#!/usr/bin/env bash # Install the Qwen3.8-Next-Flash configuration. # # Same runtime and engine build as install.sh -- the retained-PM4 ROCr/HIP prefixes and the # strix-halo llama.cpp branch -- but a different model set and a launcher tuned for the # 16384-token prefill path. Kept as a separate entry point because the disk and memory # requirements are much larger than the 27B configuration's. # # bash <(curl -fsSL https://raw.githubusercontent.com/pwilkin/strix-halo/main/install-flash-next.sh) # # Every flag install.sh accepts is forwarded, so --check-only and --model-dir work here too. set -Eeuo pipefail IFS=$'\n\t' readonly installer_url=https://raw.githubusercontent.com/pwilkin/strix-halo/main/install.sh printf '%s\n' \ 'Qwen3.8-Next-Flash: 177B parameters, 93 GiB of IQ4_NL weights plus a 2.8 GB draft.' \ 'Needs ~110 GiB free on the model disk and a 128 GB unified-memory machine.' \ 'Prefill is the tuned path; decode is still behind where it should be.' \ '' # Prefer a sibling install.sh when this script was downloaded as part of the repo, so a # local checkout is not silently bypassed in favour of whatever main happens to hold. self_dir=$(cd -- "$(dirname -- "${BASH_SOURCE[0]}")" >/dev/null 2>&1 && pwd -P || true) if [[ -n ${self_dir:-} && -r $self_dir/install.sh ]]; then STRIX_PROFILE=flash-next exec bash "$self_dir/install.sh" "$@" fi tmp=$(mktemp -d) trap 'rm -rf "$tmp"' EXIT curl -fsSL "$installer_url" -o "$tmp/install.sh" STRIX_PROFILE=flash-next exec bash "$tmp/install.sh" "$@"