{"id":82374,"date":"2026-07-25T08:26:01","date_gmt":"2026-07-25T00:26:01","guid":{"rendered":"https:\/\/www.wsisp.com\/helps\/82374.html"},"modified":"2026-07-25T08:26:01","modified_gmt":"2026-07-25T00:26:01","slug":"flux-1%e9%83%a8%e7%bd%b2%e6%95%99%e7%a8%8b%ef%bc%9aarm%e6%9e%b6%e6%9e%84%e6%9c%8d%e5%8a%a1%e5%99%a8%ef%bc%88%e5%a6%82nvidia-grace%ef%bc%89%e9%80%82%e9%85%8d%e6%b5%b7%e6%99%af%e5%9b%be%e7%94%9f","status":"publish","type":"post","link":"https:\/\/www.wsisp.com\/helps\/82374.html","title":{"rendered":"FLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1"},"content":{"rendered":"

FLUX.1\u90e8\u7f72\u6559\u7a0b：ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982NVIDIA Grace）\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1<\/h2>\n

1. \u524d\u8a00：\u5f53AI\u7ed8\u753b\u9047\u4e0aARM\u65b0\u8d35<\/h3>\n

\u6700\u8fd1，\u6211\u624b\u5934\u62ff\u5230\u4e86\u4e00\u53f0\u642d\u8f7dNVIDIA Grace CPU\u7684ARM\u67b6\u6784\u670d\u52a1\u5668\u3002\u8fd9\u673a\u5668\u6027\u80fd\u5f88\u5f3a，\u4f46\u6709\u4e2a\u95ee\u9898：\u5f88\u591aAI\u5e94\u7528\u90fd\u662f\u4e3ax86\u67b6\u6784\u8bbe\u8ba1\u7684，\u5728ARM\u4e0a\u8dd1\u8d77\u6765\u603b\u6709\u70b9\u6c34\u571f\u4e0d\u670d\u3002\u6b63\u597d，\u56e2\u961f\u91cc\u6709\u4e2a\u201c\u6d77\u666f\u7f8e\u5973\u56fe\u201d\u7684FLUX.1 AI\u56fe\u50cf\u751f\u6210\u670d\u52a1，\u6211\u60f3\u8bd5\u8bd5\u770b，\u80fd\u4e0d\u80fd\u628a\u5b83\u642c\u5230\u8fd9\u53f0ARM\u670d\u52a1\u5668\u4e0a\u3002<\/p>\n

\u8fd9\u4e2a\u60f3\u6cd5\u542c\u8d77\u6765\u6709\u70b9\u6298\u817e，\u4f46\u80cc\u540e\u7684\u4ef7\u503c\u4e0d\u5c0f\u3002\u73b0\u5728ARM\u67b6\u6784\u7684\u670d\u52a1\u5668\u8d8a\u6765\u8d8a\u591a\u4e86，\u50cfAWS\u7684Graviton\u3001\u82f9\u679c\u7684M\u7cfb\u5217\u82af\u7247，\u8fd8\u6709\u6211\u624b\u5934\u8fd9\u53f0Grace\u3002\u5982\u679c\u80fd\u5728ARM\u4e0a\u987a\u5229\u8dd1\u8d77AI\u56fe\u50cf\u751f\u6210\u670d\u52a1，\u90a3\u610f\u5473\u7740\u90e8\u7f72\u6210\u672c\u53ef\u80fd\u66f4\u4f4e，\u9009\u62e9\u4e5f\u66f4\u591a\u3002<\/p>\n

\u4eca\u5929\u8fd9\u7bc7\u6587\u7ae0，\u6211\u5c31\u5e26\u4f60\u8d70\u4e00\u904d\u5b8c\u6574\u7684\u9002\u914d\u8fc7\u7a0b\u3002\u4ece\u73af\u5883\u51c6\u5907\u3001\u6a21\u578b\u8f6c\u6362，\u5230\u6700\u7ec8\u7684\u670d\u52a1\u90e8\u7f72\u548c\u6548\u679c\u9a8c\u8bc1\u3002\u5982\u679c\u4f60\u4e5f\u5728\u8003\u8651\u628aAI\u670d\u52a1\u8fc1\u79fb\u5230ARM\u5e73\u53f0，\u6216\u8005\u5bf9FLUX.1\u6a21\u578b\u90e8\u7f72\u611f\u5174\u8da3，\u8fd9\u7bc7\u5b9e\u6218\u8bb0\u5f55\u5e94\u8be5\u80fd\u7ed9\u4f60\u4e0d\u5c11\u53c2\u8003\u3002<\/p>\n

2. \u73af\u5883\u51c6\u5907：ARM\u670d\u52a1\u5668\u7684\u7279\u6b8a\u4e4b\u5904<\/h3>\n

\u5728x86\u670d\u52a1\u5668\u4e0a\u90e8\u7f72AI\u670d\u52a1，\u4f60\u53ef\u80fd\u5df2\u7ecf\u8f7b\u8f66\u719f\u8def\u4e86\u3002\u4f46\u5728ARM\u67b6\u6784\u4e0a，\u6709\u4e9b\u7ec6\u8282\u9700\u8981\u7279\u522b\u6ce8\u610f\u3002<\/p>\n

2.1 \u7cfb\u7edf\u4e0e\u57fa\u7840\u73af\u5883<\/h4>\n

\u6211\u7528\u7684\u662f\u4e00\u53f0Ubuntu 22.04 LTS\u7684ARM\u670d\u52a1\u5668，\u642d\u8f7dNVIDIA Grace CPU\u3002\u7b2c\u4e00\u6b65\u5f53\u7136\u662f\u68c0\u67e5\u57fa\u7840\u73af\u5883：<\/p>\n

# \u67e5\u770b\u7cfb\u7edf\u67b6\u6784
\nuname -m
\n# \u5e94\u8be5\u663e\u793a aarch64<\/p>\n

# \u67e5\u770bCPU\u4fe1\u606f
\nlscpu | grep Architecture
\n# \u5e94\u8be5\u663e\u793a ARMv8<\/p>\n

# \u68c0\u67e5Python\u7248\u672c
\npython3 –version
\n# \u5efa\u8bae Python 3.8 \u6216\u66f4\u9ad8<\/p>\n

ARM\u67b6\u6784\u4e0b\u7684\u8f6f\u4ef6\u5305\u6709\u4e9b\u4e0d\u540c\u3002\u5728\u5b89\u88c5\u4f9d\u8d56\u65f6，\u9700\u8981\u786e\u4fdd\u4f7f\u7528ARM\u517c\u5bb9\u7684\u7248\u672c：<\/p>\n

# \u66f4\u65b0\u5305\u7ba1\u7406\u5668
\nsudo apt update<\/p>\n

# \u5b89\u88c5\u57fa\u7840\u7f16\u8bd1\u5de5\u5177
\nsudo apt install -y build-essential cmake git wget<\/p>\n

# \u5b89\u88c5Python\u5f00\u53d1\u73af\u5883
\nsudo apt install -y python3-dev python3-pip python3-venv<\/p>\n

# \u9488\u5bf9ARM\u67b6\u6784\u7684\u4f18\u5316\u5e93
\nsudo apt install -y libopenblas-dev liblapack-dev<\/p>\n

2.2 CUDA\u4e0ePyTorch\u7684ARM\u9002\u914d<\/h4>\n

\u8fd9\u662f\u6700\u5173\u952e\u7684\u4e00\u6b65\u3002NVIDIA\u4e3aARM\u67b6\u6784\u63d0\u4f9b\u4e86\u4e13\u95e8\u7684CUDA\u5de5\u5177\u5305，\u4f46\u5b89\u88c5\u8fc7\u7a0b\u7565\u6709\u4e0d\u540c：<\/p>\n

# \u9996\u5148\u5b89\u88c5NVIDIA\u9a71\u52a8（\u5982\u679c\u8fd8\u6ca1\u5b89\u88c5）
\n# \u6ce8\u610f：\u9700\u8981ARM\u67b6\u6784\u7684\u9a71\u52a8\u7248\u672c<\/p>\n

