor_rollout_ref.rollout.tensor_model_parallel_size="$ROLLOUT_TP" \
+  actor_rollout_ref.rollout.n="$ROLLOUT_N" \
+  actor_rollout_ref.rollout.seed=42 \
+  actor_rollout_ref.rollout.agent.num_workers=$((N_GPUS / ROLLOUT_TP)) \
+  actor_rollout_ref.rollout.agent.default_agent_loop=minimax_h3_diffusion_single_turn_agent \
+  actor_rollout_ref.rollout.load_format=safetensors \
+  actor_rollout_ref.rollout.layered_summon=True \
+  actor_rollout_ref.rollout.calculate_log_probs=False \
+  actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \
+  +actor_rollout_ref.rollout.engine_kwargs.vllm_omni.enable_cpu_offload=True \
+  actor_rollout_ref.rollout.max_prompt_embed_length="$MAX_PROMPT_EMBEDS" \
+  actor_rollout_ref.rollout.pipeline.task=fl2va \
+  actor_rollout_ref.rollout.pipeline.frame_indices="$FRAME_INDICES" \
+  actor_rollout_ref.rollout.pipeline.height="$HEIGHT" \
+  actor_rollout_ref.rollout.pipeline.width="$WIDTH" \
+  actor_rollout_ref.rollout.pipeline.num_frames="$NUM_FRAMES" \
+  actor_rollout_ref.rollout.pipeline.frame_rate=24.0 \
+  actor_rollout_ref.rollout.pipeline.num_inference_steps="$INFER_STEPS" \
+  actor_rollout_ref.rollout.pipeline.true_cfg_scale=1.0 \
+  actor_rollout_ref.rollout.pipeline.max_sequence_length="$MAX_PROMPT_EMBEDS" \
+  actor_rollout_ref.rollout.pipeline.video_flow_shift=12.0 \
+  +actor_rollout_ref.rollout.pipeline.output_type=pt \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.height="$VAL_HEIGHT" \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.width="$VAL_WIDTH" \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.num_frames="$NUM_FRAMES" \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.frame_rate=24.0 \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.num_inference_steps=40 \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.true_cfg_scale=1.0 \
+  +actor_rollout_ref.rollout.val_kwargs.pipeline.output_type=pt \
+  actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \
+  reward.reward_model.enable=False \
+  reward.num_workers=1 \
+  reward.custom_reward_function.path=pkg://verl_omni.reward_loop.reward_manager.multi \
+  reward.custom_reward_function.name=_multi_reward_placeholder \
+  reward.reward_manager.name=MultiVisualRewardManager \
+  reward.reward_manager.module.path=pkg://verl_omni.reward_loop.reward_manager \
+  "+reward.reward_functions.clap.path=$repo_root/verl_omni/utils/reward_score/clap.py" \
+  '+reward.reward_functions.clap.name=compute_score' \
+  '+reward.reward_functions.clap.weight=1.0' \
+  '+reward.reward_functions.clap.device=cuda:0' \
+  '+reward.reward_functions.clap.model_name_or_path=laion/larger_clap_general' \
+  "+reward.reward_functions.imagebind.path=$repo_root/verl_omni/utils/reward_score/imagebind.py" \
+  '+reward.reward_functions.imagebind.name=compute_score' \
+  '+reward.reward_functions.imagebind.weight=1.0' \
+  '+reward.reward_functions.imagebind.device=cuda:1' \
+  '+reward.reward_functions.imagebind.model_name_or_path=.checkpoints/imagebind_huge.pth' \
+  '+reward.reward_functions.imagebind.mode=audio_video' \
+  reward.aggregation=weighted_sum \
+  trainer.logger='["console","tensorboard","wandb"]' \
+  trainer.project_name=diffusion_nft \
+  trainer.experiment_name=minimax_h3_fl2va_lora_v1 \
+  trainer.default_local_dir="$checkpoint_dir" \
+  trainer.validation_data_dir="$output_dir/validation_data" \
+  trainer.rollout_data_dir="$output_dir/rollout_data" \
+  trainer.rollout_data_save_freq=10 \
+  trainer.log_val_generations=8 \
+  trainer.video_fps=24 \
+  trainer.val_before_train=True \
+  trainer.n_gpus_per_node="$N_GPUS" \
+  trainer.nnodes=1 \
+  trainer.save_freq=5 \
+  trainer.test_freq=10 \
+  trainer.total_epochs=15 \
+  trainer.total_training_steps="$TOTAL_TRAINING_STEPS" \
+  trainer.use_v1=true \
+  trainer.v1.trainer_mode=sync \
+  "$@"
diff --git a/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_ref2va_lora_v1.sh b/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_ref2va_lora_v1.sh
new file mode 100644
index 000000000..42d07e499
--- /dev/null
+++ b/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_ref2va_lora_v1.sh
@@ -0,0 +1,183 @@
+#!/usr/bin/env bash
+# MiniMax H3 Ref2VA (multi-reference) DiffusionNFT LoRA recipe
+# (V1 trainer: TransferQueue + ReplayBuffer + sync mode).
+# References may mix images, videos and standalone audio (up to twelve files).
+#
+# This is the v1 counterpart of run_minimax_h3_ref2va_lora.sh. It uses
+# `verl_omni.trainer.main_diffusion_v1`, which selects
+# `PolicyGradientDiffusionTrainerV1Sync` via `trainer.v1.trainer_mode=sync`.
