bschooled/Qwen-Image-2.1-text-encoder-quantized

Qwen-Image-2.1 text encoder — NF4 pre-quantized

This repository is a derived artifact, not an upstream release. It publishes the text_encoder/ component of Qwen/Qwen-Image-2.1 as a bitsandbytes NF4 4-bit double-quantized checkpoint, so a Diffusers pipeline can load a ~6 GB encoder instead of the 17.5 GB BF16 shards.

Provenance

• Source repository: Qwen/Qwen-Image-2.1
• Source revision: d26bb61231c349cf6b7896fa83353113880e1ba3
• Source subfolder: text_encoder

Source shard	Bytes	SHA-256
model-00001-of-00004.safetensors	4998056552	dde00291b5f7fb92013895310a3da0ddba78674df9f10d505d375243dc01fc6f
model-00002-of-00004.safetensors	4915962464	9047faccc0a6d98496a52d55f27be1c94a9c259d1e283fbea0128d054a948d42
model-00003-of-00004.safetensors	4915962496	8c54187654c0176b73ae73785bf791dc9a14c9df7fb4310083a09d42048cb57e
model-00004-of-00004.safetensors	2704357976	5311532aaaeae3259eb6a7b2c600636be1159adf7ded35f53579f7d0e7d43cdd

Each shard was loaded with the quantization config below on a CUDA device (device_map={"": "cuda"}, torch_dtype=bfloat16) and saved with transformers save_pretrained. The quantization_config in config.json is what makes a plain from_pretrained load this checkpoint already quantized.

Quantization

{
  "load_in_4bit": true,
  "bnb_4bit_quant_type": "nf4",
  "bnb_4bit_use_double_quant": true,
  "bnb_4bit_compute_dtype": "bfloat16",
  "bnb_4bit_quant_storage": "uint8"
}

Contents

File	Bytes	SHA-256
config.json	2028	0138d076165297a6b734b2d8ec1172c44ea489a2ebb4835411b79c8d2a5d5516
generation_config.json	212	4a291bf827c4ac8c7dbf54c43dfd31fde5b0472dc5f41e2568a62b01098667ad
model-00001-of-00002.safetensors	4978393565	2eb23adb94a2941320dfe60f0f3b26f2fd0ac22fc603f4d21e4d173f512c99b2
model-00002-of-00002.safetensors	1400036772	fb15ce4eae281106dd46988ee0fb5ccff24423e6ed96f075a329cce29815e92d
model.safetensors.index.json	267233	efe1b24791750f34911a277a8e3301acb563c0e500c3fe842474c484b8e09f34
added_tokens.json	707	c0284b582e14987fbd3d5a2cb2bd139084371ed9acbae488829a1c900833c680
chat_template.jinja	5292	3636d0f0bd6bef02654cdffdc447b79cb2cef8ab02cc75267345946291a489e4
merges.txt	1671853	8831e4f1a044471340f7c0a83d7bd71306a5b867e95fd870f74d0c5308a904d5
preprocessor_config.json	782	93585062a80db5e8ca038efc7726a3e6411d9db948472d81d63c6303993be8c5
special_tokens_map.json	613	76862e765266b85aa9459767e33cbaf13970f327a0e88d1c65846c2ddd3a1ecd
tokenizer.json	11422654	aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
tokenizer_config.json	5445	81ec7bb9530159b326c0bef1d0b6c33d392090524014ea3f0123a3c1eb9c2af5
video_preprocessor_config.json	817	59c5c9eb52182eb14c06ffb10ca9effd29adce5f238a95de23ca14a38dbd2cb1
vocab.json	2776833	ca10d7e9fb3ed18575dd1e277a2579c16d108e32f27439684afa0e10b1440910

text_encoder_nf4/ is the encoder (load it with Qwen3VLForConditionalGeneration). processor/ holds the tokenizer and processor files from the same source revision, so the pair can be used without the full base repository.

Licence and intended use

The base model is research-only; this derived checkpoint carries the same restriction. It is published to cut download size for a local, single-GPU research deployment. No weights were fine-tuned, merged, or altered apart from the quantization above.