存算一体·后摩智能科技有限公司|劲能芯·极效行

HOUMO.AI Software Platform

Combining the proprietary Houmo Dadao™ Software Stack with Houmo Open-Source AI Ecosystem Software, the platform delivers full-stack, end-to-end support from chip hardware to upper-layer applications. It supports multiple generations of heterogeneous hardware while seamlessly integrating with mainstream deep learning frameworks and LLM inference ecosystems. A complete suite of C++ and Python APIs streamlines development and simplifies deployment.

Key Advantages

Full-Stack Compatibility

Deeply optimized for mainstream LLM inference frameworks such as vLLM, with ModelZoo and domain-specific suites for seamless ecosystem integration and cross-platform migration.
Efficient Compilation

Combines a static compiler, automated search-based optimization and a high-performance operator library to deliver maximum performance without complex configuration or tuning.
Effortless Deployment

Supports native floating-point computation, eliminating complex accuracy-tuning steps, while professional quantization tools enable flexible low-bit, high-throughput deployment.
Comprehensive Tooling

Integrates visual analysis and SMI system management tools, providing full lifecycle support across development, quantization, compilation and runtime.

Platform Architecture

Solutions

Intelligent Perception & Analytics

Real-time perception and analysis of environments and behaviors based on multimodal data.

AI Computing & Performance

Heterogeneous computing and inference optimization for edge AI, improving model inference efficiency and system performance.

Embodied Intelligence

Integrating perception, reasoning and motion control to deliver edge AI solutions for robotic grasping and embodied intelligence scenarios.

Ecosystem & Certification

Building an open partner ecosystem with comprehensive adaptation and certification support.

Others

Edge AI solutions covering more scenarios for diverse application requirements.

Application Enablement

EcoSystem Domain Suites

End-to-end deployment solutions for LLMs, speech, image and video generation, integrating model configuration, pre/post-processing and optimization parameters.

PostMo-ModelZoo

A collection of vision, language, LLM and multimodal models, validated and optimized by the Houmo toolchain for efficient inference on Houmo AI chips.

Application Examples

Complete runnable applications demonstrating Houmo AI chip capabilities, including chat interfaces for quick onboarding.

Application Tools

A collection of helper tools providing abstract API interfaces and basic development utilities for domain suites, model zoo and application examples.

vLLM

A high-throughput LLM inference and serving engine supporting LLM application deployment on Houmo AI chips.

SGLang

A high-performance inference and serving framework for LLMs and multimodal models, supporting chat, Agent and RAG application deployment.

Llama.cpp

A lightweight C/C++ implementation of LLM inference, supporting multiple backends such as CPU, CUDA, Metal and Vulkan.

...

Other Inference Frameworks

Additional inference framework adaptations to meet diverse model inference and deployment requirements.

Toolchain

PostMo-Quant

Supports PTQ and Weight-Only quantization, converting floating-point models into low-precision models such as MXINT8 with accuracy validation.

PostMo-Sight

Provides operator-level detailed analysis and system-level full-stack performance analysis, collecting execution time, bandwidth utilization and other metrics to help locate performance bottlenecks.

PostMo-Bench

Provides CV, NLP, LLM and AIGC benchmark suites evaluating latency, throughput, memory and power, with single/multi-card and competitor comparisons.

PostMo-Toolkit

Provides environment installation, component detection and deployment assistance tools, simplifying development and inference environment configuration.

PostMo-SMI

Monitors device temperature, power, frequency, memory, process usage and topology in real time, and supports device reset and power management.

Hardware Enablement

PostMo-Compiler

Converts models into chip-executable files and improves execution efficiency through graph optimization, operator optimization and memory planning.

PostMo-Runtime

Handles model loading and execution, task scheduling, device memory management, data transfer and runtime status synchronization.

Firmware

Manages chip initialization, power and clock control, interrupt handling and exception recovery, providing low-level hardware control for drivers.

Drivers

Provides Houmo AI chip driver capabilities for operating systems, supporting device management, data transfer and hardware access interfaces for the runtime.