+# TransferQueue is force-enabled inside the runner, so it does not need to be
+# set on the CLI. GPU layout stays colocated (same as v0).
+#
+# Reference (legacy v0 script):
+# verl-omni/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_ref2va_lora.sh
+set -euo pipefail
+
+export WANDB_MODE=${WANDB_MODE:-online}
+export WANDB_RUN_ID=${WANDB_RUN_ID:-minimax_h3_ref2va_lora_v1}
+export WANDB_RESUME=${WANDB_RESUME:-allow}
+export RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+export PYTORCH_CUDA_ALLOC_CONF=${PYTORCH_CUDA_ALLOC_CONF:-expandable_segments:True}
+
+: "${DATA_DIR:?Set DATA_DIR to the parquet directory produced by prepare_ref2va_data.py}"
+if [[ -z "${MODEL_PATH:-}" || ! -d "$MODEL_PATH/Ref2VA" || ! -d "$MODEL_PATH/transformer_ref" ]]; then
+    echo "MODEL_PATH must point to a MiniMax-H3 repo root containing Ref2VA/ (fused rollout) and transformer_ref/ (diffusers actor) (got: '${MODEL_PATH:-<unset>}')" >&2
+    exit 1
+fi
+
+N_GPUS=${N_GPUS:-8}
+ROLLOUT_TP=${ROLLOUT_TP:-4}
+TEXT_ENCODER_TP=${TEXT_ENCODER_TP:-$ROLLOUT_TP}
+ROLLOUT_N=${ROLLOUT_N:-16}
+HEIGHT=${HEIGHT:-288}
+WIDTH=${WIDTH:-448}
+VAL_HEIGHT=${VAL_HEIGHT:-576}
+VAL_WIDTH=${VAL_WIDTH:-928}
+NUM_FRAMES=${NUM_FRAMES:-96}
+INFER_STEPS=${INFER_STEPS:-10}
+MAX_PROMPT_EMBEDS=${MAX_PROMPT_EMBEDS:-12288}
+REF_IMAGE_SHORT_EDGE=${REF_IMAGE_SHORT_EDGE:-2048}
+VAL_REF_IMAGE_SHORT_EDGE=${VAL_REF_IMAGE_SHORT_EDGE:-$REF_IMAGE_SHORT_EDGE}
+export REF_IMAGE_SHORT_EDGE
+ACTOR_ATTN_BACKEND=${ACTOR_ATTN_BACKEND:-_flash_3_varlen_hub}
+ROLLOUT_ATTN_BACKEND=${ROLLOUT_ATTN_BACKEND:-FLASH_ATTN_3_HUB}
+TOTAL_TRAINING_STEPS=${TOTAL_TRAINING_STEPS:-1000}
+
+script_path=$(readlink -f "$0")
+script_name=$(basename "$script_path" .sh)
+repo_root=$(dirname "$script_path")
+while [[ "$repo_root" != "/" && ! -f "$repo_root/LICENSE" ]]; do
+    repo_root=$(dirname "$repo_root")
+done
+output_dir=${OUTPUT_DIR:-$(dirname "$script_path")/outputs/$script_name}
+checkpoint_dir=$output_dir/checkpoints
+run_timestamp=$(date +"%Y%m%d_%H%M")
+log_file=$output_dir/logs/$run_timestamp/${NODE_RANK:-0}.log
+mkdir -p "$checkpoint_dir" "$(dirname "$log_file")"
+
+lora_warmstart_arg=()
+if [[ -n "${LORA_WARMSTART_PATH:-}" ]]; then
+    lora_warmstart_arg=(actor_rollout_ref.model.lora_adapter_path=$LORA_WARMSTART_PATH)
+fi
+
+python3 -m verl_omni.trainer.main_diffusion_v1 \
+  algorithm.trainer_type=direct_preference \
+  algorithm.sample_source=online \
+  algorithm.adv_mode=continuous \
+  algorithm.timestep_fraction=1.0 \
+  algorithm.old_policy_decay_schedule=delayed_linear_to_0_999 \
+  algorithm.old_policy_update_interval=2 \
+  data.train_files="$DATA_DIR/train.parquet" \
+  data.val_files="$DATA_DIR/test.parquet" \
+  data.train_batch_size=32 \
+  data.val_max_samples=128 \
+  data.max_prompt_length=4096 \
+  data.truncation=error \
+  data.seed=42 \
+  actor_rollout_ref.model.path="$MODEL_PATH/Ref2VA" \
+  actor_rollout_ref.model.tokenizer_path="$MODEL_PATH/Ref2VA/tokenizer" \
+  actor_rollout_ref.model.config_path="$MODEL_PATH/transformer_ref" \
+  +actor_rollout_ref.model.architecture=MiniMaxH3Pipeline \
+  actor_rollout_ref.model.external_lib=verl_omni.pipelines.minimax_h3_diffusion_nft \
+  actor_rollout_ref.model.algorithm=diffusion_nft \
+  actor_rollout_ref.model.model_type=diffusion_nft_model \
+  actor_rollout_ref.model.attn_backend="$ACTOR_ATTN_BACKEND" \
+  actor_rollout_ref.model.enable_gradient_checkpointing=True \
+  actor_rollout_ref.model.lora_rank=64 \
+  actor_rollout_ref.model.lora_alpha=128 \
+  "${lora_warmstart_arg[@]}" \
+  actor_rollout_ref.model.policy_state_adapters='["default","old"]' \
+  actor_rollout_ref.model.target_modules='["to_q","to_k","to_v","to_out.0","ff.net.0.proj","ff.net.2"]' \
+  actor_rollout_ref.model.fsdp_layer_prefixes="['transformer_blocks.','token_refiner.refiner_blocks.']" \