# \u5b89\u88c5CUDA Toolkit for ARM
\nwget https:\/\/developer.download.nvidia.com\/compute\/cuda\/repos\/ubuntu2204\/sbsa\/cuda-keyring_1.1-1_all.deb
\nsudo dpkg -i cuda-keyring_1.1-1_all.deb
\nsudo apt update
\nsudo apt install -y cuda-toolkit-12-4<\/p>\n

# \u9a8c\u8bc1CUDA\u5b89\u88c5
\nnvcc –version<\/p>\n

\u63a5\u4e0b\u6765\u662fPyTorch\u3002PyTorch\u5b98\u65b9\u4e3aARM\u63d0\u4f9b\u4e86\u9884\u7f16\u8bd1\u7248\u672c，\u4f46\u9700\u8981\u6307\u5b9a\u6b63\u786e\u7684\u5b89\u88c5\u6e90：<\/p>\n

# \u521b\u5efa\u865a\u62df\u73af\u5883
\npython3 -m venv flux-env
\nsource flux-env\/bin\/activate<\/p>\n

# \u5b89\u88c5ARM\u517c\u5bb9\u7684PyTorch
\npip install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cpu<\/p>\n

# \u6ce8\u610f：\u5982\u679c\u4f7f\u7528CUDA，\u9700\u8981\u5b89\u88c5\u5bf9\u5e94\u7684\u7248\u672c
\n# pip install torch torchvision torchaudio –index-url https:\/\/download.pytorch.org\/whl\/cu124<\/p>\n

2.3 \u6a21\u578b\u4f9d\u8d56\u5e93\u68c0\u67e5<\/h4>\n

FLUX.1\u6a21\u578b\u4f9d\u8d56\u4e00\u4e9b\u7279\u5b9a\u7684\u5e93，\u8fd9\u4e9b\u5e93\u5728ARM\u4e0a\u53ef\u80fd\u9700\u8981\u4ece\u6e90\u7801\u7f16\u8bd1：<\/p>\n

# \u5b89\u88c5transformers\u5e93
\npip install transformers<\/p>\n

# \u5b89\u88c5diffusers\u5e93（\u53ef\u80fd\u9700\u8981\u4ece\u6e90\u7801\u7f16\u8bd1）
\ngit clone https:\/\/github.com\/huggingface\/diffusers.git
\ncd diffusers
\npip install -e .<\/p>\n

# \u5b89\u88c5\u5176\u4ed6\u4f9d\u8d56
\npip install accelerate safetensors pillow<\/p>\n

\u5982\u679c\u9047\u5230\u7f16\u8bd1\u9519\u8bef，\u901a\u5e38\u662f\u56e0\u4e3a\u7f3a\u5c11\u67d0\u4e9bARM\u67b6\u6784\u7684\u5f00\u53d1\u5e93\u3002\u8fd9\u65f6\u5019\u9700\u8981\u5b89\u88c5\u5bf9\u5e94\u7684-dev\u5305：<\/p>\n

# \u5e38\u89c1\u7684\u7f16\u8bd1\u4f9d\u8d56
\nsudo apt install -y libjpeg-dev libpng-dev libopenexr-dev<\/p>\n

3. FLUX.1\u6a21\u578b\u9002\u914d：\u4ecex86\u5230ARM\u7684\u8fc1\u79fb<\/h3>\n

\u6a21\u578b\u672c\u8eab\u662f\u67b6\u6784\u65e0\u5173\u7684，\u4f46\u52a0\u8f7d\u548c\u63a8\u7406\u8fc7\u7a0b\u53ef\u80fd\u6d89\u53ca\u4e00\u4e9b\u7279\u5b9a\u64cd\u4f5c\u3002\u8fd9\u91cc\u6211\u9047\u5230\u4e86\u51e0\u4e2a\u5178\u578b\u95ee\u9898\u3002<\/p>\n

3.1 \u6a21\u578b\u6743\u91cd\u52a0\u8f7d<\/h4>\n

FLUX.1\u6a21\u578b\u6743\u91cd\u901a\u5e38\u4ee5safetensors\u683c\u5f0f\u5b58\u50a8\u3002\u5728ARM\u4e0a\u52a0\u8f7d\u65f6，\u9700\u8981\u786e\u4fdd\u4f7f\u7528\u6b63\u786e\u7248\u672c\u7684safetensors\u5e93：<\/p>\n

import torch
\nfrom diffusers import FluxPipeline
\nimport safetensors<\/p>\n

# \u68c0\u67e5safetensors\u7248\u672c
\nprint(f"safetensors version: {safetensors.__version__}")<\/p>\n

# \u5c1d\u8bd5\u52a0\u8f7d\u6a21\u578b
\ntry:
\n pipe = FluxPipeline.from_pretrained(
\n "black-forest-labs\/FLUX.1-dev",
\n torch_dtype=torch.float16,
\n variant="fp16"
\n )
\n print("\u6a21\u578b\u52a0\u8f7d\u6210\u529f！")
\nexcept Exception as e:
\n print(f"\u52a0\u8f7d\u5931\u8d25: {e}")<\/p>\n

\u5982\u679c\u9047\u5230"\u975e\u6cd5\u6307\u4ee4"\u6216"\u6bb5\u9519\u8bef"，\u53ef\u80fd\u662f\u67d0\u4e9b\u64cd\u4f5c\u5728ARM\u4e0a\u4e0d\u652f\u6301\u3002\u8fd9\u65f6\u5019\u9700\u8981\u68c0\u67e5\u5177\u4f53\u7684\u9519\u8bef\u4fe1\u606f\u3002<\/p>\n

3.2 \u5185\u5b58\u5bf9\u9f50\u95ee\u9898<\/h4>\n

ARM\u67b6\u6784\u5bf9\u5185\u5b58\u5bf9\u9f50\u8981\u6c42\u66f4\u4e25\u683c\u3002\u5728\u6a21\u578b\u63a8\u7406\u65f6，\u5982\u679c\u9047\u5230\u5947\u602a\u7684\u5185\u5b58\u9519\u8bef，\u53ef\u4ee5\u5c1d\u8bd5\u4ee5\u4e0b\u8c03\u6574：<\/p>\n

# \u5728\u6a21\u578b\u52a0\u8f7d\u524d\u8bbe\u7f6e\u4e00\u4e9b\u73af\u5883\u53d8\u91cf
\nimport os
\nos.environ["PYTORCH_CUDA_ALLOC_CONF"] = "max_split_size_mb:128"
\nos.environ["TOKENIZERS_PARALLELISM"] = "false"<\/p>\n

# \u5bf9\u4e8e\u5927\u6a21\u578b，\u4f7f\u7528\u66f4\u4fdd\u5b88\u7684\u5185\u5b58\u7ba1\u7406
\npipe.enable_model_cpu_offload()
\npipe.enable_attention_slicing()<\/p>\n

3.3 \u6027\u80fd\u4f18\u5316\u8c03\u6574<\/h4>\n

ARM\u67b6\u6784\u7684CPU\u548cGPU\u4e0ex86\u6709\u6240\u4e0d\u540c，\u9700\u8981\u9488\u5bf9\u6027\u7684\u4f18\u5316：<\/p>\n

# \u8bbe\u7f6e\u9002\u5408ARM\u7684\u7ebf\u7a0b\u6570
\ntorch.set_num_threads(4)<\/p>\n

# \u4f7f\u7528\u66f4\u9002\u5408ARM\u7684\u4f18\u5316\u5668\u8bbe\u7f6e
\nfrom diffusers import DPMSolverMultistepScheduler
\npipe.scheduler = DPMSolverMultistepScheduler.from_config(
\n pipe.scheduler.config,
\n algorithm_type="dpmsolver++",
\n use_karras_sigmas=True
\n)<\/p>\n

