diff --git a/configs/qwen3.5-4b-dta.json b/configs/qwen3.5-4b-dta.json
new file mode 100644
index 000000000..24b7b39d4
--- /dev/null
+++ b/configs/qwen3.5-4b-dta.json
@@ -0,0 +1,49 @@
+{
+    "architectures": [
+      "DFlashDraftModel"
+    ],
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "auto_map": {
+      "AutoModel": "dflash.DFlashDraftModel"
+    },
+    "block_size": 16,
+    "bos_token_id": 248043,
+    "dflash_config": {
+      "mask_token_id": 248070,
+      "target_layer_ids": [1, 8, 15, 22, 29],
+      "training_mode": "vp_drafter",
+      "prefix_weight_base": 0.9
+    },
+    "dtype": "bfloat16",
+    "eos_token_id": 248044,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 2560,
+    "initializer_range": 0.02,
+    "intermediate_size": 9728,
+    "layer_types": [
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention",
+      "full_attention"
+    ],
+    "max_position_embeddings": 262144,
+    "max_window_layers": 5,
+    "model_type": "qwen3",
+    "num_attention_heads": 32,
+    "num_hidden_layers": 5,
+    "num_key_value_heads": 8,
+    "num_target_layers": 32,
+    "pad_token_id": 248044,
+    "rms_norm_eps": 1e-06,
+    "rope_scaling": null,
+    "rope_theta": 10000000,
+    "sliding_window": null,
+    "tie_word_embeddings": true,
+    "transformers_version": "4.57.1",
+    "use_cache": true,
+    "use_sliding_window": false,
+    "vocab_size": 248320
+  }
diff --git a/examples/run_qwen3.5_4b_dta_online_npu.sh b/examples/run_qwen3.5_4b_dta_online_npu.sh
new file mode 100755
index 000000000..a4d8fc6a7
--- /dev/null
+++ b/examples/run_qwen3.5_4b_dta_online_npu.sh
@@ -0,0 +1,48 @@
+#!/bin/bash
+# VP-Drafter (D2SD) training for Qwen3.5-4B on Ascend NPU
+# Backend: HF + SDPA attention + HCCL distributed
+#
+# Required environment variables (override before invoking the script):
+#   TARGET_MODEL_PATH  Path to Qwen3.5-4B weights
+#   TRAIN_DATA_PATH    Path to the training jsonl
+
+set -eu
+
+SCRIPT_DIR=$( cd -- "$( dirname -- "${BASH_SOURCE[0]}" )" &> /dev/null && pwd )
+ROOT_DIR=$(dirname "$SCRIPT_DIR")
+
+: "${TARGET_MODEL_PATH:?Set TARGET_MODEL_PATH to the Qwen3.5-4B weights directory}"
+: "${TRAIN_DATA_PATH:?Set TRAIN_DATA_PATH to the training jsonl file}"
+
+NPU_DEVICES=${1:-0,1,2,3,4,5,6,7}
+NUM_DEVICES=$(echo "$NPU_DEVICES" | tr ',' '\n' | wc -l)
+
+export ASCEND_RT_VISIBLE_DEVICES=$NPU_DEVICES
+export PYTORCH_NPU_ALLOC_CONF=max_split_size_mb:32
+
+torchrun \
+    --standalone \
+    --nproc_per_node "$NUM_DEVICES" \
+    "$ROOT_DIR/scripts/train_dflash.py" \
+    --target-model-path "$TARGET_MODEL_PATH" \
+    --target-model-backend hf \
+    --draft-config-path "$ROOT_DIR/configs/qwen3.5-4b-dta.json" \
+    --train-data-path "$TRAIN_DATA_PATH" \
+    --output-dir "$ROOT_DIR/outputs/qwen3.5-4b-dta-npu" \
+    --num-epochs 6 \
+    --batch-size 1 \
+    --accumulation-steps 4 \
+    --learning-rate 6e-4 \
+    --warmup-ratio 0.04 \
+    --max-grad-norm 1.0 \
+    --max-length 1024 \
+    --chat-template qwen3.5 \
+    --attention-backend sdpa \
+    --num-anchors 16 \
+    --loss-decay-gamma 7.0 \
+    --log-interval 50 \
+    --save-interval 3000 \
+    --report-to tensorboard \
+    --block-size 16 \
+    --embedding-key model.language_model.embed_tokens.weight \
+    --trust-remote-code