Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.
Python agentaichatgptdeepseekfine-tuninggemmaimage-generationllamallmllmsopenaipythonqwenreinforcement-learningself-hostedstable-diffusiontext-to-speechttsuiunsloth
Installers (ROCm on WSL): fetch the helper from the commit with the LD_LIBRARY_PATH fix (#14449)
Daniel Han69d84037d2a1
diff --git a/install.sh b/install.shindex b5b3d244a09..0d70687bcac 100755--- a/install.sh+++ b/install.sh@@ -6754,7 +6754,7 @@ _maybe_bootstrap_rocm_wsl() {substep "One-time, uses sudo and a large download. (skip: re-run with UNSLOTH_SKIP_ROCM_WSL_SETUP=1)"# Locate the helper: prefer the copy shipped beside install.sh, else fetch it. The local copy counts only for a --local checkout run, since this executes with no prompt and _REPO_ROOT may otherwise be the caller's cwd. PINNED, never a branch: this runs unattended and installs with sudo, so a moving ref would turn any rewrite of that branch into root code on every affected WSL box. Bump it whenever the helper changes; lagging only means an older helper, and the gate below rejects one too old to be safe.- _ROCM_WSL_HELPER_REF="b1d829182f8c490a326cf7690f156d2de06381f2"+ _ROCM_WSL_HELPER_REF="8f3d1c01e6c543f45f9a698f45bc48b5a262c928"# librocdxg pin (v1.2.2), forwarded to the helper. The ref IS the commit, so an older helper that ignores the SHA still resolves this exact revision: its `--branch <sha>` attempt fails and the full clone plus checkout land on it. Kept equal to the helper's defaults; a test enforces that. A user-set ref wins and, with no SHA of its own, turns the helper's check off rather than failing against our pin._rw_dxg_ref="${UNSLOTH_LIBROCDXG_REF:-}"_rw_dxg_sha="${UNSLOTH_LIBROCDXG_SHA:-}"