+  actor_rollout_ref.actor.strategy=fsdp2 \
+  actor_rollout_ref.actor.optim.lr=3e-4 \
+  actor_rollout_ref.actor.optim.weight_decay=1e-4 \
+  actor_rollout_ref.actor.optim.betas="[0.9,0.999]" \
+  actor_rollout_ref.actor.optim.override_optimizer_config="{eps: 1e-8}" \
+  actor_rollout_ref.actor.optim.clip_grad=1.0 \
+  actor_rollout_ref.actor.ppo_mini_batch_size=16 \
+  actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=16 \
+  actor_rollout_ref.actor.diffusion_loss.loss_mode=diffusion_nft \
+  actor_rollout_ref.actor.diffusion_loss.clip_ratio=1e-5 \
+  actor_rollout_ref.actor.diffusion_loss.mix_beta=0.1 \
+  actor_rollout_ref.actor.diffusion_loss.ref_kl_coef=0.0001 \
+  actor_rollout_ref.actor.diffusion_loss.adv_clip_max=5.0 \
+  actor_rollout_ref.actor.use_kl_loss=False \
+  actor_rollout_ref.actor.fsdp_config.model_dtype=bfloat16 \
+  actor_rollout_ref.actor.fsdp_config.param_offload=True \
+  actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+  actor_rollout_ref.actor.fsdp_config.ulysses_sequence_parallel_size=1 \
+  actor_rollout_ref.rollout.name=vllm_omni \
+  actor_rollout_ref.rollout.max_num_seqs=1 \
+  actor_rollout_ref.rollout.rollout_attn_backend="$ROLLOUT_ATTN_BACKEND" \
+  actor_rollout_ref.rollout.rollout_adapter=old \
+  actor_rollout_ref.rollout.tensor_model_parallel_size="$ROLLOUT_TP" \
+  +actor_rollout_ref.rollout.engine_kwargs.vllm_omni.text_encoder_tp_size="$TEXT_ENCODER_TP" \
+  actor_rollout_ref.rollout.n="$ROLLOUT_N" \
+  actor_rollout_ref.rollout.seed=42 \
+  actor_rollout_ref.rollout.agent.num_workers=$((N_GPUS / ROLLOUT_TP)) \
+  actor_rollout_ref.rollout.agent.default_agent_loop=minimax_h3_diffusion_single_turn_agent \
+  actor_rollout_ref.rollout.load_format=safetensors \
+  actor_rollout_ref.rollout.layered_summon=True \
+  actor_rollout_ref.rollout.calculate_log_probs=False \
+  actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \
+  +actor_rollout_ref.rollout.engine_kwargs.vllm_omni.enable_layerwise_offload=True \
+  actor_rollout_ref.rollout.max_prompt_embed_length="$MAX_PROMPT_EMBEDS" \
+  actor_rollout_ref.rollout.pipeline.task=ref2va \
+  actor_rollout_ref.rollout.pipeline.height="$HEIGHT" \
+  actor_rollout_ref.rollout.pipeline.width="$WIDTH" \
+  actor_rollout_ref.rollout.pipeline.num_frames="$NUM_FRAMES" \
+  actor_rollout_ref.rollout.pipeline.frame_rate=24.0 \
+  actor_rollout_ref.rollout.pipeline.num_inference_steps="$INFER_STEPS" \
+  actor_rollout_ref.rollout.pipeline.true_cfg_scale=1.0 \
+  actor_rollout_ref.rollout.pipeline.max_sequence_length="$MAX_PROMPT_EMBEDS" \
+  actor_rollout_ref.rollout.pipeline.reference_image_short_edge="$REF_IMAGE_SHORT_EDGE" \
+  actor_rollout_ref.rollout.pipeline.video_flow_shift=12.0 \
+  +actor_rollout_ref.rollout.pipeline.output_type=pt \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.height="$VAL_HEIGHT" \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.width="$VAL_WIDTH" \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.num_frames="$NUM_FRAMES" \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.frame_rate=24.0 \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.num_inference_steps=40 \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.true_cfg_scale=1.0 \
+  actor_rollout_ref.rollout.val_kwargs.pipeline.reference_image_short_edge="$VAL_REF_IMAGE_SHORT_EDGE" \
+  +actor_rollout_ref.rollout.val_kwargs.pipeline.output_type=pt \
+  actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \
+  reward.reward_model.enable=False \
+  reward.num_workers=1 \
+  reward.custom_reward_function.path=pkg://verl_omni.reward_loop.reward_manager.multi \
+  reward.custom_reward_function.name=_multi_reward_placeholder \
+  reward.reward_manager.name=MultiVisualRewardManager \
+  reward.reward_manager.module.path=pkg://verl_omni.reward_loop.reward_manager \
