Benign huggingface Download

azukivc/Qwen3.8-27B-Abliterated-SFT main

Shell changes working directoryExecutes a child process

SHA-256f3b23b8fd6cdfd5bb37e9b228b8af26ee2f55f327035c36ea9623a3dd8cc2e98

URLhttps://huggingface.co/azukivc/Qwen3.8-27B-Abliterated-SFT

Evidence

English function-word token "this" lines 1–6

1#!/usr/bin/env bash
2# Qwen3.8-27B-Abliterated-SFT — L40-optimized serving (llama.cpp)
3#
4# Default profile: DUAL L40 (2x 46GiB), validated on this exact hardware —
5#   Q8_0 weights + 8 concurrent slots x 262,144 tokens each (full native
6#   …

Shell changes working directory lines 9–16

9:35… OFILE=single (Q5_K_M, 2 x 64k slots).
10set -euo pipefail
11
12DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
13PROFILE="${LLAMA_PROFILE:-dual-maxctx}"
14
15if [[ "$PROFILE" == "single" ]]; then
16    MODEL="${LLAMA_MODEL:-$DIR/gguf/qwen38- …

English function-word token "with" lines 20–22

21… ve (max quality, dual only): MODEL=...bf16.gguf with
  #   -c 1114112 (8 x 139k) — full 262k/slot does not fit with bf16 weights.
  PARALLEL=8 …

English function-word token "and" lines 46–48

2…    # qwen35 SSM/attention hybrids; server warns and disables).
  --cache-reuse 256
  --spec-type ngram-mod        # $0 speculative decode (~1.3 …

Uses shell exec to replace process lines 65–68

65:32… PROJ")"
66echo "Endpoint: http://$HOST:$PORT"
67echo "============================================================"
68exec llama-server "${ARGS[@]}"