Swift-1.5-Qwen3.8-27B-QSA-bf16

Overview

This model employs the methodology described at https://huggingface.co/higa2515/Qwen3.8-27B-QSA-BF16, replacing the Global Attention layerโ€”which traditionally consists of Grouped Query Attentionโ€”with QSA (Qwen Sparse Attention) as used in Qwen3.8-Flash-Next.

This model was created for the purpose of verifying QSA replacement technology. As it is not intended for inference in real-world environments, please be aware of this before use.

Limitations

We have not developed an Indexer for image-based mapping models. Additionally, please note that since no modifications have been made to MTP, Vision and MTP cannot be enabled.

The following product is required to run this model: https://github.com/kuraneko1/Qwen3.8-27B-QSA-llama.cpp Please install it as needed before trying it out.

Licensing

This model is treated as a derivative of Swift-1.5-Qwen3.8-27B and is therefore subject to the license terms defined therein. Specifically, it follows the [Swift Open License v1.0].

Accordingly, those intending to use this model for commercial purposes are requested to consult with ukisAI separately.

Qwen3.8-27B-QSA

Full multimodal backbone, with trained QSA indexers on text decoder full-attention layers. Text-only requests use QSA (K=2048, r=4). Image/video requests, including subsequent cached generation, use original dense attention. Padded requests also use dense attention. Backbone weights, vision encoder, MTP tensors, tokenizer, processor and chat template are preserved.

Load

Install the tested runtime with pip install -r requirements.txt. The indexers remain FP32; backbone weights remain BF16.

from transformers import AutoModelForImageTextToText, AutoProcessor

path = '/absolute/path/to/Qwen3.8-27B-QSA'
processor = AutoProcessor.from_pretrained(path, trust_remote_code=True)
model = AutoModelForImageTextToText.from_pretrained(
    path, trust_remote_code=True, dtype='auto', device_map='auto',
    attn_implementation='sdpa',
)
# Use processor.apply_chat_template(...) and model.generate(...) as for the base model.

Status

This runtime uses a dense Boolean mask as a correctness reference, not an optimized sparse kernel. No inference speedup is claimed. Use the default dynamic cache; speculative rollback, MTP execution, static/quantized caches and continuous batching are not implemented here. The original MTP weights are retained for future work. See conversion_manifest.json for provenance and validation limits, and BASE_MODEL_CARD.md for the original model card.

Downloads last month
1,040
Safetensors
Model size
28B params
Tensor type
BF16
ยท
F32
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for yue-neg-613/Swift-1.5-Qwen3.8-27b-QSA

Base model

Qwen/Qwen3.8-27B
Quantized
(93)
this model