+  "+reward.reward_functions.clap.path=$repo_root/verl_omni/utils/reward_score/clap.py" \
+  '+reward.reward_functions.clap.name=compute_score' \
+  '+reward.reward_functions.clap.weight=1.0' \
+  '+reward.reward_functions.clap.device=cuda:0' \
+  '+reward.reward_functions.clap.model_name_or_path=laion/larger_clap_general' \
+  "+reward.reward_functions.imagebind.path=$repo_root/verl_omni/utils/reward_score/imagebind.py" \
+  '+reward.reward_functions.imagebind.name=compute_score' \
+  '+reward.reward_functions.imagebind.weight=1.0' \
+  '+reward.reward_functions.imagebind.device=cuda:1' \
+  '+reward.reward_functions.imagebind.model_name_or_path=.checkpoints/imagebind_huge.pth' \
+  '+reward.reward_functions.imagebind.mode=audio_video' \
+  reward.aggregation=weighted_sum \
+  trainer.logger='["console","tensorboard","wandb"]' \
+  trainer.project_name=diffusion_nft \
+  trainer.experiment_name=minimax_h3_ref2va_lora_v1 \
+  trainer.default_local_dir="$checkpoint_dir" \
+  trainer.validation_data_dir="$output_dir/validation_data" \
+  trainer.rollout_data_dir="$output_dir/rollout_data" \
+  trainer.rollout_data_save_freq=10 \
+  trainer.rollout_data_max_samples=8 \
+  trainer.log_val_generations=8 \
+  trainer.video_fps=24 \
+  trainer.val_before_train=True \
+  trainer.n_gpus_per_node="$N_GPUS" \
+  trainer.nnodes=1 \
+  trainer.save_freq=5 \
+  trainer.test_freq=10 \
+  trainer.total_epochs=15 \
+  trainer.total_training_steps="$TOTAL_TRAINING_STEPS" \
+  trainer.use_v1=true \
+  trainer.v1.trainer_mode=sync \
+  "$@"
diff --git a/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_t2va_lora_v1.sh b/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_t2va_lora_v1.sh
new file mode 100644
index 000000000..b41bcd071
--- /dev/null
+++ b/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_t2va_lora_v1.sh
@@ -0,0 +1,181 @@
+#!/usr/bin/env bash
+# MiniMax-H3 text-to-audio-video (t2va) DiffusionNFT LoRA recipe
+# (V1 trainer: TransferQueue + ReplayBuffer + sync mode).
+#
+# This is the v1 counterpart of run_minimax_h3_t2va_lora.sh. It uses
+# `verl_omni.trainer.main_diffusion_v1`, which selects
+# `PolicyGradientDiffusionTrainerV1Sync` via `trainer.v1.trainer_mode=sync`.
+# TransferQueue is force-enabled inside the runner, so it does not need to be
+# set on the CLI. GPU layout stays colocated (same as v0).
+#
+# Reference (legacy v0 script):
+# verl-omni/examples/diffusionnft_trainer/minimax_h3/run_minimax_h3_t2va_lora.sh
+set -x
+
+export WANDB_MODE=${WANDB_MODE:-offline}
+
+WORKSPACE=${WORKSPACE:-$HOME}
+MODEL_PATH=${MODEL_PATH:-}
+DATA_DIR=${DATA_DIR:-$WORKSPACE/data/vid_prompt/verl_omni}
+NUM_GPUS=${NUM_GPUS:-8}
+ROLLOUT_TP=${ROLLOUT_TP:-2}
+ROLLOUT_N=${ROLLOUT_N:-16}
+TOTAL_TRAINING_STEPS=${TOTAL_TRAINING_STEPS:-1000}
+HEIGHT=${HEIGHT:-256}
+WIDTH=${WIDTH:-384}
+NUM_FRAMES=${NUM_FRAMES:-121}
+INFER_STEPS=${INFER_STEPS:-10}
+VAL_HEIGHT=${VAL_HEIGHT:-512}
+VAL_WIDTH=${VAL_WIDTH:-768}
+ACTOR_ATTN_BACKEND=${ACTOR_ATTN_BACKEND:-_flash_3_varlen_hub}
+ROLLOUT_ATTN_BACKEND=${ROLLOUT_ATTN_BACKEND:-FLASH_ATTN_3_HUB}
+
+if [[ -z "$MODEL_PATH" || ! -d "$MODEL_PATH/FL2VA" || ! -d "$MODEL_PATH/transformer" ]]; then
+    echo "MODEL_PATH must point to a local MiniMax-H3 repo root containing FL2VA/ and transformer/ (got: '${MODEL_PATH:-<unset>}')" >&2
+    exit 1
+fi
+
+train_path=$DATA_DIR/train.parquet
+test_path=$DATA_DIR/test.parquet
+
+script_path=$(readlink -f "$0")
+script_name=$(basename "$script_path" .sh)
+repo_root=$(dirname "$script_path")
+while [[ "$repo_root" != "/" && ! -f "$repo_root/LICENSE" ]]; do
+    repo_root=$(dirname "$repo_root")
+done
+if [[ ! -f "$repo_root/LICENSE" ]]; then
+    echo "Unable to locate repo root from $script_path: no LICENSE found" >&2
+    exit 1
+fi
+
+output_dir=${OUTPUT_DIR:-$repo_root/outputs/$script_name}
+checkpoint_dir=$output_dir/checkpoints
+run_timestamp=$(date +"%Y%m%d_%H%M")
