This model is an INT4 model of Qwen3.8-Flash-Next generated by intel/auto-round with Tuning mode. Please follow the license of the original model.
| Configuration | GSM8K | MMLU | PIQA | HelleSwag | Average | Relative to BF16 |
|---|---|---|---|---|---|---|
| BF16 | 0.9673 | 0.8651 | 0.8193 | 0.6927 | 0.8362 | - |
| INT4 | 0.9682 | 0.8560 | 0.8215 | 0.6873 | 0.8332 | 99.64% |
auto-round \
--model_name Qwen/Qwen3.8-Flash-Next \
--scheme W4A16 \
--iters 200 \
--ignore_layers \
lm_head,embed_tokens,visual,linear_attn,self_attn,hyper_connection,mlp.gate,shared_expert,in_proj_a,in_proj_b,ple,mtp,indexer \
--device_map auto \
--format auto_round \
--output_dir Qwen3.8-Flash-Next-W4A16-AutoRound
The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.
Therefore, before deploying any applications of the model, developers should perform safety testing.
Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.
Here are a couple of useful links to learn more about Intel's AI software:
The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.
@article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of llms}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi}, journal={arXiv preprint arXiv:2309.05516}, year={2023} }
model-00016-of-00017.safetensors 102 GBmodel-00007-of-00017.safetensors 5.4 GBmodel-00010-of-00017.safetensors 5.4 GBmodel-00013-of-00017.safetensors 5.4 GBmodel-00004-of-00017.safetensors 5.4 GBmodel-00003-of-00017.safetensors 5.4 GBmodel-00012-of-00017.safetensors 5.4 GBmodel-00009-of-00017.safetensors 5.4 GBmodel-00006-of-00017.safetensors 5.4 GBmodel-00005-of-00017.safetensors 5.4 GBmodel-00008-of-00017.safetensors 5.3 GBmodel-00014-of-00017.safetensors 5.3 GBmodel-00011-of-00017.safetensors 5.3 GBmodel_extra_tensors.safetensors 5.2 GBmodel-00015-of-00017.safetensors 5.1 GBmodel-00001-of-00017.safetensors 2.9 GBmodel-00017-of-00017.safetensors 1.3 GB有文件?提交种子。
提交种子docker run --rm -d --name lsv-cb-wufn-dflash2-tp4 --net=host --shm-size=32g --ipc=host --device /dev/dri:/dev/dri -v /dev/dri:/dev/dri -v /models/weights:/models:ro -v /models/bench/cb-20261023/session/cb-wufn-dflash2-tp4-c1/session/cache:/cache --security-opt seccomp=unconfined --security-opt label=disable --mount type=bind,source=/dev/dri/by-path,target=/dev/dri/by-path,readonly --env-file /models/bench/cb-20261023/session/cb-wufn-dflash2-tp4-c1/session/server-configs/cb-wufn-dflash2-tp4.env ghcr.io/wu1ff/qwen38-flashnext-b70:dflash2@sha256:6e2ab4e686680a5208ff0119c4637341cba81942bd39b6576ea14dc8f547b83b vllm serve /models/Qwen3.8-Flash-Next-W4A16-AutoRound-full --served-model-name qwen38-flash-next --dtype bfloat16 --enable-expert-parallel --max-num-seqs 4 --max-num-batched-tokens 1024 --gpu-memory-utilization 0.91 --limit-mm-per-prompt '{"image":2,"video":0}' --trust-remote-code --enable-prefix-caching --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3 --exclude-tools-when-tool-choice-none --enforce-eager --num-gpu-blocks-override 7785 --speculative-config '{"method":"dflash","model":"/models/Qwen3.8-Flash-Next-DFlash2-incoai","num_speculative_tokens":7,"draft_tensor_parallel_size":4,"draft_sample_method":"probabilistic","rejection_sample_method":"standard","attention_backend":"TRITON_ATTN"}' --tensor-parallel-size 4 --max-model-len 262144在 LocalMaxxing 上查看测试 git clone https://github.com/AlexBrandes/cmp170hx-qwen38-flash-next && cd cmp170hx-qwen38-flash-next && git checkout c0db9613d1 && HOST_TUNE=0 ./start.sh
# start.sh launches: vllm serve Qwen3.8-Flash-Next-W4A16-AutoRound --pipeline-parallel-size 2 (VLLM_PP_LAYER_PARTITION=26,22) --max-model-len 262144 --block-size 1680 --mamba-cache-mode align --mamba-ssm-cache-dtype float32 --max-num-seqs 8 --max-num-batched-tokens 2048 --gpu-memory-utilization 0.95 --enable-prefix-caching -cc=<config/compilation.json> --speculative-config {"method":"mtp","num_speculative_tokens":5} (int4 draft head)在 LocalMaxxing 上查看测试 来自 LocalMaxxing 的 3 次社区测试(2 种配置)的中位数。
自己运行:配置指南