Team Red LLM：AMD GPU 本地大模型推理的实战指南与基准数据库

Team Red LLM is a community-maintained project focusing on AMD GPU local LLM deployment. It provides detailed ROCm/HIP inference steps, common pitfalls, real performance benchmarks, and supports consumer Radeon, data center Instinct, and Strix Halo APU. The project aims to help AMD users avoid ROCm-related pitfalls and improve their local AI experience.

The current LLM open-source ecosystem is CUDA-centric—most tutorials assume NVIDIA GPUs, toolchains often fallback to CPU on ROCm detection, and solutions are scattered in old Reddit posts. AMD users face a 'second-class citizen' experience. Team Red LLM was created as a community cookbook and benchmark database to solve these issues, allowing developers to share ROCm pitfalls and help others.

The project uses a modular structure:

• COOKBOOK.md: Step-by-step deployment guide with verified steps and marked pitfalls.
• benchmarks/results.csv: Community-contributed token generation speed data (not vendor marketing).
• hardware/: Per-GPU docs (BIOS, drivers,散热, best models).
• models/: Model-specific configs (launch params, problematic quant formats, MoE offload tips).
• scripts/: Wrapper scripts for llama-server, model switching, and benchmarking.

The project provides RX7900 GRE 16GB performance data:

GPU	架构	模型	量化	模式	生成 tok/s	提示 tok/s
RX7900 GRE	gfx1100	Moonlight-16B-A3B-Instruct	Q6_K	Full GPU	100.2	188.1
RX7900 GRE	gfx1100	gemma-4-26B-A4B-it	UD-Q4_K_M	MoE offload (-ncmoe6)	31.0	61.3
RX7900 GRE	gfx1100	Qwen3.6-35B-A3B-UD	Q4_K_S	MoE offload (-ncmoe32)	22.7	41.7
RX7900 GRE	gfx1100	gemma-4-26B-A4B-it	UD-Q6_K	MoE offload (-ncmoe16)	17.3	80.7

Modes: Full GPU (max speed, VRAM-limited) vs MoE offload (runs larger models but 3-5x slower due to DDR5 bandwidth).

The project maintains a support matrix:

Code	架构家族	示例型号	ROCm 支持状态
gfx1100	RDNA3	RX7900XTX/XT/GRE	✅成熟
gfx1101	RDNA3	RX7800XT/7700XT	✅可用
gfx1102	RDNA3	RX7600	⚠️部分支持
gfx1200/1201	RDNA4	RX9070XT/9060	✅近期支持
gfx1150/1151	Strix Halo	Ryzen AI Max+395	⚠️前沿支持
gfx942/950	CDNA3	MI300X/MI325X	✅数据中心

Strix Halo APU needs manual patches or specific kernel versions for ROCm support.

Contribute via:

• Benchmarks: Submit PR to results.csv or use GitHub Issue template.
• Pitfalls: Add to COOKBOOK.md.
• Models: Create model-specific docs in models/.
• Hardware: Add GPU docs in hardware/.

GitHub Discussions for general topics (hardware suggestions, 'worth it' questions); Issues for bugs.

Team Red LLM proves AMD hardware can handle local LLM inference via community collaboration. It lowers entry barriers for budget users (e.g., used RX6000/7000 series) and reduces AMD users' reliance on cloud APIs. The project is MIT-licensed, encouraging free use and modification.