# tyllm

**Repository Path**: Intellifusion_com/tyllm

## Basic Information

- **Project Name**: tyllm
- **Description**: A high-performance LLM library on Intellifusion Edgex processor.
- **Primary Language**: Unknown
- **License**: Apache-2.0
- **Default Branch**: master
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 3
- **Forks**: 0
- **Created**: 2025-02-26
- **Last Updated**: 2026-07-06

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# TYLLM

TYLLM is a high-performance operator library designed to accelerate LLM inference on Intellifusion Edgex processor through optimized kernels written in TYLang, a kernel-level programming language derives from TVMScript.

Currently supported:
- MLA(Multi-head Latent Attention)
- w4a16 dense
- MHA/GQA/ with flash attention
- Cache-aware operations (KV cache management) 
- Commonly used Collective Communication Library
  
## Quick Start

### Test
```bash
python tests/test_mla.py
```

## Requirements
- Edgex hardware
- TVM with Edgex support
- PyTorch >= 2.4.0

## Acknowledgement

MLA is inspired by [FlashMLA](https://github.com/deepseek-ai/FlashMLA)