[Roadmap] NPU Support Roadmap

[Kirrito-k423]

📅 [Roadmap] NPU Support Roadmap

💡 Overview

This roadmap outlines the strategic plan for NPU (Neural Processing Unit) ecosystem support within the VeOmni framework, targeting Ascend C and Triton backends. The initiative aims to deliver comprehensive NPU coverage for mainstream open-source models in both inference and training scenarios, spanning language, vision-language, and video generation domains.

---

🎯 1. Core Infrastructure

• Ascend C Kernel Adaptation: Implement custom operator kernels in Ascend C dialect, achieving hardware-level operator fusion for prevalent model architectures.
• Triton Backend Integration: Construct a high-performance Triton-Ascend C backend with standardized operator registration and code generation pipelines.

---

🤖 2. Open-Source Model Support

🔥 In Progress

Model Family	Scope	Description
Qwen3.5 Series	Full support	Complete coverage of Qwen3.5
Qwen Image Model (Multimodal)	Full support	Support for Qwen Image
Wan2.2 Model	Full support	Adapter for Wan 2.2

📋 Planned

Model Family	Status	Description
Bagel Model	Planned	Multilingual / multimodal model support to further expand NPU ecosystem coverage.

---

🔧 3. Performance & Tooling

• NPU Profiling Toolchain: Integrate NPUProfiler for operator-level performance analysis and bottleneck identification.
• Standardized Benchmark Suite: Publish reproducible performance baselines for major models on NPU hardware (throughput, latency, memory utilization).
• Documentation & Tutorials: End-to-end deployment guides from HuggingFace / ModelScope checkpoints to NPU inference.

---

📌 Milestone Overview

Milestone	Objective	Target Quarter
M1	Ascend C + Triton Backend foundational adaptation	Early Q3
M2	Qwen3.5 & Qwen Image Model NPU support	Early Q3
M3	Wan2.2 video generation NPU support	Q3
M4	Bagel model support + comprehensive performance optimization	Q3

---

📎 References

• VeOmni Q2 2026 Roadmap: #669
• Ascend C Documentation: https://www.hiascend.com/
• Triton Compiler: https://openai.github.io/triton/

---

This roadmap is proposed for community discussion. Contributions and feedback are welcome.