Index Translate models for ComfyUI
中文
新增完整 2B CONVROT INT8 与 9B CONVROT INT8，权重约 2.67 / 10.37 GB。量化来自上述固定官方源版本，保存逐文件 SHA256、Hadamard 旋转分组、逐通道 scale 与浮点保留层；详细 recipe 在各目录的 index-quantization.json。这是 Comfy CONVROT INT8 布局，不是 GGUF Q8 / NF4，也不是只改文件名。原始模型卡保留，量化和运行时适配由本社区项目完成。
INT8 文件夹下载后无需改名。ComfyUI 用户原样放入 ComfyUI/models/index_translate/，用自己的 Python 安装节点 install_convrot.py（Windows）或 requirements-convrot.txt，节点精度选择 auto / convrot-int8，model_path 指向完整目录。Windows 0.1.17 完整整合包内置 2B INT8，可在模型页「选择文件夹…」选择任意盘符的 2B / 9B INT8，保存后「加载并预热」。INT8 本版要求 NVIDIA CUDA；CPU 用官方目录。两个产品分别推理，节点不连接整合包。
直接下载到正确目录：hf download t8star/Index-Translate-Comfy --include "Index-Translate-2B-ConvRot-INT8/*" --local-dir "ComfyUI/models/index_translate"；9B 将 2B 改为 9B。请下载整个文件夹，不能缺分片、配置、分词器或量化清单。本机 Windows 常驻服务的 2B 首次编译约 40–60 秒，9B 首次编译约 116 秒、加载另需约 20 秒；2B 预热后六类样例 12 次生成中位数约 0.297 秒；这是小样本实测，不是所有 GPU 的保证，也不是独立 ComfyUI 工作流加载耗时。
本仓库为 ComfyUI Index Translate · T8star-Aix 提供完整的 2B 和 9B 模型文件，均来自 IndexTeam 官方模型。2B 固定来源修订版 47682cf7adbb20e7834c2e43960b1b85c369713a；9B 固定来源修订版 bd6d8a91199a2a72c009fbd01d575066249d89ba。原始模型卡分别保留在各目录的 README.md。这是社区镜像与节点集成，模型版权归原作者。
下载完整的 Index-Translate-2B 或 Index-Translate-9B 文件夹，原样放入 ComfyUI/models/index_translate/ 即可，无需重命名。也可使用任意完整模型目录，在节点 model_path 填入绝对路径。请勿只下载权重文件：配置、分词器及预处理文件也必须完整。2B 权重约 4.55 GB，9B 分片合计约 19.31 GB。GitHub Release 仅含代码，不含权重或 Python。35B preview 源仓库分片不完整，未镜像。
命令行可直接下载到正确目录：hf download t8star/Index-Translate-Comfy --include "Index-Translate-2B/*" --local-dir "ComfyUI/models/index_translate"。9B 用户把命令中的 2B 改为 9B 即可。
English
Complete 2B and 9B CONVROT INT8 directories are also available (approximately 2.67 / 10.37 GB of weights). They derive from the same pinned official revisions. Each index-quantization.json records source hashes, output hashes, row scales, Hadamard rotation groups, and floating-point layers. Original model cards remain unchanged. This is the Comfy ConvRot layout, not GGUF Q8 or NF4.
Keep the directory names as downloaded. Place them under ComfyUI/models/index_translate/, install install_convrot.py with ComfyUI's own Python on Windows (or requirements-convrot.txt on other systems), and select auto / convrot-int8 plus the complete directory in model_path. The independent Windows 0.1.17 full bundle includes 2B INT8 and supports selecting any complete 2B/9B directory via its folder picker; save and warm up before viewing. INT8 currently requires NVIDIA CUDA. Use the original model directory for CPU. ComfyUI never calls the standalone service.
Download directly with hf download t8star/Index-Translate-Comfy --include "Index-Translate-2B-ConvRot-INT8/*" --local-dir "ComfyUI/models/index_translate". Replace 2B with 9B as needed. All shards, configuration, tokenizer, and quantization manifest are required. On the tested Windows persistent service, first compilation takes roughly 40–60 seconds for 2B and 116 seconds for 9B, with another 20 seconds for 9B model loading; the median of 12 warmed 2B generations across six examples was 0.297 seconds. This small local benchmark is not a universal GPU or ComfyUI workflow latency guarantee.
Complete 2B and 9B weights for ComfyUI Index Translate · T8star-Aix, copied from the official IndexTeam models. The pinned upstream revisions are 47682cf7adbb20e7834c2e43960b1b85c369713a for 2B and bd6d8a91199a2a72c009fbd01d575066249d89ba for 9B. Original model cards remain in each directory. This is a community mirror and node integration; the original authors retain model credit.
Download the entire Index-Translate-2B or Index-Translate-9B directory and place it as-is under ComfyUI/models/index_translate/; no rename is needed. Alternatively, enter any complete model directory as an absolute model_path in the node. Configuration, tokenizer, and preprocessor files are required alongside the weights. GitHub Releases contain code only, without weights or Python. The incomplete upstream 35B preview is not mirrored.
For a direct download into the correct location, run hf download t8star/Index-Translate-Comfy --include "Index-Translate-2B/*" --local-dir "ComfyUI/models/index_translate". Replace 2B with 9B for the larger model.
By T8star-Aix · Bilibili · YouTube · API · Free gallery · Online AI apps · ComfyUI bundle · Hugging Face