RUM-F2K-4B 1296000 GGUF

ComfyUI-RUM GGUF 节点仓库 · patched llama-quantize 工具 · English

中文模型卡

这是 RUM-FLUX.2-klein-4B-preview 的非官方 GGUF 量化版本，对应原始 checkpoint：

model-checkpoint-1296000.safetensors

本仓库目前提供两个量化版本：

文件	建议用途
`rum-1296000-rum-native-Q4_K_M.gguf`	推荐日常使用，体积较小，显存压力较低
`rum-1296000-rum-native-Q8_0.gguf`	更接近原始权重，适合质量对比和验证转换正确性

本 GGUF 转换版本已获得原作者允许上传。
这不是 RUM 官方仓库，也不包含 Qwen、WAI teacher CLIP、VAE 或原始 safetensors checkpoint。

重要说明

这个文件不是普通 FLUX.2 Klein 4B GGUF。

RUM 主模型需要保留两路文本投影：

context_embedder.weight
context_embedder_2.weight

普通 FLUX/Klein GGUF 转换器可能只保留或合并为：

txt_in.weight

这样会丢掉 RUM 需要的 teacher CLIP extra projection。
如果缺少 context_embedder_2.weight，该 GGUF 就不是完整 RUM 主模型，不能安全替代原始 RUM checkpoint。

本仓库中的 GGUF 使用 RUM 专用转换器生成，并保留了完整的 RUM 双文本投影结构。

文件说明

文件	说明
`rum-1296000-rum-native-Q4_K_M.gguf`	Q4_K_M 量化，推荐日常使用
`rum-1296000-rum-native-Q8_0.gguf`	Q8_0 量化，更接近原始 checkpoint
`quantization_config.json`	转换和量化信息
`assets/preview.jpg`	预览图

需要的 ComfyUI 节点

请使用支持 RUM-native GGUF 的 ComfyUI-RUM fork：

https://github.com/shira-qwq/ComfyUI-RUM-gguf

需要的节点：

RUM FLUX.2 Load Native Model GGUF

不要用普通 FLUX GGUF loader 加载这个文件。

转换与量化信息

转换流程：

model-checkpoint-1296000.safetensors
→ rum-1296000-rum-native-BF16.gguf
→ rum-1296000-rum-native-Q4_K_M.gguf
→ rum-1296000-rum-native-Q8_0.gguf

转换器：

ComfyUI-RUM/tools/convert_rum_to_gguf.py

量化工具：

patched llama-quantize

patched 工具 Release：

https://github.com/shira-qwq/ComfyUI-RUM-gguf/releases/tag/rum-gguf-tools-v0.1

普通 llama.cpp 预编译版 llama-quantize.exe 可能会报：

unknown model architecture: flux

需要使用基于 ComfyUI-GGUF/tools/lcpp.patch 编译的 patched llama-quantize。

已知限制

这是非官方实验性 GGUF 发布
Q4_K_M 输出可能和原始 BF16 / safetensors checkpoint 有差异
Q8_0 通常更接近原始 checkpoint，但文件更大
暂不建议量化 teacher CLIP
Qwen GGUF 支持仍然是实验性质，建议先使用 qwen_3_4b.safetensors 验证 baseline
本仓库不包含 Qwen、WAI teacher CLIP、VAE 或原始 RUM safetensors 权重
自定义 workflow 需要保留 RUM 的 200 + 77 token conditioning 路线

授权和归属说明

原始模型和项目：

RUM by RimoChan
ComfyUI-RUM by peter119lee
GGUF / 量化工具基于 ComfyUI-GGUF 和 llama.cpp

本 GGUF 转换版本已获得 RUM 原作者允许上传。

本仓库暂时标记为 license: other，因为具体使用仍应参考原模型、原项目和相关组件的上游说明。

请遵守原作者和所有依赖组件的许可要求。

English Version

This repository contains unofficial GGUF quantizations of RUM-FLUX.2-klein-4B-preview, based on:

model-checkpoint-1296000.safetensors

Two quantized files are provided:

File	Suggested use
`rum-1296000-rum-native-Q4_K_M.gguf`	Recommended daily version, smaller and more VRAM-friendly
`rum-1296000-rum-native-Q8_0.gguf`	Closer to the original checkpoint, useful for quality comparison

This GGUF release was uploaded with permission from the original RUM author.
This is not the official RUM repository.

Important Notice

This is not a normal FLUX.2 Klein 4B GGUF.

RUM requires two text projection tensors:

context_embedder.weight
context_embedder_2.weight

A normal FLUX/Klein GGUF converter may collapse the text projection into txt_in.weight and lose the required teacher CLIP extra projection.

If context_embedder_2.weight is missing, the GGUF is not a complete RUM main model.

This GGUF was created with a RUM-specific converter and preserves the required dual text projection.

Required ComfyUI Node

Use this file with the ComfyUI-RUM GGUF fork:

https://github.com/shira-qwq/ComfyUI-RUM-gguf

Required node:

RUM FLUX.2 Load Native Model GGUF

Do not load this file with a normal FLUX GGUF loader.

Recommended Setup

Place the main model in:

ComfyUI/models/diffusion_models/

Recommended companion files:

qwen_3_4b.safetensors
waiIllustriousSDXL_v140_clip_l.safetensors
waiIllustriousSDXL_v140_clip_g.safetensors
flux2-vae.safetensors

Recommended settings:

base_text_tokens = 200
extra_text_tokens = 77
qwen_layers = 10,20,30

Use the RUM Diffusers Noise / Scheduler / CFGuider workflow path. Do not replace it with a normal KSampler workflow.

Quantization

Converter:

ComfyUI-RUM/tools/convert_rum_to_gguf.py

Quantizer:

patched llama-quantize

Tool release:

https://github.com/shira-qwq/ComfyUI-RUM-gguf/releases/tag/rum-gguf-tools-v0.1

A stock llama.cpp llama-quantize.exe may fail with:

unknown model architecture: flux

A patched llama-quantize built with ComfyUI-GGUF/tools/lcpp.patch is required for Flux/RUM image-model GGUF quantization.

Known Limitations

Unofficial and experimental
Q4_K_M may differ from the original BF16 / safetensors checkpoint
Q8_0 is usually closer to the original checkpoint but larger
Teacher CLIP GGUF is not recommended at this stage
Qwen GGUF support is experimental
This repository does not include Qwen, WAI teacher CLIP, VAE, or original RUM safetensors weights
Custom workflows should preserve the RUM 200 + 77 token conditioning path

Attribution and License Note

Original model and project:

RUM by RimoChan
ComfyUI-RUM by peter119lee
GGUF / quantization tooling based on ComfyUI-GGUF and llama.cpp

The uploader received permission from the original RUM author to upload these GGUF quantizations.

This repository is marked as license: other because users should still check the original project and model pages for upstream terms.

Please follow the original author's terms and the licenses of all required components.

Downloads last month: 23

GGUF

Model size

4B params

Architecture

flux

Hardware compatibility

4-bit

8-bit

Model tree for shira-qwq/RUM-F2K-4B-1296000-GGUF

Base model

rimochan/RUM-FLUX.2-klein-4B-preview

Quantized

(2)

this model

shira-qwq
/

RUM-F2K-4B-1296000-GGUF