WARNING: I have not directly tested the .patch included here. The patch is for Linux and I am a Windows user so my NInfer is more customized. For me it works, hopefully it works for you, but know that you might be wasting your time.
I wanted to create a NInfer quant of QUASAR-QAT/Qwen3.8-27B-QUASAR-NVFP4 that is as true to the original as possible.
The difference in comparison to the two other NInfer quants you can find on HF of QUASAR-NVFP4 is that this one keeps both token_embedding and lm_head in BF16 just as they are in the original QUASAR model does.
I originally started with MirkoCovizzi's model and code, so thanks to MirkoCovizzi.
The artifact carries the identity qwen3.8-27b / nvfp4-quasar-bf16, which stock NInfer does not know. It loads only with a patch applied on top of:
NInfer at 6e8b2e2ad5d53597c3ba8e7989f9546d40b921fc.
1. Get the engine
git clone https://github.com/Neroued/ninfer.git
cd ninfer
git checkout 6e8b2e2ad5d53597c3ba8e7989f9546d40b921fc
2. Apply the patch (one, not both)
git apply quasar.patch
PS. Qwen said something about using --ignore-whitespace when running the patch command, so try that if it's complains.