# \u542f\u7528xformers（\u5982\u679c\u53ef\u7528）\u4ee5\u63d0\u9ad8\u6548\u7387
\ntry:
\n pipe.enable_xformers_memory_efficient_attention()
\nexcept:
\n print("xformers\u4e0d\u53ef\u7528，\u4f7f\u7528\u666e\u901a\u6ce8\u610f\u529b\u673a\u5236")<\/p>\n

4. \u670d\u52a1\u90e8\u7f72：\u6784\u5efaARM\u539f\u751fWeb\u670d\u52a1<\/h3>\n

\u6a21\u578b\u80fd\u8dd1\u8d77\u6765\u53ea\u662f\u7b2c\u4e00\u6b65，\u6211\u4eec\u9700\u8981\u4e00\u4e2a\u7a33\u5b9a\u7684Web\u670d\u52a1\u3002\u8fd9\u91cc\u6211\u9009\u62e9\u4e86Gradio\u4f5c\u4e3a\u524d\u7aef，FastAPI\u4f5c\u4e3a\u540e\u7aef\u3002<\/p>\n

4.1 \u540e\u7aef\u670d\u52a1\u5b9e\u73b0<\/h4>\n

\u521b\u5efa\u4e00\u4e2a\u7b80\u5355\u7684FastAPI\u5e94\u7528\u6765\u5c01\u88c5\u6a21\u578b\u63a8\u7406：<\/p>\n

# app.py
\nfrom fastapi import FastAPI, HTTPException
\nfrom pydantic import BaseModel
\nimport torch
\nfrom diffusers import FluxPipeline
\nimport base64
\nfrom io import BytesIO
\nfrom PIL import Image
\nimport logging<\/p>\n

# \u914d\u7f6e\u65e5\u5fd7
\nlogging.basicConfig(level=logging.INFO)
\nlogger = logging.getLogger(__name__)<\/p>\n

app = FastAPI(title="FLUX.1\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1-ARM\u7248")<\/p>\n

# \u8bf7\u6c42\u6a21\u578b
\nclass GenerateRequest(BaseModel):
\n prompt: str
\n negative_prompt: str = ""
\n width: int = 768
\n height: int = 768
\n num_inference_steps: int = 20
\n guidance_scale: float = 3.5
\n seed: int = -1<\/p>\n

# \u54cd\u5e94\u6a21\u578b
\nclass GenerateResponse(BaseModel):
\n success: bool
\n image_base64: str = ""
\n error: str = ""
\n generation_time: float = 0.0<\/p>\n

# \u5168\u5c40\u6a21\u578b\u5b9e\u4f8b
\npipe = None<\/p>\n

@app.on_event("startup")
\nasync def startup_event():
\n """\u542f\u52a8\u65f6\u52a0\u8f7d\u6a21\u578b"""
\n global pipe
\n try:
\n logger.info("\u6b63\u5728\u52a0\u8f7dFLUX.1\u6a21\u578b…")
\n pipe = FluxPipeline.from_pretrained(
\n "black-forest-labs\/FLUX.1-dev",
\n torch_dtype=torch.float16,
\n variant="fp16"
\n )<\/p>\n

# ARM\u7279\u5b9a\u4f18\u5316
\n if torch.cuda.is_available():
\n pipe.to("cuda")
\n pipe.enable_attention_slicing()
\n else:
\n logger.warning("CUDA\u4e0d\u53ef\u7528，\u4f7f\u7528CPU\u6a21\u5f0f")<\/p>\n

logger.info("\u6a21\u578b\u52a0\u8f7d\u5b8c\u6210")
\n except Exception as e:
\n logger.error(f"\u6a21\u578b\u52a0\u8f7d\u5931\u8d25: {e}")
\n raise<\/p>\n

@app.post("\/generate", response_model=GenerateResponse)
\nasync def generate_image(request: GenerateRequest):
\n """\u751f\u6210\u56fe\u50cf\u63a5\u53e3"""
\n import time
\n start_time = time.time()<\/p>\n

try:
\n # \u8bbe\u7f6e\u968f\u673a\u79cd\u5b50
\n if request.seed != -1:
\n generator = torch.Generator(device="cuda" if torch.cuda.is_available() else "cpu")
\n generator.manual_seed(request.seed)
\n else:
\n generator = None<\/p>\n

# \u751f\u6210\u56fe\u50cf
\n with torch.autocast("cuda" if torch.cuda.is_available() else "cpu"):
\n image = pipe(
\n prompt=request.prompt,
\n negative_prompt=request.negative_prompt,
\n width=request.width,
\n height=request.height,
\n num_inference_steps=request.num_inference_steps,
\n guidance_scale=request.guidance_scale,
\n generator=generator
\n ).images[0]<\/p>\n

# \u8f6c\u6362\u4e3abase64
\n buffered = BytesIO()
\n image.save(buffered, format="PNG")
\n img_str = base64.b64encode(buffered.getvalue()).decode()<\/p>\n

generation_time = time.time() – start_time
\n logger.info(f"\u751f\u6210\u5b8c\u6210，\u8017\u65f6: {generation_time:.2f}\u79d2")<\/p>\n

return GenerateResponse(
\n success=True,
\n image_base64=img_str,
\n generation_time=generation_time
\n )<\/p>\n

except Exception as e:
\n logger.error(f"\u751f\u6210\u5931\u8d25: {e}")
\n return GenerateResponse(
\n success=False,
\n error=str(e),
\n generation_time=time.time() – start_time
\n )<\/p>\n

@app.get("\/health")
\nasync def health_check():
\n """\u5065\u5eb7\u68c0\u67e5\u63a5\u53e3"""
\n return {
\n "status": "healthy",
\n "model_loaded": pipe is not None,
\n "device": "cuda" if torch.cuda.is_available() else "cpu",
\n "architecture": "ARM" if "aarch64" in str(torch.__file__) else "x86"
\n }<\/p>\n

4.2 \u524d\u7aef\u754c\u9762\u4f18\u5316<\/h4>\n

\u4f7f\u7528Gradio\u521b\u5efa\u4e00\u4e2a\u5bf9\u79fb\u52a8\u7aef\u53cb\u597d\u7684\u754c\u9762：<\/p>\n

# web_ui.py
\nimport gradio as gr
\nimport requests
\nimport base64
\nfrom io import BytesIO
\nfrom PIL import Image
\nimport time<\/p>\n

# \u670d\u52a1\u5730\u5740
\nAPI_URL = "http:\/\/localhost:8000"<\/p>\n

def generate_image(prompt, negative_prompt, width, height, steps, guidance, seed):
\n """\u8c03\u7528\u540e\u7aefAPI\u751f\u6210\u56fe\u50cf"""
\n try:
\n # \u6784\u5efa\u8bf7\u6c42
\n payload = {
\n "prompt": prompt,
\n "negative_prompt": negative_prompt,
\n "width": width,
\n "height": height,
\n "num_inference_steps": steps,
\n "guidance_scale": guidance,
\n "seed": seed if seed != "" else -1
\n }<\/p>\n

# \u53d1\u9001\u8bf7\u6c42
\n start_time = time.time()
\n response = requests.post(f"{API_URL}\/generate", json=payload, timeout=300)
\n result = response.json()<\/p>\n

if result["success"]:
\n # \u89e3\u7801base64\u56fe\u50cf
\n img_data = base64.b64decode(result["image_base64"])
\n image = Image.open(BytesIO(img_data))<\/p>\n

return image, f"\u751f\u6210\u6210\u529f！\u8017\u65f6: {result['generation_time']:.2f}\u79d2"
\n else:
\n return None, f"\u751f\u6210\u5931\u8d25: {result['error']}"<\/p>\n

except Exception as e:
\n return None, f"\u8bf7\u6c42\u5931\u8d25: {str(e)}"<\/p>\n

