Skip to content

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.

README: say UNSLOTH_LLAMA_CPP_BACKEND picks the GGUF chat build, not PyTorch (#14256)

Leo Borcherding59631d2e68aa
diff --git a/README.md b/README.md
index 42b7a027812..296bbd2bf6f 100644
--- a/README.md
+++ b/README.md
@@ -409,7 +409,7 @@ Cap Unsloth's native CPU thread pools on high-core hosts: `UNSLOTH_CPU_THREADS=8
#### Vulkan, custom llama.cpp backends:
-You can force the backend during installation:
+Pick the llama.cpp build GGUF chat runs on during installation. PyTorch is picked separately from your GPU, so this does not change it:
```bash
export UNSLOTH_LLAMA_CPP_BACKEND=vulkan # or cpu, cuda, rocm, auto
curl -fsSL https://unsloth.ai/install.sh | sh
@@ -419,6 +419,15 @@ $env:UNSLOTH_LLAMA_CPP_BACKEND="vulkan" # or cpu, cuda, rocm, auto
irm https://unsloth.ai/install.ps1 | iex
```
+To choose PyTorch yourself, pin its index:
+```bash
+curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_TORCH_INDEX_FAMILY=cpu sh # or a family like rocm7.2
+```
+Or skip PyTorch for GGUF chat only:
+```bash
+curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_NO_TORCH=1 sh
+```
+
#### audio.cpp audio engine:
Setup also installs [audio.cpp](https://github.com/0xShug0/audio.cpp) (prebuilt by [unslothai/audio.cpp](https://github.com/unslothai/audio.cpp)), which runs the audio.cpp speech, music and speech-to-text models on the Audio page, in Voice settings and behind `/v1/audio/*`. It is optional and never blocks setup: