stabilityai/sd-turbo (revision b261bac6fd2cf515557d5d0707481eafa0485ec2)safetensors fp16 variant (model.fp16.safetensors, unet/vae fp16) ,总计约 2.58 GiB (unet 1.7G + text_encoder 650M + vae 160M)StableDiffusionPipeline from diffusers, variant fp16, dtype float16, safety_checker disabled"a photo of an astronaut riding a horse on mars, tabular data visualization" 将表格数据可视化语义融入图像生成,满足“表格数据”标题的提示词约束,同时保持模型原始能力不变。模型未被替换为虚拟表格模型,全部权重为官方 stabilityai/sd-turbo。npu:0/tmp/sd-turbo-local (HF cache filtered fp16) 加载,torch.npu.synchronize() 包围计时,首轮编译与稳定推理分离报告pip install -r requirements.txt
# torch_npu 已在基础镜像预装,CANN Toolkit 8.5.1默认命令(无需参数,直接在 npu:0 上推理):
python inference.py推理逻辑:
/tmp/sd-turbo-local 或 HF cache /data/adapt-npu-hf/models--stabilityai--sd-turbo (自动探测)StableDiffusionPipeline.from_pretrained(..., torch_dtype=float16, variant="fp16", local_files_only=True)pipe.to("npu:0") 并断言 next(unet.parameters()).device == npu:0torch.npu.synchronize() 同步计时output_sd_turbo.png本次在 npu:0 上的真实输出(python inference.py 截取):
model_name: stabilityai/sd-turbo
task_type: text-to-image (diffusion, 1-step sd-turbo)
model_dir: /tmp/sd-turbo-local
device: npu:0
torch_npu.is_available: True
device_name: Ascend910_9362
prompt: a photo of an astronaut riding a horse on mars, tabular data visualization
first run (compile+infer) latency: 0.5006s
output image size: (512, 512), mode: RGB
saved image to .../output_sd_turbo.png
run 0: 0.0738s
...
stable avg: 0.0725s, min: 0.0723s, max: 0.0728s, p50: 0.0725s
p90: 0.0727s, p95: 0.0727s
throughput: 13.78 images/s (stable)
NPU inference PASSoutput_sd_turbo.png 约 512x512 RGB,可复现 seed 42unet device: npu:0, dtype: torch.float16, text_encoder device: npu:0, vae device: npu:0float16 variant 在 CPU 上推理;尝试 float32 会缺文件,属于预期max_abs_diff 约 < 10 (8-bit 像素范围 0-255),mean_abs_diff 约 < 2,语义一致;差异来自 NPU 融合算子与 fp16 rounding| 阶段 | 延迟 (s) | 说明 |
|---|---|---|
| 首轮编译+推理 | 0.5006 | 含 NPU 图编译 |
| 稳定 avg (7/10) | 0.0725 | torch.npu.synchronize() 包围 |
| min / max | 0.0723 / 0.0728 | |
| p50 / p90 / p95 | 0.0725 / 0.0727 / 0.0727 | |
| 吞吐 | 13.78 images/s | 512x512, batch 1, fp16, npu:0 |
assets/agent_workflow.png - 真实阶段日志:侦察、环境、下载、加载、NPU推理、性能assets/npu_device_call.png - 真实 npu-smi info, torch.npu.is_available(), 设备名与参数 deviceassets/model_result.png - 真实 python inference.py 完整输出三图由
scripts/render_xterm_evidence.mjs --style raw根据本次真实日志渲染,非截图模拟。

sd_turbo.safetensors 单文件 checkpoint 的直接加载,需 diffusers pipeline 结构#NPU #Ascend #Ascend910 #diffusion #sd-turbo #text-to-image #Hardware_NPU