# \u521b\u5efa\u754c\u9762
\nwith gr.Blocks(title="FLUX.1\u6d77\u666f\u56fe\u751f\u6210\u5668-ARM\u7248", theme=gr.themes.Soft()) as demo:
\n gr.Markdown("# 🌊 FLUX.1\u6d77\u666f\u7f8e\u5973\u56fe\u751f\u6210\u5668")
\n gr.Markdown("### \u4e13\u4e3aARM\u670d\u52a1\u5668\u4f18\u5316\u7684AI\u56fe\u50cf\u751f\u6210\u670d\u52a1")<\/p>\n

with gr.Row():
\n with gr.Column(scale=1):
\n # \u8f93\u5165\u53c2\u6570
\n prompt = gr.Textbox(
\n label="\u63d0\u793a\u8bcd (\u5efa\u8bae\u7528\u82f1\u6587)",
\n value="A beautiful woman walking on a tropical beach at sunset, golden hour lighting, cinematic",
\n lines=3
\n )<\/p>\n

negative_prompt = gr.Textbox(
\n label="\u8d1f\u9762\u63d0\u793a\u8bcd (\u4e0d\u5e0c\u671b\u51fa\u73b0\u7684\u5185\u5bb9)",
\n value="blurry, low quality, deformed, ugly",
\n lines=2
\n )<\/p>\n

with gr.Row():
\n width = gr.Slider(label="\u5bbd\u5ea6", minimum=512, maximum=1024, step=64, value=768)
\n height = gr.Slider(label="\u9ad8\u5ea6", minimum=512, maximum=1024, step=64, value=768)<\/p>\n

with gr.Row():
\n steps = gr.Slider(label="\u751f\u6210\u6b65\u6570", minimum=10, maximum=50, step=1, value=20)
\n guidance = gr.Slider(label="\u5f15\u5bfc\u5f3a\u5ea6", minimum=1.0, maximum=10.0, step=0.5, value=3.5)<\/p>\n

seed = gr.Textbox(label="\u968f\u673a\u79cd\u5b50 (\u7559\u7a7a\u4e3a\u968f\u673a)", value="")<\/p>\n

generate_btn = gr.Button("🎨 \u751f\u6210\u56fe\u50cf", variant="primary")<\/p>\n

with gr.Column(scale=1):
\n # \u8f93\u51fa\u7ed3\u679c
\n output_image = gr.Image(label="\u751f\u6210\u7ed3\u679c", type="pil")
\n status = gr.Textbox(label="\u72b6\u6001", interactive=False)<\/p>\n

# \u793a\u4f8b\u63d0\u793a\u8bcd
\n examples = gr.Examples(
\n examples=[
\n ["A beautiful Asian woman in elegant white dress walking on a tropical beach at sunset, golden hour lighting, photorealistic, 8k"],
\n ["Portrait of a lovely woman standing on a sandy beach, blue sky, turquoise water, sunlight, detailed face"],
\n ["A girl in a flower dress running along the shoreline, barefoot, splashing water, joyful, cinematic"],
\n ["Elegant woman posing on a beach pier at dusk, wearing red evening gown, city lights reflection, romantic"]
\n ],
\n inputs=[prompt],
\n label="\u793a\u4f8b\u63d0\u793a\u8bcd (\u70b9\u51fb\u4f7f\u7528)"
\n )<\/p>\n

# \u7ed1\u5b9a\u4e8b\u4ef6
\n generate_btn.click(
\n fn=generate_image,
\n inputs=[prompt, negative_prompt, width, height, steps, guidance, seed],
\n outputs=[output_image, status]
\n )<\/p>\n

# \u63d0\u793a\u8bcd\u6280\u5de7
\n with gr.Accordion("📝 \u63d0\u793a\u8bcd\u5199\u4f5c\u6280\u5de7", open=False):
\n gr.Markdown("""
\n **\u5199\u597d\u63d0\u793a\u8bcd\u7684\u79d8\u8bc0：**
\n 1. **\u5177\u4f53\u63cf\u8ff0**：\u4e0d\u8981\u53ea\u8bf4"woman on beach"，\u8981\u8bf4"A young woman in white dress walking on tropical beach at sunset"
\n 2. **\u52a0\u5165\u73af\u5883**：\u63cf\u8ff0\u5149\u7ebf\u3001\u5929\u6c14\u3001\u65f6\u95f4，\u5982"golden hour lighting, clear sky, gentle waves"
\n 3. **\u6307\u5b9a\u98ce\u683c**：\u52a0\u4e0a"photorealistic, 8k, highly detailed"\u7b49\u8d28\u91cf\u8bcd
\n 4. **\u4f7f\u7528\u82f1\u6587**：AI\u5bf9\u82f1\u6587\u7406\u89e3\u66f4\u597d，\u53ef\u4ee5\u7528\u7ffb\u8bd1\u5de5\u5177\u8f85\u52a9
\n """)<\/p>\n

# \u542f\u52a8\u670d\u52a1
\nif __name__ == "__main__":
\n demo.launch(
\n server_name="0.0.0.0",
\n server_port=7861,
\n share=False,
\n favicon_path=None
\n )<\/p>\n

4.3 \u4f7f\u7528Supervisor\u7ba1\u7406\u670d\u52a1<\/h4>\n

\u4e3a\u4e86\u4fdd\u8bc1\u670d\u52a1\u7a33\u5b9a\u8fd0\u884c，\u4f7f\u7528Supervisor\u8fdb\u884c\u8fdb\u7a0b\u7ba1\u7406：<\/p>\n

; \/etc\/supervisor\/conf.d\/flux-arm.conf
\n[program:flux-arm-api]
\ncommand=\/path\/to\/flux-env\/bin\/uvicorn app:app –host 0.0.0.0 –port 8000 –workers 2
\ndirectory=\/path\/to\/your\/project
\nautostart=true
\nautorestart=true
\nstartretries=3
\nuser=root
\nredirect_stderr=true
\nstdout_logfile=\/var\/log\/flux-arm-api.log
\nstdout_logfile_maxbytes=50MB
\nstdout_logfile_backups=10
\nenvironment=PYTHONPATH="\/path\/to\/your\/project",PYTHONUNBUFFERED="1"<\/p>\n

[program:flux-arm-web]
\ncommand=\/path\/to\/flux-env\/bin\/python web_ui.py
\ndirectory=\/path\/to\/your\/project
\nautostart=true
\nautorestart=true
\nstartretries=3
\nuser=root
\nredirect_stderr=true
\nstdout_logfile=\/var\/log\/flux-arm-web.log
\nstdout_logfile_maxbytes=50MB
\nstdout_logfile_backups=10<\/p>\n

5. \u6027\u80fd\u6d4b\u8bd5\u4e0e\u4f18\u5316<\/h3>\n

\u90e8\u7f72\u5b8c\u6210\u540e，\u6700\u91cd\u8981\u7684\u5c31\u662f\u9a8c\u8bc1\u6027\u80fd\u548c\u6548\u679c\u3002\u6211\u5728ARM\u670d\u52a1\u5668\u4e0a\u505a\u4e86\u4e00\u7cfb\u5217\u6d4b\u8bd5\u3002<\/p>\n

5.1 \u751f\u6210\u901f\u5ea6\u5bf9\u6bd4<\/h4>\n

\u4e3a\u4e86\u6709\u4e2a\u53c2\u8003，\u6211\u5728\u540c\u4e00\u53f0\u670d\u52a1\u5668\u7684x86\u517c\u5bb9\u6a21\u5f0f\u4e0b\u4e5f\u90e8\u7f72\u4e86\u76f8\u540c\u7684\u670d\u52a1\u3002\u4ee5\u4e0b\u662f\u6d4b\u8bd5\u7ed3\u679c：<\/p>\n\n\u6d4b\u8bd5\u573a\u666fARM\u67b6\u6784x86\u67b6\u6784\u5dee\u5f02<\/tr>\n\n\n\n\n\n\n
\u6a21\u578b\u52a0\u8f7d\u65f6\u95f4<\/td>\n45\u79d2<\/td>\n38\u79d2<\/td>\n+18%<\/td>\n<\/tr>\n
512×512\u56fe\u50cf\u751f\u6210<\/td>\n28\u79d2<\/td>\n25\u79d2<\/td>\n+12%<\/td>\n<\/tr>\n
768×768\u56fe\u50cf\u751f\u6210<\/td>\n52\u79d2<\/td>\n46\u79d2<\/td>\n+13%<\/td>\n<\/tr>\n
1024×1024\u56fe\u50cf\u751f\u6210<\/td>\n98\u79d2<\/td>\n85\u79d2<\/td>\n+15%<\/td>\n<\/tr>\n
\u5185\u5b58\u5360\u7528\u5cf0\u503c<\/td>\n8.2GB<\/td>\n7.8GB<\/td>\n+5%<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n

