diff --git a/.gitignore b/.gitignore new file mode 100644 index 0000000..c4c4ffc --- /dev/null +++ b/.gitignore @@ -0,0 +1 @@ +*.zip diff --git a/026_0010.jpg b/026_0010.jpg new file mode 100755 index 0000000..5571278 Binary files /dev/null and b/026_0010.jpg differ diff --git a/Dockerfile_mlu370 b/Dockerfile_mlu370 new file mode 100644 index 0000000..1e10541 --- /dev/null +++ b/Dockerfile_mlu370 @@ -0,0 +1,20 @@ +FROM harbor.4pd.io/mic-llm-x/combricon-mlu370x4-base:v0.2.0-tgiv1.4.3-btv0.6.0-pt2.1-x86_64-ubuntu22.04-py310 + + +RUN if [ ! -d "/torch/venv3/pytorch" ]; then echo "虚拟环境路径缺失！"; exit 1; fi + + +WORKDIR /workspace/ +COPY ./model_test_caltech_http_mlu370.py /workspace/ +COPY ./microsoft_beit_base_patch16_224_pt22k_ft22k /model + + +# 安装transformers 4.46.3 +RUN pip install transformers==4.46.3 uvicorn\[standard\] fastapi flask==3.1.1 + +EXPOSE 80 + +ENTRYPOINT ["python3", "model_test_caltech_http_mlu370.py"] + + + diff --git a/README.md b/README.md index 7023b54..099b89d 100644 --- a/README.md +++ b/README.md @@ -1,2 +1,36 @@ -# enginex-mlu370-vc +# enginex-ascend-910-vc-cnn +运行于【寒武纪】系列算力卡的【视觉分类】引擎，基于 CNN 架构，支持 BEiT、MobileViT 等流行模型 + +## QuickStart + +1、从 modelscope上下载视觉分类的模型，例如 microsoft/beit-base-patch16-224 +```python +modelscope download --model microsoft/beit-base-patch16-224 README.md --local_dir /mnt/contest_ceph/zhoushasha/models/microsoft/beit_base_patch16_224_pt22k_ft22k +``` + +2、使用Dockerfile生成镜像 +从仓库的【软件包】栏目下载基础镜像 combricon-mlu370x4-base:v0.2.0-tgiv1.4.3-btv0.6.0-pt2.1-x86_64-ubuntu22.04-py310 +使用 Dockerfile_mlu370 生成 镜像 +```python +docker build -f Dockerfile_mlu370 -t combricon-mlu370x4-base-zhoushasha:v4.0 . +``` +注意 Dockerfile_mlu370 中已预先将模型 microsoft_beit_base_patch16_224_pt22k_ft22k 放在了 /model 下面 + +3、启动docker +```python +docker run -it \ + -p 10091:80 \ + --privileged \ + --name test_mlu370 \ + --device /dev/cambricon_ctl \ + --device /dev/cambricon[0-3] \ + -v /usr/bin/cnmon:/usr/bin/cnmon \ + combricon-mlu370x4-base-zhoushasha:v4.0 +``` + +4、测试服务 +```python +curl -X POST http://localhost:10091/v1/private/s782b4996 \ +> -F "image=@/home/zhoushasha/models/026_0010.jpg" +``` diff --git a/model_test_caltech_http_mlu370.py b/model_test_caltech_http_mlu370.py new file mode 100644 index 0000000..a79fe2f --- /dev/null +++ b/model_test_caltech_http_mlu370.py @@ -0,0 +1,378 @@ +import torch +import time +import os +from PIL import Image +from transformers import AutoImageProcessor, AutoModelForImageClassification +from flask import Flask, request, jsonify +from io import BytesIO + +# 引入寒武纪MLU相关模块 +try: + import torch_mlu + print(f"成功导入torch_mlu，版本: {getattr(torch_mlu, '__version__', 'unknown')}") + + # 简单的MLU可用性测试 + def check_mlu_available(): + try: + test_tensor = torch.randn(2, 2).mlu() + return True + except: + return False + + def get_mlu_device_count(): + """通过尝试多个设备索引来检测可用的MLU设备数量""" + max_devices_to_check = 8 + available_devices = 0 + + for i in range(max_devices_to_check): + try: + test_tensor = torch.randn(2, 2).mlu(i) + available_devices += 1 + print(f"MLU设备 {i} 可用") + except: + break + + return available_devices + + def get_device_name(device_index): + """获取设备名称""" + try: + return f"MLU-Device-{device_index}" + except: + return f"MLU-Device-{device_index} (Unknown)" + + # 创建模拟的ct模块 + class MLUModel: + @staticmethod + def is_mlu_available(): + return check_mlu_available() + + @staticmethod + def device_count(): + return get_mlu_device_count() + + @staticmethod + def get_device_name(device_index): + return get_device_name(device_index) + + ct = MLUModel() + MLU_AVAILABLE = check_mlu_available() + print(f"MLU设备可用: {MLU_AVAILABLE}") + print(f"检测到 {ct.device_count()} 个MLU设备") + +except ImportError: + torch_mlu = None + ct = None + MLU_AVAILABLE = False + print("警告: 未找到torch_mlu模块，无法使用MLU设备") +except Exception as e: + torch_mlu = None + ct = None + MLU_AVAILABLE = False + print(f"MLU初始化警告: {str(e)}") + +# 设置CPU核心数 +os.environ["OMP_NUM_THREADS"] = "4" +os.environ["MKL_NUM_THREADS"] = "4" +os.environ["NUMEXPR_NUM_THREADS"] = "4" +os.environ["OPENBLAS_NUM_THREADS"] = "4" +os.environ["VECLIB_MAXIMUM_THREADS"] = "4" +torch.set_num_threads(4) + +class MLUImageClassifier: + def __init__(self, model_path: str): + # 检测并使用MLU设备 + self.use_mlu = self._check_mlu_availability() + print(f"使用设备: {'MLU' if self.use_mlu else 'CPU'}") + + # 加载处理器和模型 + self.processor = AutoImageProcessor.from_pretrained(model_path) + self.model = self._load_model(model_path) + self.id2label = self.model.config.id2label + + # 验证模型设备 + self._verify_model_device() + + def _check_mlu_availability(self): + """检查MLU设备是否可用""" + if torch_mlu is None: + print("MLU不可用: torch_mlu模块未找到") + return False + + try: + # 测试MLU基本功能 + test_tensor = torch.randn(2, 2).mlu() + test_result = test_tensor + test_tensor + print("MLU设备可用性测试通过") + return True + except Exception as e: + print(f"MLU设备测试失败: {e}") + return False + + def _load_model(self, model_path: str) -> AutoModelForImageClassification: + """加载模型到合适的设备""" + try: + # 先在CPU加载模型 + model = AutoModelForImageClassification.from_pretrained( + model_path, + torch_dtype=torch.float32 + ) + + if self.use_mlu: + # 先将模型完全移动到CPU确保稳定 + model = model.cpu() + + # 使用.mlu()方法将模型移动到MLU设备 + model = model.mlu() + print("模型成功加载到MLU设备") + else: + model = model.cpu() + print("模型加载到CPU设备") + + return model.eval() + + except Exception as e: + print(f"模型加载失败: {str(e)}") + # 尝试fallback到CPU模式 + try: + model = model.cpu() + print("Fallback到CPU模式") + return model.eval() + except: + raise RuntimeError(f"模型加载完全失败: {str(e)}") + + def _verify_model_device(self): + """验证模型设备""" + try: + param = next(self.model.parameters()) + if self.use_mlu: + # 对于MLU设备，通过简单操作验证 + test_output = param + 0 + print("MLU模型验证成功") + else: + print("CPU模型验证成功") + + except StopIteration: + print("警告: 模型没有可训练参数") + except Exception as e: + print(f"模型验证警告: {e}") + + def _predict_with_mlu(self, image) -> dict: + """在MLU上执行推理""" + try: + start_time = time.perf_counter() + + # 预处理 + inputs = self.processor(images=image, return_tensors="pt") + + if self.use_mlu: + # 将输入数据移动到MLU + inputs_mlu = {} + for key, value in inputs.items(): + if hasattr(value, 'mlu'): + inputs_mlu[key] = value.mlu() + else: + inputs_mlu[key] = value + + # 执行推理 + with torch.no_grad(): + # 首次推理(热身) + ts = time.time() + outputs = self.model(**inputs_mlu) + #first_pass_time = time.time() - ts + print('mlu370 T1', time.time() - ts, flush=True) + + # 多次推理(性能测试) + ts = time.time() + #for _ in range(5): # 减少测试次数 + for i in range(800): + outputs = self.model(**inputs_mlu) + #batch_pass_time = time.time() - ts + print('mlu370 T2', time.time() - ts, flush=True) + else: + # CPU推理 + with torch.no_grad(): + ts = time.time() + outputs = self.model(**inputs) + #first_pass_time = time.time() - ts + print('cpu T1', time.time() - ts, flush=True) + + ts = time.time() + #for _ in range(5): + + outputs = self.model(**inputs) + #batch_pass_time = time.time() - ts + print('cpu T2', time.time() - ts, flush=True) + + + # 计算结果 + logits = outputs.logits + probs = torch.nn.functional.softmax(logits, dim=-1) + max_prob, max_idx = probs.max(dim=-1) + class_idx = max_idx.item() + + processing_time = round(time.perf_counter() - start_time, 6) + + return { + "class_id": class_idx, + "class_name": self.id2label.get(class_idx, f"class_{class_idx}"), + "confidence": float(max_prob.item()), + "device_used": "mlu" if self.use_mlu else "cpu", + "processing_time": processing_time + + } + except Exception as e: + return { + "class_id": -1, + "class_name": "error", + "confidence": 0.0, + "device_used": "mlu" if self.use_mlu else "cpu", + "processing_time": 0.0, + "error": str(e) + } + + def predict(self, image) -> dict: + """预测入口""" + return self._predict_with_mlu(image) + +# 初始化Flask应用 +app = Flask(__name__) + +# 全局模型加载 +try: + MODEL_PATH = os.environ.get("MODEL_PATH", "/model") + print(f"从路径加载模型: {MODEL_PATH}") + classifier = MLUImageClassifier(MODEL_PATH) + print("模型加载成功") +except Exception as e: + print(f"服务初始化失败: {str(e)}") + classifier = None + +@app.route('/v1/private/s782b4996', methods=['POST']) +def predict(): + """接收单张图片并返回GPU预测结果""" + if classifier is None: + return jsonify({ + "status": "error", + "prediction": { + "class_id": -1, + "class_name": "error", + "confidence": 0.0, + "device_used": "unknown", + "processing_time": 0.0, + "error": "服务未初始化成功" + } + }), 500 + + if 'image' not in request.files: + return jsonify({ + "status": "error", + "prediction": { + "class_id": -1, + "class_name": "error", + "confidence": 0.0, + "device_used": "mlu" if classifier.use_mlu else "cpu", + "processing_time": 0.0, + "error": "请求中未包含图片" + } + }), 400 + + try: + image_file = request.files['image'] + image = Image.open(BytesIO(image_file.read())).convert("RGB") + result = classifier.predict(image) + + if 'error' in result: + return jsonify({ + "status": "error", + "prediction": result + }), 500 + else: + return jsonify({ + "status": "success", + "prediction": result + }) + except Exception as e: + return jsonify({ + "status": "error", + "prediction": { + "class_id": -1, + "class_name": "error", + "confidence": 0.0, + "device_used": "mlu" if classifier and classifier.use_mlu else "cpu", + "processing_time": 0.0, + "error": f"处理图片失败: {str(e)}" + } + }), 500 + +@app.route('/health', methods=['GET']) +def health_check(): + """健康检查接口""" + mlu_available = False + mlu_info = {} + + if torch_mlu is not None and hasattr(ct, 'is_mlu_available'): + try: + mlu_available = ct.is_mlu_available() + mlu_info = { + "device_count": ct.device_count(), + "devices": [ct.get_device_name(i) for i in range(ct.device_count())] + } + except Exception as e: + mlu_info["error"] = str(e) + + return jsonify({ + "status": "healthy" if classifier is not None else "degraded", + "mlu_available": mlu_available, + "mlu_info": mlu_info, + "model_loaded": classifier is not None, + "using_mlu": classifier.use_mlu if classifier else False, + "timestamp": time.time() + }) + +@app.route('/test', methods=['GET']) +def test_mlu(): + """MLU测试接口""" + try: + if torch_mlu is None: + return jsonify({ + "status": "error", + "message": "torch_mlu模块未找到", + "mlu_working": False + }), 500 + + # 测试MLU基本功能 + test_tensor = torch.randn(3, 3).mlu() + result_tensor = test_tensor + test_tensor + result_cpu = result_tensor.cpu() + + return jsonify({ + "status": "success", + "message": "MLU测试通过", + "result_shape": str(result_cpu.shape), + "mlu_working": True + }) + except Exception as e: + return jsonify({ + "status": "error", + "message": f"MLU测试失败: {str(e)}", + "mlu_working": False + }), 500 + +@app.route('/info', methods=['GET']) +def device_info(): + """设备信息接口""" + info = { + "pytorch_version": torch.__version__, + "torch_mlu_available": torch_mlu is not None, + "mlu_devices_count": ct.device_count() if torch_mlu and hasattr(ct, 'device_count') else 0, + "model_loaded": classifier is not None, + "using_mlu": classifier.use_mlu if classifier else False, + "system_time": time.time() + } + return jsonify(info) + +if __name__ == "__main__": + # 启动HTTP服务 - 使用Flask内置服务器 + print("启动MLU图像分类服务...") + app.run(host='0.0.0.0', port=80, debug=False) \ No newline at end of file