Benign huggingface Download

kaushikvira/Qwen3.8-27B-swift15-smoothquant-nvfp4mlp-dflash2-NInfer-v3 main

Markdown-style heading precedes a code fenceAssessment or defense section heading

SHA-25632d43b3ed05735629e2f1a87d09245d23fce3056f65494607d798384a606ae57

URLhttps://huggingface.co/kaushikvira/Qwen3.8-27B-swift15-smoothquant-nvfp4mlp-dflash2-NInfer-v3

Evidence

English function-word token "the" lines 1–4

1:69… er v3 container
2
3Everything needed to reproduce the artifact in this repo. The scripts in `build/` are the
4ones actually used; `build/run.sh` drives them. No paths  …

English function-word token "for" lines 4–11

4:65… d-coded — set the env vars
5in the box below.
6
7```bash
8MODELS=/path/to/models        # BF16 base, DFlash2 drafter, outputs
9NINFER=/path/to/ninfer        # NInfer engine source (for tools.convert + shape gate)
10CALIB=/path/to/calib          # calibration corpus (see §2)
11IMGPY= …

At least four Markdown section headings lines 12–20

12IMGCONV=pytorch/pytorch:2.7.1-cuda12.8-cudnn9-runtime
13```
14
15---
16
17## 1. What this artifact is
18
19A 27B hybrid-attention VLM (Qwen3.8-27B) post-trained by **ukisai** into Swift-1.5, packed
20for the **NInfer** v3 container. Th …

English function-word token "and" lines 30–32

11… 27 blocks) | **BF16**, except `merger/fc2` (Q8) and `position_embedding` (Q5) | OCR/detail headroom; the two exceptions have no BF16 kernel (see §6) |
MTP | Q8 + BF16 norms | stored, not loaded (no `--spec mtp`) |
D …

Documentation references model or LLM output lines 120–123

56… l.safe_load` and raises
 `ConstructorError` — **after** the model has been written. The output is fine; delete the
 file after stage 1 (run.sh does).
**Stage 1 saves no vision weights.** `AutoModelForCausalLM` loads t …