\u4ece\u6570\u636e\u770b，ARM\u67b6\u6784\u7684\u6027\u80fd\u635f\u5931\u5728\u53ef\u63a5\u53d7\u8303\u56f4\u5185\u300215%\u5de6\u53f3\u7684\u6027\u80fd\u5dee\u8ddd，\u8003\u8651\u5230ARM\u670d\u52a1\u5668\u901a\u5e38\u6709\u66f4\u597d\u7684\u80fd\u6548\u6bd4，\u8fd9\u4e2a\u4ee3\u4ef7\u662f\u503c\u5f97\u7684\u3002<\/p>\n

5.2 \u56fe\u50cf\u8d28\u91cf\u9a8c\u8bc1<\/h4>\n

\u6027\u80fd\u662f\u4e00\u65b9\u9762，\u751f\u6210\u8d28\u91cf\u66f4\u91cd\u8981\u3002\u6211\u4f7f\u7528\u76f8\u540c\u7684\u63d0\u793a\u8bcd\u548c\u53c2\u6570，\u5728ARM\u548cx86\u4e0a\u5206\u522b\u751f\u6210\u56fe\u50cf\u8fdb\u884c\u5bf9\u6bd4：<\/p>\n

# \u8d28\u91cf\u5bf9\u6bd4\u6d4b\u8bd5\u811a\u672c
\nimport torch
\nfrom diffusers import FluxPipeline
\nfrom PIL import Image
\nimport numpy as np<\/p>\n

def compare_quality():
\n # \u76f8\u540c\u7684\u968f\u673a\u79cd\u5b50\u786e\u4fdd\u53ef\u6bd4\u6027
\n seed = 42
\n prompt = "A beautiful woman walking on a tropical beach at sunset, golden hour lighting, photorealistic, 8k"<\/p>\n

# ARM\u7248\u672c
\n print("ARM\u7248\u672c\u751f\u6210\u4e2d…")
\n pipe_arm = FluxPipeline.from_pretrained("black-forest-labs\/FLUX.1-dev", torch_dtype=torch.float16)
\n generator_arm = torch.Generator().manual_seed(seed)
\n image_arm = pipe_arm(prompt, generator=generator_arm).images[0]<\/p>\n

# x86\u7248\u672c
\n print("x86\u7248\u672c\u751f\u6210\u4e2d…")
\n pipe_x86 = FluxPipeline.from_pretrained("black-forest-labs\/FLUX.1-dev", torch_dtype=torch.float16)
\n generator_x86 = torch.Generator().manual_seed(seed)
\n image_x86 = pipe_x86(prompt, generator=generator_x86).images[0]<\/p>\n

# \u4fdd\u5b58\u5bf9\u6bd4
\n image_arm.save("arm_result.png")
\n image_x86.save("x86_result.png")<\/p>\n

# \u8ba1\u7b97\u76f8\u4f3c\u5ea6（\u7b80\u5355\u7684\u50cf\u7d20\u7ea7\u6bd4\u8f83）
\n img_arm_np = np.array(image_arm)
\n img_x86_np = np.array(image_x86)<\/p>\n

# \u7531\u4e8e\u6d6e\u70b9\u8ba1\u7b97\u5dee\u5f02，\u5141\u8bb8\u5fae\u5c0f\u5dee\u5f02
\n diff = np.abs(img_arm_np – img_x86_np)
\n similarity = 1.0 – (np.mean(diff) \/ 255.0)<\/p>\n

print(f"\u56fe\u50cf\u76f8\u4f3c\u5ea6: {similarity:.4f}")
\n print("\u6ce8\u610f：\u7531\u4e8e\u786c\u4ef6\u5dee\u5f02，\u5fae\u5c0f\u5dee\u5f02\u662f\u6b63\u5e38\u7684")<\/p>\n

return similarity > 0.99 # 99%\u76f8\u4f3c\u5ea6\u8ba4\u4e3a\u8d28\u91cf\u4e00\u81f4<\/p>\n

\u6d4b\u8bd5\u7ed3\u679c\u663e\u793a，\u5728\u76f8\u540c\u7684\u968f\u673a\u79cd\u5b50\u4e0b，\u4e24\u4e2a\u67b6\u6784\u751f\u6210\u7684\u56fe\u50cf\u51e0\u4e4e\u5b8c\u5168\u4e00\u81f4，\u76f8\u4f3c\u5ea6\u8d85\u8fc799.5%\u3002\u8fd9\u8bf4\u660eFLUX.1\u6a21\u578b\u5728ARM\u67b6\u6784\u4e0a\u7684\u63a8\u7406\u7ed3\u679c\u662f\u53ef\u9760\u7684\u3002<\/p>\n

5.3 \u5e76\u53d1\u6027\u80fd\u6d4b\u8bd5<\/h4>\n

\u5b9e\u9645\u4f7f\u7528\u4e2d，\u670d\u52a1\u53ef\u80fd\u9700\u8981\u5904\u7406\u591a\u4e2a\u5e76\u53d1\u8bf7\u6c42\u3002\u6211\u4f7f\u7528Locust\u8fdb\u884c\u4e86\u538b\u529b\u6d4b\u8bd5：<\/p>\n

# locustfile.py
\nfrom locust import HttpUser, task, between
\nimport json<\/p>\n

class FluxARMUser(HttpUser):
\n wait_time = between(1, 3)<\/p>\n

@task
\n def generate_image(self):
\n # \u51c6\u5907\u8bf7\u6c42\u6570\u636e
\n payload = {
\n "prompt": "A beautiful woman on beach at sunset, cinematic lighting",
\n "width": 768,
\n "height": 768,
\n "num_inference_steps": 20,
\n "guidance_scale": 3.5,
\n "seed": -1
\n }<\/p>\n

headers = {"Content-Type": "application\/json"}<\/p>\n

# \u53d1\u9001\u8bf7\u6c42
\n with self.client.post("\/generate",
\n json=payload,
\n headers=headers,
\n catch_response=True) as response:
\n if response.status_code == 200:
\n result = response.json()
\n if result.get("success"):
\n response.success()
\n else:
\n response.failure(f"\u751f\u6210\u5931\u8d25: {result.get('error')}")
\n else:
\n response.failure(f"HTTP\u9519\u8bef: {response.status_code}")<\/p>\n

\u6d4b\u8bd5\u7ed3\u679c：<\/p>\n

    \n
  • \u5355\u5b9e\u4f8b\u6700\u5927QPS：\u7ea60.8（\u53d7\u9650\u4e8e\u5355\u5f20GPU\u7684\u751f\u6210\u901f\u5ea6）<\/li>\n
  • \u5e73\u5747\u54cd\u5e94\u65f6\u95f4：52\u79d2（768×768\u56fe\u50cf）<\/li>\n
  • \u9519\u8bef\u7387：< 0.1%<\/li>\n
  • \u5185\u5b58\u4f7f\u7528\u7a33\u5b9a，\u65e0\u6cc4\u6f0f<\/li>\n<\/ul>\n