+log_file=$output_dir/logs/$run_timestamp/${NODE_RANK:-0}.log
+mkdir -p "$checkpoint_dir" "$(dirname "$log_file")"
+exec > >(tee -a "$log_file") 2>&1
+
+export RAY_ACCEL_ENV_VAR_OVERRIDE_ON_ZERO=0
+
+h3_lora_targets="['to_q','to_k','to_v','to_out.0','ff.net.0.proj','ff.net.2']"
+
+lora_warmstart_arg=()
+if [[ -n "${LORA_WARMSTART_PATH:-}" ]]; then
+    lora_warmstart_arg=(actor_rollout_ref.model.lora_adapter_path=$LORA_WARMSTART_PATH)
+fi
+
+python3 -m verl_omni.trainer.main_diffusion_v1 \
+    data.train_files=$train_path \
+    data.val_files=$test_path \
+    data.train_batch_size=32 \
+    data.val_max_samples=128 \
+    data.max_prompt_length=1024 \
+    data.truncation=error \
+    data.seed=42 \
+    algorithm.trainer_type=direct_preference \
+    algorithm.sample_source=online \
+    algorithm.timestep_fraction=1.0 \
+    algorithm.old_policy_decay_schedule=delayed_linear_to_0_999 \
+    algorithm.old_policy_update_interval=2 \
+    algorithm.adv_mode=continuous \
+    actor_rollout_ref.model.path=$MODEL_PATH/FL2VA \
+    actor_rollout_ref.model.config_path=$MODEL_PATH/transformer \
+    +actor_rollout_ref.model.architecture=MiniMaxH3Pipeline \
+    actor_rollout_ref.model.algorithm=diffusion_nft \
+    actor_rollout_ref.model.model_type=diffusion_nft_model \
+    actor_rollout_ref.model.attn_backend=$ACTOR_ATTN_BACKEND \
+    actor_rollout_ref.model.enable_gradient_checkpointing=True \
+    actor_rollout_ref.model.lora_rank=64 \
+    actor_rollout_ref.model.lora_alpha=128 \
+    "${lora_warmstart_arg[@]}" \
+    actor_rollout_ref.model.policy_state_adapters='["default","old"]' \
+    actor_rollout_ref.model.target_modules="$h3_lora_targets" \
+    actor_rollout_ref.model.fsdp_layer_prefixes="['transformer_blocks.','token_refiner.refiner_blocks.']" \
+    '+actor_rollout_ref.actor.fsdp_config.wrap_policy.transformer_layer_cls_to_wrap=[MiniMaxH3TransformerBlock,MiniMaxH3TokenRefinerBlock]' \
+    actor_rollout_ref.actor.strategy=fsdp2 \
+    actor_rollout_ref.actor.optim.lr=3e-4 \
+    actor_rollout_ref.actor.optim.weight_decay=1e-4 \
+    actor_rollout_ref.actor.optim.betas="[0.9,0.999]" \
+    actor_rollout_ref.actor.optim.override_optimizer_config="{eps: 1e-8}" \
+    actor_rollout_ref.actor.optim.clip_grad=1.0 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=16 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.actor.diffusion_loss.loss_mode=diffusion_nft \
+    actor_rollout_ref.actor.diffusion_loss.clip_ratio=1e-5 \
+    actor_rollout_ref.actor.diffusion_loss.mix_beta=0.1 \
+    actor_rollout_ref.actor.diffusion_loss.ref_kl_coef=0.0001 \
+    actor_rollout_ref.actor.diffusion_loss.adv_clip_max=5.0 \
+    actor_rollout_ref.actor.use_kl_loss=False \
+    actor_rollout_ref.actor.fsdp_config.model_dtype=bfloat16 \
+    actor_rollout_ref.actor.fsdp_config.param_offload=True \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+    actor_rollout_ref.actor.fsdp_config.ulysses_sequence_parallel_size=1 \
+    actor_rollout_ref.rollout.name=vllm_omni \
+    actor_rollout_ref.rollout.max_num_seqs=1 \
+    actor_rollout_ref.rollout.rollout_attn_backend=$ROLLOUT_ATTN_BACKEND \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=$ROLLOUT_TP \
+    actor_rollout_ref.rollout.n=$ROLLOUT_N \
+    actor_rollout_ref.rollout.seed=42 \
+    actor_rollout_ref.rollout.agent.num_workers=$((NUM_GPUS / ROLLOUT_TP)) \
+    actor_rollout_ref.rollout.agent.default_agent_loop=minimax_h3_diffusion_single_turn_agent \
+    actor_rollout_ref.rollout.load_format=safetensors \
+    actor_rollout_ref.rollout.layered_summon=True \
+    actor_rollout_ref.rollout.calculate_log_probs=False \
+    actor_rollout_ref.rollout.rollout_adapter=old \
+    actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=16 \
+    actor_rollout_ref.rollout.pipeline.aspect_ratio=16:9 \
