2toINF/X-VLA-Libero(X-VLA 0.9B Libero Edition)的端到端昇腾 NPU 推理候选。
在物理 npu:2(Ascend910B4-1)上运行真实 checkpoint 前向,输出完整的 30 步
ee6d 动作轨迹。全程不发布、不启动仿真/机器人/服务端。
模型 ID:2toINF/X-VLA-Libero
上游来源:https://huggingface.co/2toINF/X-VLA-Libero
不可变修订:129e71460678b7236cee6fc9707f09d9fa0c3590
许可证:Apache-2.0(上游模型卡声明,见源码文件头)
架构(原生 X-VLA LIBERO 严格 checkpoint 路线):
Florence2ForConditionalGeneration,DaViT 视觉塔 + BART-like 语言编码器)ee6d 动作空间(dim_action=20,gripper_idx=(9,19),末端夹爪后处理 sigmoid)关键配置:num_actions=30、num_domains=30、len_soft_prompts=32、use_proprio=True、depth=24。
任务类型:vision-language-action(pipeline_tag: robotics,library_name: transformers)。
权重完整性(严格加载):
| 项目 | 值 |
|---|---|
| 权重文件 | model.safetensors |
| 字节数 | 3,519,068,172 |
| SHA256 | 260cc58869125b826e93bcaa60bca3ea37bcc7aaf893d368a991ca14fde9f0c8 |
| 严格加载张量 | 903(load_state_dict(strict=True) → missing=[] unexpected=[]) |
| 参数量 | 879,482,456 |
| 绑定嵌入别名 | vlm.language_model.model.encoder.embed_tokens.weight ← vlm.language_model.model.shared.weight |
本候选在已有 CANN 容器内运行,torch 2.9.0+cpu 与 torch_npu 2.9.0 已匹配安装,
禁止安装或替换为普通 PyPI wheel。
实测版本:
| 组件 | 版本 |
|---|---|
| Python | 3.11.14 |
| torch | 2.9.0+cpu(容器自带,不重装) |
| torch_npu | 2.9.0(容器自带,不重装) |
| transformers | 5.15.1 |
| safetensors | 0.8.0 |
| numpy | 1.26.4 |
| tokenizers | 0.22.2 |
| Pillow | 12.3.0 |
| opencv-python-headless | 4.9.0.80 |
| fastapi / uvicorn / json-numpy | 0.141.1 / 0.52.4 / 2.1.1 |
| CANN / npu-smi | 25.5.1 |
| 硬件 | 8× Ascend910B4-1,本候选使用物理 npu:2 |
仅当缺失非 torch 依赖时,使用 Aliyun pip 镜像安装:
pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/注意:
requirements.txt中torch==2.9.0+cpu/torch-npu==2.9.0以注释形式记录, 绝不通过pip重装,避免破坏 CANN 匹配的框架对。
确认 CANN 环境与 NPU 可见:
npu-smi info
python3 -c "import torch, torch_npu; print(torch.npu.device_count(), torch.npu.get_device_name(2))"进入交付目录:
cd /work/pipeline/jobs/rl26/x-vla-libero/x-vla-libero-npu校验权重(可选,流式 SHA256):
sha256sum /work/pipeline/models/rl26/x-vla-libero/model.safetensors
# 期望:260cc58869125b826e93bcaa60bca3ea37bcc7aaf893d368a991ca14fde9f0c8运行一次真实 NPU 推理(前向完全在 npu:2,失败即非零退出):
python3 inference.py > logs/npu_inference.log 2>&1
echo "exit=$?"查看结构化结果与原始日志:
cat logs/npu_result.json
cat logs/npu_smi_during_hold.txt # 持窗内 npu-smi 的 PID 关联证据脚本行为:
npu:2。[1, num_actions=30, dim_action=20] 动作轨迹。npu-smi info 关联推理 Python PID 在物理 npu:2 上的显存占用。确定性测试输入(可再分发、合成):
pick up the black bowl and place it on the plate(BART tokenizer,[1, 50])[1, 3, 3, 224, 224])[1, 20])0(来自固定 LIBERO 处理管线 XVLAAddDomainIdProcessorStep 默认值)运行命令:
cd /work/pipeline/jobs/rl26/x-vla-libero/x-vla-libero-npu
python3 inference.py实际输出(真实运行,确定性):
[2026-08-22T11:10:56] npu device_count=8
[2026-08-22T11:10:56] npu:2 = Ascend910B4-1
[2026-08-22T11:10:58] contract: model_id=2toINF/X-VLA-Libero engine=torch_npu npu_available=True npu_device_name=Ascend910B4-1