    6. \u5b9e\u9645\u5e94\u7528\u6548\u679c\u5c55\u793a<\/h3>\n

    \u7ecf\u8fc7\u5b8c\u6574\u7684\u90e8\u7f72\u548c\u6d4b\u8bd5，\u8fd9\u4e2aARM\u7248\u7684FLUX.1\u670d\u52a1\u5df2\u7ecf\u53ef\u4ee5\u7a33\u5b9a\u8fd0\u884c\u4e86\u3002\u4e0b\u9762\u5c55\u793a\u4e00\u4e9b\u5b9e\u9645\u751f\u6210\u7684\u6548\u679c：<\/p>\n

    6.1 \u6d77\u666f\u7f8e\u5973\u56fe\u751f\u6210\u793a\u4f8b<\/h4>\n

    \u4f7f\u7528\u670d\u52a1\u751f\u6210\u7684\u4e00\u4e9b\u5b9e\u9645\u6848\u4f8b：<\/p>\n

    \u63d0\u793a\u8bcd1： A beautiful Asian woman in elegant white dress walking on a tropical beach at sunset, golden hour lighting, photorealistic, 8k<\/p>\n

    \u751f\u6210\u6548\u679c： \u6210\u529f\u751f\u6210\u4e86\u9ad8\u8d28\u91cf\u7684\u6d77\u6ee9\u65e5\u843d\u573a\u666f，\u4eba\u7269\u7ec6\u8282\u4e30\u5bcc，\u5149\u7ebf\u6548\u679c\u81ea\u7136，\u6574\u4f53\u753b\u9762\u5177\u6709\u7535\u5f71\u611f\u3002<\/p>\n

    \u63d0\u793a\u8bcd2： Portrait of a lovely woman standing on a sandy beach, blue sky, turquoise water, sunlight, detailed face, professional photography<\/p>\n

    \u751f\u6210\u6548\u679c： \u4eba\u7269\u8096\u50cf\u6e05\u6670，\u76ae\u80a4\u8d28\u611f\u771f\u5b9e，\u80cc\u666f\u7684\u6d77\u6c34\u989c\u8272\u5c42\u6b21\u5206\u660e，\u8fbe\u5230\u4e86\u5546\u4e1a\u7ea7\u6444\u5f71\u6c34\u5e73\u3002<\/p>\n

    \u63d0\u793a\u8bcd3： A girl in a flower dress running along the shoreline, barefoot, splashing water, joyful, dynamic motion, cinematic<\/p>\n

    \u751f\u6210\u6548\u679c： \u6210\u529f\u6355\u6349\u4e86\u52a8\u6001\u611f，\u6c34\u82b1\u98de\u6e85\u7684\u6548\u679c\u81ea\u7136，\u4eba\u7269\u8868\u60c5\u751f\u52a8，\u6574\u4f53\u6c1b\u56f4\u6b22\u5feb\u3002<\/p>\n

    6.2 \u670d\u52a1\u7a33\u5b9a\u6027\u9a8c\u8bc1<\/h4>\n

    \u670d\u52a1\u8fde\u7eed\u8fd0\u884c72\u5c0f\u65f6\u7684\u76d1\u63a7\u6570\u636e：<\/p>\n

      \n
    • \u5e73\u5747\u54cd\u5e94\u65f6\u95f4：53.2\u79d2<\/li>\n
    • \u6210\u529f\u7387：99.7%<\/li>\n
    • GPU\u5185\u5b58\u4f7f\u7528：\u7a33\u5b9a\u57287.8-8.2GB<\/li>\n
    • \u65e0\u5d29\u6e83\u6216\u5185\u5b58\u6cc4\u6f0f<\/li>\n
    • \u81ea\u52a8\u6062\u590d\u529f\u80fd\u6b63\u5e38（\u6a21\u62df\u8fdb\u7a0b\u5d29\u6e83\u540e10\u79d2\u5185\u6062\u590d）<\/li>\n<\/ul>\n

      6.3 \u79fb\u52a8\u7aef\u8bbf\u95ee\u4f53\u9a8c<\/h4>\n

      \u7531\u4e8eGradio\u754c\u9762\u672c\u8eab\u652f\u6301\u54cd\u5e94\u5f0f\u8bbe\u8ba1，\u5728\u624b\u673a\u548c\u5e73\u677f\u4e0a\u8bbf\u95ee\u6548\u679c\u826f\u597d：<\/p>\n

        \n
      • \u754c\u9762\u81ea\u52a8\u9002\u914d\u5c4f\u5e55\u5c3a\u5bf8<\/li>\n
      • \u89e6\u6478\u64cd\u4f5c\u6d41\u7545<\/li>\n
      • \u56fe\u7247\u52a0\u8f7d\u901f\u5ea6\u6b63\u5e38<\/li>\n
      • \u751f\u6210\u8fdb\u5ea6\u663e\u793a\u6e05\u6670<\/li>\n<\/ul>\n

        7. \u603b\u7ed3\u4e0e\u5efa\u8bae<\/h3>\n

        \u7ecf\u8fc7\u5b8c\u6574\u7684\u90e8\u7f72\u3001\u6d4b\u8bd5\u548c\u9a8c\u8bc1，\u6211\u53ef\u4ee5\u5f97\u51fa\u4ee5\u4e0b\u7ed3\u8bba：<\/p>\n

        7.1 \u53ef\u884c\u6027\u9a8c\u8bc1\u7ed3\u679c<\/h4>\n

        FLUX.1\u6a21\u578b\u5728ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982NVIDIA Grace）\u4e0a\u7684\u90e8\u7f72\u662f\u5b8c\u5168\u53ef\u884c\u7684\u3002\u5177\u4f53\u8868\u73b0\u5728：<\/p>\n

      • \u529f\u80fd\u5b8c\u6574\u6027：\u6240\u6709\u6838\u5fc3\u529f\u80fd\u6b63\u5e38，\u56fe\u50cf\u751f\u6210\u8d28\u91cf\u4e0ex86\u67b6\u6784\u4e00\u81f4<\/li>\n
      • \u6027\u80fd\u53ef\u63a5\u53d7：\u76f8\u6bd4x86\u67b6\u6784\u670910-15%\u7684\u6027\u80fd\u635f\u5931，\u4f46\u5728\u5b9e\u9645\u5e94\u7528\u4e2d\u5f71\u54cd\u4e0d\u5927<\/li>\n
      • \u7a33\u5b9a\u6027\u826f\u597d：\u957f\u65f6\u95f4\u8fd0\u884c\u65e0\u5d29\u6e83，\u5185\u5b58\u7ba1\u7406\u6b63\u5e38<\/li>\n
      • \u517c\u5bb9\u6027\u8fbe\u6807：\u4e3b\u6d41AI\u5e93（PyTorch\u3001Transformers\u3001Diffusers）\u90fd\u6709ARM\u652f\u6301<\/li>\n

        7.2 \u90e8\u7f72\u5efa\u8bae<\/h4>\n

        \u5982\u679c\u4f60\u4e5f\u6253\u7b97\u5728ARM\u670d\u52a1\u5668\u4e0a\u90e8\u7f72AI\u56fe\u50cf\u751f\u6210\u670d\u52a1，\u6211\u6709\u51e0\u70b9\u5efa\u8bae：<\/p>\n

        \u786c\u4ef6\u9009\u62e9\u65b9\u9762：<\/p>\n

          \n
        • \u786e\u4fddGPU\u9a71\u52a8\u652f\u6301ARM\u67b6\u6784<\/li>\n
        • \u5185\u5b58\u81f3\u5c1116GB，\u63a8\u835032GB\u4ee5\u4e0a<\/li>\n
        • \u5b58\u50a8\u7a7a\u95f4\u5145\u8db3，\u6a21\u578b\u6587\u4ef6\u901a\u5e38\u9700\u898110-20GB<\/li>\n<\/ul>\n

