fix: ACL bytes_to_ptr GC bug + 添加 NPU 模型转换指南

python/infer_server.py:
- 修复 acl.util.bytes_to_ptr() 后 bytes 被 GC 回收导致全零输出的 bug
- 保持 bytes 引用直到 memcpy 完成
- 正确解析 FP16 模型输出 (half precision -> float32)

docs/CONVERSION_GUIDE.md:
- 完整的 YOLOv8 ONNX->OM 转换流程文档
- ATC 转换参数详解及低内存设备配置
- 已知问题排查表 (OOM/全零输出/算子不支持)
- 其他 YOLOv8 尺寸模型参考

start.sh: 一键启动脚本

python/infer_server.py

@@ -112,10 +112,15 @@ class AclModel:
     def run(self, blob):
         """Push blob to NPU, execute, pull outputs back as numpy arrays."""
         in_ds = acl.mdl.create_dataset()
+        # Host input buffer - MUST keep reference alive during memcpy
+        host_buf = np.ascontiguousarray(blob)
+        host_bytes = host_buf.tobytes()
+        host_addr = acl.util.bytes_to_ptr(host_bytes)
+        # Device input buffer
         in_buf, ret = acl.rt.malloc(self.input_size, ACL_MEM_MALLOC_NORMAL_ONLY)
         assert ret == 0
         ret = acl.rt.memcpy(in_buf, self.input_size,
-                            blob.tobytes(), self.input_size,
+                            host_addr, self.input_size,
                             ACL_MEMCPY_HOST_TO_DEVICE)
         assert ret == 0
         db = acl.create_data_buffer(in_buf, self.input_size)
@@ -135,11 +140,27 @@ class AclModel:
         for i in range(self.output_num):
             sz   = self.out_sizes[i]
             host = np.zeros(sz, dtype=np.uint8)
-            ret  = acl.rt.memcpy(host.ctypes.data, sz,
+            # Keep reference alive during memcpy
+            host_bytes = host.tobytes()
+            host_addr = acl.util.bytes_to_ptr(host_bytes)
+            ret  = acl.rt.memcpy(host_addr, sz,
                                  self.out_bufs[i], sz,
                                  ACL_MEMCPY_DEVICE_TO_HOST)
             assert ret == 0
-            outputs.append(host.view(np.float32).reshape(self.output_shapes[i]))
+            # Read back from host_bytes (memcpy modified it)
+            host_from_bytes = np.frombuffer(host_bytes, dtype=np.uint8)
+            # FP16 model outputs are half precision (2 bytes)
+            num_fp16 = sz // 2
+            num_fp32 = sz // 4
+            expected = self.output_shapes[i][0] * self.output_shapes[i][1] * self.output_shapes[i][2]
+            if num_fp16 == expected:
+                out_fp16 = host_from_bytes.view(np.float16)
+                outputs.append(out_fp16.astype(np.float32).reshape(self.output_shapes[i]))
+            elif num_fp32 == expected:
+                outputs.append(host_from_bytes.view(np.float32).reshape(self.output_shapes[i]))
+            else:
+                out_fp16 = host_from_bytes.view(np.float16)
+                outputs.append(out_fp16.astype(np.float32).reshape(self.output_shapes[i]))
 
         acl.rt.free(in_buf)
         acl.mdl.destroy_dataset(in_ds)