+    actor_rollout_ref.rollout.pipeline.height=$HEIGHT \
+    actor_rollout_ref.rollout.pipeline.width=$WIDTH \
+    actor_rollout_ref.rollout.pipeline.num_frames=$NUM_FRAMES \
+    actor_rollout_ref.rollout.pipeline.frame_rate=24.0 \
+    actor_rollout_ref.rollout.pipeline.num_inference_steps=$INFER_STEPS \
+    actor_rollout_ref.rollout.pipeline.true_cfg_scale=1.0 \
+    actor_rollout_ref.rollout.pipeline.max_sequence_length=1024 \
+    actor_rollout_ref.rollout.pipeline.video_flow_shift=12.0 \
+    +actor_rollout_ref.rollout.pipeline.output_type=pt \
+    actor_rollout_ref.rollout.val_kwargs.pipeline.height=$VAL_HEIGHT \
+    actor_rollout_ref.rollout.val_kwargs.pipeline.width=$VAL_WIDTH \
+    actor_rollout_ref.rollout.val_kwargs.pipeline.num_frames=$NUM_FRAMES \
+    actor_rollout_ref.rollout.val_kwargs.pipeline.frame_rate=24.0 \
+    actor_rollout_ref.rollout.val_kwargs.pipeline.num_inference_steps=40 \
+    actor_rollout_ref.rollout.val_kwargs.pipeline.true_cfg_scale=1.0 \
+    +actor_rollout_ref.rollout.val_kwargs.pipeline.output_type=pt \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=16 \
+    reward.num_workers=1 \
+    reward.reward_model.enable=False \
+    reward.custom_reward_function.path=pkg://verl_omni.reward_loop.reward_manager.multi \
+    reward.custom_reward_function.name=_multi_reward_placeholder \
+    reward.reward_manager.name=MultiVisualRewardManager \
+    reward.reward_manager.module.path=pkg://verl_omni.reward_loop.reward_manager \
+    "+reward.reward_functions.clap.path=$repo_root/verl_omni/utils/reward_score/clap.py" \
+    '+reward.reward_functions.clap.name=compute_score' \
+    '+reward.reward_functions.clap.weight=1.0' \
+    '+reward.reward_functions.clap.device=cuda:0' \
+    '+reward.reward_functions.clap.model_name_or_path=laion/larger_clap_general' \
+    "+reward.reward_functions.imagebind.path=$repo_root/verl_omni/utils/reward_score/imagebind.py" \
+    '+reward.reward_functions.imagebind.name=compute_score' \
+    '+reward.reward_functions.imagebind.weight=1.0' \
+    '+reward.reward_functions.imagebind.device=cuda:1' \
+    '+reward.reward_functions.imagebind.model_name_or_path=.checkpoints/imagebind_huge.pth' \
+    '+reward.reward_functions.imagebind.mode=audio_video' \
+    reward.aggregation=weighted_sum \
+    trainer.logger='["console","tensorboard","wandb"]' \
+    trainer.project_name=diffusion_nft \
+    trainer.experiment_name=minimax_h3_t2va_lora_v1 \
+    trainer.default_local_dir=$checkpoint_dir \
+    trainer.validation_data_dir=$output_dir/validation_data \
+    trainer.rollout_data_dir=$output_dir/rollout_data \
+    trainer.rollout_data_save_freq=10 \
+    trainer.log_val_generations=8 \
+    trainer.video_fps=24 \
+    trainer.val_before_train=True \
+    trainer.n_gpus_per_node=$NUM_GPUS \
+    trainer.nnodes=1 \
+    trainer.save_freq=5 \
+    trainer.max_actor_ckpt_to_keep=5 \
+    trainer.test_freq=10 \
+    trainer.total_epochs=15 \
+    trainer.total_training_steps=$TOTAL_TRAINING_STEPS \
+    trainer.use_v1=true \
+    trainer.v1.trainer_mode=sync "$@"
diff --git a/examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_npu_v1.sh b/examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_npu_v1.sh
new file mode 100644
index 000000000..f358dd244
--- /dev/null
+++ b/examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_npu_v1.sh
@@ -0,0 +1,103 @@
+#!/usr/bin/env bash
+# Qwen-Image DiffusionNFT LoRA RL on Ascend NPU (V1 trainer: TransferQueue + ReplayBuffer + sync mode).
+#
+# This is the v1 counterpart of run_qwen_image_ocr_lora_npu.sh. It uses
+# `verl_omni.trainer.main_diffusion_v1`, which selects
+# `PolicyGradientDiffusionTrainerV1Sync` via `trainer.v1.trainer_mode=sync`.
+# TransferQueue is force-enabled inside the runner, so it does not need to be
+# set on the CLI. Attention / TP layout stays on the v0 NPU script.