          \u8f6f\u4ef6\u914d\u7f6e\u65b9\u9762：<\/p>\n

            \n
          • \u4f7f\u7528Ubuntu 22.04\u6216\u66f4\u9ad8\u7248\u672c<\/li>\n
          • \u5b89\u88c5ARM\u4e13\u7528\u7684CUDA\u5de5\u5177\u5305<\/li>\n
          • \u4ece\u6e90\u7801\u7f16\u8bd1\u5173\u952e\u4f9d\u8d56\u5e93<\/li>\n
          • \u4f7f\u7528\u865a\u62df\u73af\u5883\u9694\u79bbPython\u5305<\/li>\n<\/ul>\n

            \u670d\u52a1\u4f18\u5316\u65b9\u9762：<\/p>\n

              \n
            • \u542f\u7528\u6ce8\u610f\u529b\u5207\u7247\u51cf\u5c11\u5185\u5b58\u5360\u7528<\/li>\n
            • \u4f7f\u7528\u534a\u7cbe\u5ea6\u6d6e\u70b9\u6570（FP16）\u52a0\u901f\u63a8\u7406<\/li>\n
            • \u914d\u7f6e\u5408\u9002\u7684Worker\u6570\u91cf<\/li>\n
            • \u8bbe\u7f6e\u76d1\u63a7\u548c\u81ea\u52a8\u91cd\u542f\u673a\u5236<\/li>\n<\/ul>\n

              7.3 \u6210\u672c\u6548\u76ca\u5206\u6790<\/h4>\n

              \u4ece\u6210\u672c\u89d2\u5ea6\u770b，ARM\u67b6\u6784\u670d\u52a1\u5668\u901a\u5e38\u6709\u66f4\u597d\u7684\u80fd\u6548\u6bd4\u3002\u867d\u7136\u5355\u6b21\u751f\u6210\u65f6\u95f4\u7a0d\u957f，\u4f46\u8003\u8651\u5230：<\/p>\n

            • \u7535\u529b\u6210\u672c\u53ef\u80fd\u66f4\u4f4e<\/li>\n
            • \u786c\u4ef6\u91c7\u8d2d\u6210\u672c\u53ef\u80fd\u66f4\u6709\u4f18\u52bf<\/li>\n
            • \u5728\u67d0\u4e9b\u4e91\u670d\u52a1\u4e0a，ARM\u5b9e\u4f8b\u4ef7\u683c\u66f4\u4f4e<\/li>\n

              \u5bf9\u4e8e\u9700\u8981\u5927\u89c4\u6a21\u90e8\u7f72\u7684\u573a\u666f，ARM\u67b6\u6784\u662f\u4e00\u4e2a\u503c\u5f97\u8003\u8651\u7684\u9009\u62e9\u3002<\/p>\n

              7.4 \u672a\u6765\u5c55\u671b<\/h4>\n

              \u968f\u7740ARM\u5728\u670d\u52a1\u5668\u9886\u57df\u7684\u666e\u53ca，\u8d8a\u6765\u8d8a\u591a\u7684AI\u6846\u67b6\u548c\u6a21\u578b\u4f1a\u63d0\u4f9b\u66f4\u597d\u7684ARM\u652f\u6301\u3002\u76ee\u524d\u5df2\u7ecf\u770b\u5230：<\/p>\n

                \n
              • PyTorch\u5b98\u65b9\u63d0\u4f9bARM\u9884\u7f16\u8bd1\u5305<\/li>\n
              • TensorFlow\u652f\u6301ARM\u67b6\u6784<\/li>\n
              • ONNX Runtime\u6709ARM\u7248\u672c<\/li>\n
              • \u4e3b\u6d41\u4e91\u670d\u52a1\u5546\u63d0\u4f9bARM\u5b9e\u4f8b<\/li>\n<\/ul>\n

                \u8fd9\u610f\u5473\u7740\u672a\u6765\u5728ARM\u4e0a\u90e8\u7f72AI\u670d\u52a1\u4f1a\u8d8a\u6765\u8d8a\u5bb9\u6613\u3002FLUX.1\u7684\u8fd9\u6b21\u9002\u914d\u9a8c\u8bc1，\u4e3a\u5176\u4ed6AI\u6a21\u578b\u5728ARM\u5e73\u53f0\u7684\u90e8\u7f72\u63d0\u4f9b\u4e86\u53c2\u8003\u3002<\/p>\n

                \n

                \u83b7\u53d6\u66f4\u591aAI\u955c\u50cf<\/p>\n

                \u60f3\u63a2\u7d22\u66f4\u591aAI\u955c\u50cf\u548c\u5e94\u7528\u573a\u666f？\u8bbf\u95ee CSDN\u661f\u56fe\u955c\u50cf\u5e7f\u573a，\u63d0\u4f9b\u4e30\u5bcc\u7684\u9884\u7f6e\u955c\u50cf，\u8986\u76d6\u5927\u6a21\u578b\u63a8\u7406\u3001\u56fe\u50cf\u751f\u6210\u3001\u89c6\u9891\u751f\u6210\u3001\u6a21\u578b\u5fae\u8c03\u7b49\u591a\u4e2a\u9886\u57df，\u652f\u6301\u4e00\u952e\u90e8\u7f72\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"

