cafonez commited on
Commit
1146119
·
verified ·
1 Parent(s): 714d033

Update README.md

Browse files

<div align="center">

# **Nemotron 3 30B A3B**
**ROCmFPX Agentic Quant**

[![GGUF](https://img.shields.io/badge/GGUF-ROCmFPX-blue?style=for-the-badge)](https://github.com/ggerganov/llama.cpp)
[![ROCm](https://img.shields.io/badge/AMD-ROCm-red?style=for-the-badge)](https://rocm.docs.amd.com/)
[![Vulkan](https://img.shields.io/badge/Vulkan-Supported-9F7CFF?style=for-the-badge)](https://www.vulkan.org/)

<div align="center" style="line-height: 1;">
<a href="https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/" style="margin: 2px;">
<img alt="License" src="https://img.shields.io/badge/License-NVIDIA Open Model License-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/>
</a>
</div>

![c0f43f9f-4a7b-47df-a0a5-ac9b79fb68ec](https://cdn-uploads.huggingface.co/production/uploads/66c3d6e296583c59b06a1dd3/NUilCo9URWpy9-q5sPwy6.png)

**Experimental ROCmFPX quantized GGUF** optimized for **AMD ROCm & Vulkan** hardware with enhanced agentic & structured output performance.
ROCmFPX Build
Use the official ROCmFPX-enabled llama.cpp fork:
🔗 https://github.com/charlie12345/ROCmFPX
This fork includes full ROCmFPX quant support, agentic presets, and Dynamic Drafting.
</div>

---

## Model Description

This repository contains an **experimental ROCmFPX quantized GGUF** build of **NVIDIA Nemotron 3 30B A3B**, prepared for llama.cpp-compatible runtimes with strong AMD ROCm and Vulkan support.

> This is **not** the original BF16 model. It is a specialized quantized derivative designed to dramatically reduce model size while preserving high practical coherency for chat, coding, JSON output, tool-calling, and agentic workflows.

### Key Features

- **ROCmFPX Quantization**: AMD-oriented FP-style low-bit block quantization with native ROCm/Vulkan kernel support and intelligent tensor-aware routing.
- **Agentic Preset**: Higher-risk tensors are preserved more carefully than standard quants, improving quality on:
- Structured JSON responses
- Tool / function calling
- Coding & multi-step reasoning
- Agent workflows (Hermes, OpenClaw, opencode-style agents, etc.)
- **High Context Support**: Excellent compatibility with extended context (tested up to 65k+ tokens)
- **Hardware Optimized**: Built for AMD Strix Halo, ROCm, and Vulkan backends

---

## Intended Use

Best suited for **local inference** on AMD hardware using ROCmFPX-enabled llama.cpp builds.

**Recommended applications:**
- Local chat assistants
- Advanced coding assistants
- Tool-calling & agentic systems
- Structured data generation (JSON)
- Multi-agent workflows (LangGraph, CrewAI, etc.)
- Research & testing on ROCm / Vulkan

⚠️ **Experimental format** — evaluate thoroughly before production use.

---

## Quick Start

### ROCm Backend
```bash
./llama-server \
-m nemotron-3-30b-a3b-rocmfp6-agent.gguf \
-ngl 999 \
-dev ROCm0 \
-c 65536 \
-fa on \
--cache-type-k q8_0 \
--cache-type-v q8_0

---

### Vulkan Backend

./llama-server \
-m nemotron-3-30b-a3b-rocmfp6-agent.gguf \
-ngl 999 \
-dev Vulkan0 \
-c 65536 \
-fa on \
--cache-type-k q8_0 \
--cache-type-v q8_0

---

Full ROCmFPX quant support
Agentic quantization presets
Optimized ROCm & Vulkan paths
Dynamic Drafting & speculative decoding support


Performance Notes

Excellent prompt processing and decode speeds on AMD hardware
Strong preservation of reasoning and structured output capabilities
Perplexity and downstream benchmarks available upon request


Built with ❤️ for the AMD open-source LLM community

Files changed (1) hide show
  1. README.md +21 -1
README.md CHANGED
@@ -3,4 +3,24 @@ license: other
3
  license_name: nvidia-nemotron-open-model-license
4
  license_link: >-
5
  https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/
6
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3
  license_name: nvidia-nemotron-open-model-license
4
  license_link: >-
5
  https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/
6
+ base_model:
7
+ - nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
8
+ pipeline_tag: text-generation
9
+ tags:
10
+ - nemotron
11
+ - nemotron_h
12
+ - mamba2
13
+ - quantization
14
+ - gguf
15
+ - llama.cpp
16
+ - agentic-quant
17
+ - amd
18
+ - json
19
+ - coding
20
+ - tools
21
+ - tool-calling
22
+ - nemotron-3-nano-amd
23
+ - rocmfpx
24
+ - rocmfp6
25
+ - rocmfp6-agent
26
+ ---