+#
+# Reference (legacy v0 script):
+# verl-omni/examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_npu.sh
+set -x
+export VERL_DATAPROTO_SERIALIZATION_METHOD=numpy
+
+# Set WORKSPACE to any writable directory; defaults to $HOME
+WORKSPACE=${WORKSPACE:-$HOME}
+
+ocr_train_path=$WORKSPACE/data/ocr/train.parquet
+ocr_test_path=$WORKSPACE/data/ocr/test.parquet
+
+model_name=Qwen/Qwen-Image
+reward_model_name=Qwen/Qwen3-VL-8B-Instruct
+reward_function_path=verl_omni/utils/reward_score/genrm_ocr.py
+
+NUM_GPUS_ACTOR_ROLLOUT_REWARD=16
+ROLLOUT_TP=2
+REWARD_TP=4
+IMAGE_RESOLUTION=512
+
+ENGINE=vllm_omni
+REWARD_ENGINE=vllm
+
+
+python3 -m verl_omni.trainer.main_diffusion_v1 \
+    data.train_files="$ocr_train_path" \
+    data.val_files="$ocr_test_path" \
+    data.train_max_samples=7200 \
+    data.train_batch_size=24 \
+    data.max_prompt_length=256 \
+    actor_rollout_ref.model.attn_backend='_native_npu' \
+    actor_rollout_ref.model.algorithm=diffusion_nft \
+    actor_rollout_ref.model.model_type=diffusion_nft_model \
+    actor_rollout_ref.model.path=$model_name \
+    actor_rollout_ref.model.lora_rank=64 \
+    actor_rollout_ref.model.lora_alpha=128 \
+    actor_rollout_ref.model.policy_state_adapters='["default","old"]' \
+    actor_rollout_ref.model.target_modules="['to_q','to_k','to_v','to_out.0','add_q_proj','add_k_proj','add_v_proj','to_add_out','img_mlp.net.0.proj','img_mlp.net.2','txt_mlp.net.0.proj','txt_mlp.net.2']" \
+    actor_rollout_ref.actor.optim.lr=3e-4 \
+    actor_rollout_ref.actor.optim.weight_decay=0.0001 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=12 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=12 \
+    actor_rollout_ref.actor.diffusion_loss.loss_mode=diffusion_nft \
+    actor_rollout_ref.actor.diffusion_loss.clip_ratio=1e-5 \
+    actor_rollout_ref.actor.diffusion_loss.mix_beta=0.1 \
+    actor_rollout_ref.actor.diffusion_loss.ref_kl_coef=0.0001 \
+    actor_rollout_ref.actor.diffusion_loss.adv_clip_max=5.0 \
+    actor_rollout_ref.actor.fsdp_config.param_offload=True \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+    actor_rollout_ref.actor.fsdp_config.model_dtype=bfloat16 \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=$ROLLOUT_TP \
+    actor_rollout_ref.rollout.rollout_attn_backend=TORCH_SDPA \
+    actor_rollout_ref.rollout.name=$ENGINE \
+    actor_rollout_ref.rollout.n=16 \
+    actor_rollout_ref.rollout.agent.num_workers=$((NUM_GPUS_ACTOR_ROLLOUT_REWARD / ROLLOUT_TP)) \
+    actor_rollout_ref.rollout.load_format=safetensors \
+    actor_rollout_ref.rollout.layered_summon=True \
+    actor_rollout_ref.rollout.calculate_log_probs=False \
+    actor_rollout_ref.rollout.rollout_adapter=old \
+    actor_rollout_ref.rollout.pipeline.num_inference_steps=10 \
+    actor_rollout_ref.rollout.pipeline.true_cfg_scale=1.0 \
+    actor_rollout_ref.rollout.pipeline.height=$IMAGE_RESOLUTION \
+    actor_rollout_ref.rollout.pipeline.width=$IMAGE_RESOLUTION \
+    actor_rollout_ref.rollout.pipeline.max_sequence_length=256 \
+    actor_rollout_ref.rollout.val_kwargs.pipeline.num_inference_steps=40 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \
+    algorithm.trainer_type=direct_preference \
+    algorithm.sample_source=online \
+    algorithm.timestep_fraction=1.0 \
+    algorithm.old_policy_decay_schedule=delayed_linear_to_0_999 \
+    algorithm.old_policy_update_interval=2 \
+    algorithm.adv_mode=continuous \
+    reward.num_workers=$((NUM_GPUS_ACTOR_ROLLOUT_REWARD / REWARD_TP)) \
+    reward.reward_model.enable=True \
+    reward.reward_model.model_path=$reward_model_name \
+    reward.reward_model.rollout.enforce_eager=False \
+    reward.reward_model.rollout.name=$REWARD_ENGINE \
+    reward.reward_model.rollout.tensor_model_parallel_size=$REWARD_TP \
+    reward.custom_reward_function.path=$reward_function_path \
+    reward.custom_reward_function.name=compute_score_ocr \
+    trainer.logger='["console", "tensorboard"]' \
+    trainer.project_name=diffusion_nft \
+    trainer.experiment_name=qwen_image_ocr_lora_npu_v1 \
+    trainer.log_val_generations=8 \
+    trainer.val_before_train=False \
+    trainer.n_gpus_per_node=$NUM_GPUS_ACTOR_ROLLOUT_REWARD \
+    trainer.nnodes=1 \
+    trainer.save_freq=60 \
+    trainer.test_freq=20 \
+    trainer.total_epochs=1 \
+    trainer.total_training_steps=300 \
+    trainer.use_v1=true \
+    trainer.v1.trainer_mode=sync "$@"
+    
diff --git a/examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_v1.sh b/examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_v1.sh
new file mode 100644
index 000000000..cd4567e29
--- /dev/null
+++ b/examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora_v1.sh
@@ -0,0 +1,98 @@
+#!/usr/bin/env bash
+# Qwen-Image DiffusionNFT LoRA RL (V1 trainer: TransferQueue + ReplayBuffer + sync mode).
+#
+# This is the v1 counterpart of run_qwen_image_ocr_lora.sh. It uses
+# `verl_omni.trainer.main_diffusion_v1`, which selects
+# `PolicyGradientDiffusionTrainerV1Sync` via `trainer.v1.trainer_mode=sync`.
+# TransferQueue is force-enabled inside the runner, so it does not need to be
+# set on the CLI.