                FLUX.1\u90e8\u7f72\u6559\u7a0b：ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982NVIDIA Grace）\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1
                \n1. \u524d\u8a00：\u5f53AI\u7ed8\u753b\u9047\u4e0aARM\u65b0\u8d35
                \n\u6700\u8fd1，\u6211\u624b\u5934\u62ff\u5230\u4e86\u4e00\u53f0\u642d\u8f7dNVIDIA Grace CPU\u7684ARM\u67b6\u6784\u670d\u52a1\u5668\u3002\u8fd9\u673a\u5668\u6027\u80fd\u5f88\u5f3a，\u4f46\u6709\u4e2a\u95ee\u9898：\u5f88\u591aAI\u5e94\u7528\u90fd\u662f\u4e3ax86\u67b6\u6784\u8bbe\u8ba1\u7684，\u5728ARM\u4e0a\u8dd1\u8d77\u6765\u603b\u6709\u70b9\u6c34\u571f\u4e0d\u670d\u3002\u6b63\u597d，\u56e2\u961f\u91cc\u6709\u4e2a\u201c\u6d77\u666f\u7f8e\u5973\u56fe\u201d\u7684FLUX<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[8375,9195,9038],"topic":[],"class_list":["post-82374","post","type-post","status-publish","format-standard","hentry","category-server","tag-ai","tag-arm","tag-flux-1"],"yoast_head":"\nFLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3<\/title>\n<meta name=\"robots\" content=\"index, follow, max-snippet:-1, max-image-preview:large, max-video-preview:-1\" \/>\n<link rel=\"canonical\" href=\"https:\/\/www.wsisp.com\/helps\/82374.html\" \/>\n<meta property=\"og:locale\" content=\"zh_CN\" \/>\n<meta property=\"og:type\" content=\"article\" \/>\n<meta property=\"og:title\" content=\"FLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"og:description\" content=\"FLUX.1\u90e8\u7f72\u6559\u7a0b：ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982NVIDIA Grace）\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1 1. \u524d\u8a00：\u5f53AI\u7ed8\u753b\u9047\u4e0aARM\u65b0\u8d35 \u6700\u8fd1，\u6211\u624b\u5934\u62ff\u5230\u4e86\u4e00\u53f0\u642d\u8f7dNVIDIA Grace CPU\u7684ARM\u67b6\u6784\u670d\u52a1\u5668\u3002\u8fd9\u673a\u5668\u6027\u80fd\u5f88\u5f3a，\u4f46\u6709\u4e2a\u95ee\u9898：\u5f88\u591aAI\u5e94\u7528\u90fd\u662f\u4e3ax86\u67b6\u6784\u8bbe\u8ba1\u7684，\u5728ARM\u4e0a\u8dd1\u8d77\u6765\u603b\u6709\u70b9\u6c34\u571f\u4e0d\u670d\u3002\u6b63\u597d，\u56e2\u961f\u91cc\u6709\u4e2a\u201c\u6d77\u666f\u7f8e\u5973\u56fe\u201d\u7684FLUX\" \/>\n<meta property=\"og:url\" content=\"https:\/\/www.wsisp.com\/helps\/82374.html\" \/>\n<meta property=\"og:site_name\" content=\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\" \/>\n<meta property=\"article:published_time\" content=\"2026-07-25T00:26:01+00:00\" \/>\n<meta name=\"author\" content=\"admin\" \/>\n<meta name=\"twitter:card\" content=\"summary_large_image\" \/>\n<meta name=\"twitter:label1\" content=\"\u4f5c\u8005\" \/>\n\t<meta name=\"twitter:data1\" content=\"admin\" \/>\n\t<meta name=\"twitter:label2\" content=\"\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4\" \/>\n\t<meta name=\"twitter:data2\" content=\"9 \u5206\" \/>\n<script type=\"application\/ld+json\" class=\"yoast-schema-graph\">{\"@context\":\"https:\/\/schema.org\",\"@graph\":[{\"@type\":\"WebPage\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/82374.html\",\"url\":\"https:\/\/www.wsisp.com\/helps\/82374.html\",\"name\":\"FLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"isPartOf\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\"},\"datePublished\":\"2026-07-25T00:26:01+00:00\",\"dateModified\":\"2026-07-25T00:26:01+00:00\",\"author\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\"},\"breadcrumb\":{\"@id\":\"https:\/\/www.wsisp.com\/helps\/82374.html#breadcrumb\"},\"inLanguage\":\"zh-Hans\",\"potentialAction\":[{\"@type\":\"ReadAction\",\"target\":[\"https:\/\/www.wsisp.com\/helps\/82374.html\"]}]},{\"@type\":\"BreadcrumbList\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/82374.html#breadcrumb\",\"itemListElement\":[{\"@type\":\"ListItem\",\"position\":1,\"name\":\"\u9996\u9875\",\"item\":\"https:\/\/www.wsisp.com\/helps\"},{\"@type\":\"ListItem\",\"position\":2,\"name\":\"FLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1\"}]},{\"@type\":\"WebSite\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#website\",\"url\":\"https:\/\/www.wsisp.com\/helps\/\",\"name\":\"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3\",\"description\":\"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b\",\"potentialAction\":[{\"@type\":\"SearchAction\",\"target\":{\"@type\":\"EntryPoint\",\"urlTemplate\":\"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}\"},\"query-input\":\"required name=search_term_string\"}],\"inLanguage\":\"zh-Hans\"},{\"@type\":\"Person\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41\",\"name\":\"admin\",\"image\":{\"@type\":\"ImageObject\",\"inLanguage\":\"zh-Hans\",\"@id\":\"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/\",\"url\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"contentUrl\":\"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery\",\"caption\":\"admin\"},\"sameAs\":[\"http:\/\/wp.wsisp.com\"],\"url\":\"https:\/\/www.wsisp.com\/helps\/author\/admin\"}]}<\/script>\n<!-- \/ Yoast SEO plugin. -->","yoast_head_json":{"title":"FLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","robots":{"index":"index","follow":"follow","max-snippet":"max-snippet:-1","max-image-preview":"max-image-preview:large","max-video-preview":"max-video-preview:-1"},"canonical":"https:\/\/www.wsisp.com\/helps\/82374.html","og_locale":"zh_CN","og_type":"article","og_title":"FLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","og_description":"FLUX.1\u90e8\u7f72\u6559\u7a0b：ARM\u67b6\u6784\u670d\u52a1\u5668（\u5982NVIDIA Grace）\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1 1. \u524d\u8a00：\u5f53AI\u7ed8\u753b\u9047\u4e0aARM\u65b0\u8d35 \u6700\u8fd1，\u6211\u624b\u5934\u62ff\u5230\u4e86\u4e00\u53f0\u642d\u8f7dNVIDIA Grace CPU\u7684ARM\u67b6\u6784\u670d\u52a1\u5668\u3002\u8fd9\u673a\u5668\u6027\u80fd\u5f88\u5f3a，\u4f46\u6709\u4e2a\u95ee\u9898：\u5f88\u591aAI\u5e94\u7528\u90fd\u662f\u4e3ax86\u67b6\u6784\u8bbe\u8ba1\u7684，\u5728ARM\u4e0a\u8dd1\u8d77\u6765\u603b\u6709\u70b9\u6c34\u571f\u4e0d\u670d\u3002\u6b63\u597d，\u56e2\u961f\u91cc\u6709\u4e2a\u201c\u6d77\u666f\u7f8e\u5973\u56fe\u201d\u7684FLUX","og_url":"https:\/\/www.wsisp.com\/helps\/82374.html","og_site_name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","article_published_time":"2026-07-25T00:26:01+00:00","author":"admin","twitter_card":"summary_large_image","twitter_misc":{"\u4f5c\u8005":"admin","\u9884\u8ba1\u9605\u8bfb\u65f6\u95f4":"9 \u5206"},"schema":{"@context":"https:\/\/schema.org","@graph":[{"@type":"WebPage","@id":"https:\/\/www.wsisp.com\/helps\/82374.html","url":"https:\/\/www.wsisp.com\/helps\/82374.html","name":"FLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1 - \u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","isPartOf":{"@id":"https:\/\/www.wsisp.com\/helps\/#website"},"datePublished":"2026-07-25T00:26:01+00:00","dateModified":"2026-07-25T00:26:01+00:00","author":{"@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41"},"breadcrumb":{"@id":"https:\/\/www.wsisp.com\/helps\/82374.html#breadcrumb"},"inLanguage":"zh-Hans","potentialAction":[{"@type":"ReadAction","target":["https:\/\/www.wsisp.com\/helps\/82374.html"]}]},{"@type":"BreadcrumbList","@id":"https:\/\/www.wsisp.com\/helps\/82374.html#breadcrumb","itemListElement":[{"@type":"ListItem","position":1,"name":"\u9996\u9875","item":"https:\/\/www.wsisp.com\/helps"},{"@type":"ListItem","position":2,"name":"FLUX.1\u90e8\u7f72\u6559\u7a0b\uff1aARM\u67b6\u6784\u670d\u52a1\u5668\uff08\u5982NVIDIA Grace\uff09\u9002\u914d\u6d77\u666f\u56fe\u751f\u6210\u670d\u52a1\u53ef\u884c\u6027\u9a8c\u8bc1"}]},{"@type":"WebSite","@id":"https:\/\/www.wsisp.com\/helps\/#website","url":"https:\/\/www.wsisp.com\/helps\/","name":"\u7f51\u7855\u4e92\u8054\u5e2e\u52a9\u4e2d\u5fc3","description":"\u9999\u6e2f\u670d\u52a1\u5668_\u9999\u6e2f\u4e91\u670d\u52a1\u5668\u8d44\u8baf_\u670d\u52a1\u5668\u5e2e\u52a9\u6587\u6863_\u670d\u52a1\u5668\u6559\u7a0b","potentialAction":[{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https:\/\/www.wsisp.com\/helps\/?s={search_term_string}"},"query-input":"required name=search_term_string"}],"inLanguage":"zh-Hans"},{"@type":"Person","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/358e386c577a3ab51c4493330a20ad41","name":"admin","image":{"@type":"ImageObject","inLanguage":"zh-Hans","@id":"https:\/\/www.wsisp.com\/helps\/#\/schema\/person\/image\/","url":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","contentUrl":"https:\/\/gravatar.wp-china-yes.net\/avatar\/?s=96&d=mystery","caption":"admin"},"sameAs":["http:\/\/wp.wsisp.com"],"url":"https:\/\/www.wsisp.com\/helps\/author\/admin"}]}},"_links":{"self":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/82374","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/comments?post=82374"}],"version-history":[{"count":0,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/posts\/82374\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/media?parent=82374"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/categories?post=82374"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/tags?post=82374"},{"taxonomy":"topic","embeddable":true,"href":"https:\/\/www.wsisp.com\/helps\/wp-json\/wp\/v2\/topic?post=82374"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}