Orange Pi Zero3 1G RAM Allwinner H618 64-bit USB2.0 Dual Band WiFi 5 BT 5.0 Gigabit LAN Port Mini PC Open Source Single Board $13.00 - $16.00 Min. order: 6 pieces STM32F407VET6 Development Board Cortex-M4 STM32 Minimum System Learning Board ARM Core Board +3.2 Inch LCD TFT With To...
6B的参数量虽然能够做到本地部署,但是目前的实现依赖库较多,如Pytorch, transfomer;对于端侧部署来说要求仍然较高。因此我们尝试将该模型转换为MNN模型,极大降低了部署时的依赖项,能够更方便的在各类端侧设备上部署与测试;同时我们对MNN模型进行了低bit量化,并实现了反量化与计算融合的计算kernel,大大降低了内存需求。
Small download size The size of eBlock is < 30MB Multiple devices Not only Arduino based boards now you can code micro:bit, SAMD51, NRF5, STM32, ESP8266, ESP32 ... Full device customization Each device can have one or more firmware to flash, it's own drivers to install or custom...
NanoPi NEO Air 512 mb RAM WIFI & BT, 8 gb/32 gb eMMC Allwinner H3 Quad-core Cortex-A7 $26.00 - $35.00 Min. order: 5 pieces Orange Pi 5 Rockchip RK3588S 8-core 64-bit processor NPU 4G 8G OPI 5 $70.00 - $94.00 Min. order: 1 piece Orange Pi R1 Plus LTS 1GB RAM, Uses Rock...
因此我们尝试将该模型转换为MNN模型,极大降低了部署时的依赖项,能够更方便的在各类端侧设备上部署与测试;同时我们对MNN模型进行了低bit量化,并实现了反量化与计算融合的计算kernel,大大降低了内存需求。实测PC端小显存显卡能够成流畅运行浮点模型,在Android手机上能够流畅运行量化模型。