+#
+# Reference (legacy v0 script):
+# verl-omni/examples/diffusionnft_trainer/qwen_image/run_qwen_image_ocr_lora.sh
+set -x
+
+# Set WORKSPACE to any writable directory; defaults to $HOME
+WORKSPACE=${WORKSPACE:-$HOME}
+
+ocr_train_path=$WORKSPACE/data/ocr/train.parquet
+ocr_test_path=$WORKSPACE/data/ocr/test.parquet
+
+model_name=Qwen/Qwen-Image
+reward_model_name=Qwen/Qwen3-VL-8B-Instruct
+reward_function_path=verl_omni/utils/reward_score/genrm_ocr.py
+
+NUM_GPUS_ACTOR_ROLLOUT_REWARD=4
+ROLLOUT_TP=1
+REWARD_TP=4
+IMAGE_RESOLUTION=512
+
+ENGINE=vllm_omni
+REWARD_ENGINE=vllm
+
+
+python3 -m verl_omni.trainer.main_diffusion_v1 \
+    data.train_files=$ocr_train_path \
+    data.val_files=$ocr_test_path \
+    data.train_max_samples=7200 \
+    data.train_batch_size=24 \
+    data.max_prompt_length=256 \
+    actor_rollout_ref.model.algorithm=diffusion_nft \
+    actor_rollout_ref.model.model_type=diffusion_nft_model \
+    actor_rollout_ref.model.path=$model_name \
+    actor_rollout_ref.model.lora_rank=64 \
+    actor_rollout_ref.model.lora_alpha=128 \
+    actor_rollout_ref.model.policy_state_adapters='["default","old"]' \
+    actor_rollout_ref.model.target_modules="['to_q','to_k','to_v','to_out.0','add_q_proj','add_k_proj','add_v_proj','to_add_out','img_mlp.net.0.proj','img_mlp.net.2','txt_mlp.net.0.proj','txt_mlp.net.2']" \
+    actor_rollout_ref.actor.optim.lr=3e-4 \
+    actor_rollout_ref.actor.optim.weight_decay=0.0001 \
+    actor_rollout_ref.actor.ppo_mini_batch_size=12 \
+    actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=12 \
+    actor_rollout_ref.actor.diffusion_loss.loss_mode=diffusion_nft \
+    actor_rollout_ref.actor.diffusion_loss.clip_ratio=1e-5 \
+    actor_rollout_ref.actor.diffusion_loss.mix_beta=0.1 \
+    actor_rollout_ref.actor.diffusion_loss.ref_kl_coef=0.0001 \
+    actor_rollout_ref.actor.diffusion_loss.adv_clip_max=5.0 \
+    actor_rollout_ref.actor.fsdp_config.param_offload=True \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+    actor_rollout_ref.actor.fsdp_config.model_dtype=bfloat16 \
+    actor_rollout_ref.rollout.tensor_model_parallel_size=$ROLLOUT_TP \
+    actor_rollout_ref.rollout.name=$ENGINE \
+    actor_rollout_ref.rollout.n=16 \
+    actor_rollout_ref.rollout.agent.num_workers=$((NUM_GPUS_ACTOR_ROLLOUT_REWARD / ROLLOUT_TP)) \
+    actor_rollout_ref.rollout.load_format=safetensors \
+    actor_rollout_ref.rollout.layered_summon=True \
+    actor_rollout_ref.rollout.calculate_log_probs=False \
+    actor_rollout_ref.rollout.rollout_adapter=old \
+    actor_rollout_ref.rollout.pipeline.num_inference_steps=10 \
+    actor_rollout_ref.rollout.pipeline.true_cfg_scale=1.0 \
+    actor_rollout_ref.rollout.pipeline.height=$IMAGE_RESOLUTION \
+    actor_rollout_ref.rollout.pipeline.width=$IMAGE_RESOLUTION \
+    actor_rollout_ref.rollout.pipeline.max_sequence_length=256 \
+    actor_rollout_ref.rollout.val_kwargs.pipeline.num_inference_steps=40 \
+    actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \
+    algorithm.trainer_type=direct_preference \
+    algorithm.sample_source=online \
+    algorithm.timestep_fraction=1.0 \
+    algorithm.old_policy_decay_schedule=delayed_linear_to_0_999 \
+    algorithm.old_policy_update_interval=2 \
+    algorithm.adv_mode=continuous \
+    reward.num_workers=$((NUM_GPUS_ACTOR_ROLLOUT_REWARD / REWARD_TP)) \
+    reward.reward_model.enable=True \
+    reward.reward_model.model_path=$reward_model_name \
+    reward.reward_model.rollout.name=$REWARD_ENGINE \
+    reward.reward_model.rollout.tensor_model_parallel_size=$REWARD_TP \
+    reward.custom_reward_function.path=$reward_function_path \
+    reward.custom_reward_function.name=compute_score_ocr \
+    trainer.logger='["console", "wandb"]' \
+    trainer.project_name=diffusion_nft \
+    trainer.experiment_name=qwen_image_ocr_lora_v1 \
+    trainer.log_val_generations=8 \
+    trainer.val_before_train=False \
+    trainer.n_gpus_per_node=$NUM_GPUS_ACTOR_ROLLOUT_REWARD \
+    trainer.nnodes=1 \
+    trainer.save_freq=60 \
+    trainer.test_freq=20 \
+    trainer.total_epochs=1 \
+    trainer.total_training_steps=300 \
+    trainer.use_v1=true \
+    trainer.v1.trainer_mode=sync "$@"