[2026-08-22T11:11:01] artifact bytes=3519068172 sha256=260cc58869125b826e93bcaa60bca3ea37bcc7aaf893d368a991ca14fde9f0c8 (verified in 2.9s)
[2026-08-22T11:11:35] model constructed in 32.4s, params=879,482,456
[2026-08-22T11:11:38] strict load_state_dict: missing=[] unexpected=[]
[2026-08-22T11:11:38] all model parameters on npu:2: OK
[2026-08-22T11:11:39] input devices: {'input_ids': 'npu:2', 'image_input': 'npu:2', 'image_mask': 'npu:2', 'proprio': 'npu:2', 'domain_id': 'npu:2'}
[2026-08-22T11:11:47] generate_actions elapsed_s=7.78
[2026-08-22T11:11:47] output shape=[1, 30, 20] device=npu:2
[2026-08-22T11:11:47] output stats: min=-0.4503 max=0.5709 mean=0.0160 std=0.2279 absmean=0.1756
[2026-08-22T11:11:47] action trajectory (first row): [-0.0377, -0.1313, 0.1995, 0.1125, -0.1990, 0.1346, 0.0421, -0.0167, -0.4503, 0.4032, -0.0512, 0.1860, -0.1696, -0.0370, -0.1805, -0.2177, 0.2699, -0.1672, -0.0017, 0.5671]
[2026-08-22T11:11:47] action trajectory (last row): [-0.0697, -0.1391, 0.2190, 0.0692, -0.2307, 0.1118, 0.0406, 0.0067, -0.4331, 0.4021, -0.0575, 0.1802, -0.1598, -0.0279, -0.1519, -0.2180, 0.2643, -0.1547, -0.0095, 0.5625]
[2026-08-22T11:11:47] gripper channel[9] every 5 steps (post-sigmoid): 0.4032, 0.4039, 0.4041, 0.4044, 0.4025, 0.4018
[2026-08-22T11:11:47] pid=105072
[2026-08-22T11:12:04] PID CORRELATION on npu:2: pid=105072 name=python3 mem_mb=4294
[2026-08-22T11:12:19] wrote /work/pipeline/jobs/rl26/x-vla-libero/x-vla-libero-npu/logs/npu_result.json (exit_code=0)PID 会随每次运行变化;最新一次运行的真实 PID 关联记录见
logs/npu_smi_during_hold.txt。
logs/npu_result.json 顶层契约(节选):
{
"model_id": "2toINF/X-VLA-Libero",
"engine": "torch_npu",
"exit_code": 0,
"npu_available": true,
"npu_device_name": "Ascend910B4-1",
"source_revision": "129e71460678b7236cee6fc9707f09d9fa0c3590",
"input_device": "npu:2",
"model_device": "npu:2",
"output_device": "npu:2",
"output": {
"task": "vision-language-action flow-matching trajectory",
"shape": [1, 30, 20],
"flow_steps": 10,
"statistics": {"min": -0.450349, "max": 0.570891, "mean": 0.015971, "std": 0.227943, "absmean": 0.175551},
"elapsed_s": 7.779,
"pid": 105072
}
}generate_actions 前向约 7.5–7.8s。npu:2 上的单卡推理路线,未做多卡或性能调优。# 语法
python3 -m py_compile inference.py models/*.py timm_shim/timm/__init__.py timm_shim/timm/layers/__init__.py
# NPU 环境预检(维护的 checker)
python3 /root/ascend-model-pipeline/upstreams/AscendSkills/ai4s-migration-base/scripts/check_npu_basics.py --device npu:2
# 交付校验(维护的 validator)
python3 /root/ascend-model-pipeline/skills/adapting-models-for-ascend-npu/scripts/validate_delivery.py .
# 校验和
sha256sum -c SHA256SUMSlogs/npu_result.jsonlogs/npu_inference.loglogs/npu_smi_during_hold.txtSHA256SUMS