unsloth

Version	2026.144.1443
Repo	superproject

Patched vLLM nightly runtime layered into the pixi env for Unsloth LLM fine-tuning (pip-only, requires pixi env from parent layer) Layers a vLLM cu130 nightly wheel into the parent layer’s pixi default environment (installed –no-deps; the matching runtime deps live in the parent jupyter-ml pixi.toml), then patches vLLM’s _decompose_size_nodes bug (upstream vllm-project/vllm#38360) so the torch.compile graph passes stop crashing. It also exports UNSLOTH_SKIP_LLAMA_CPP_INSTALL=1 and HF_HOME (the HuggingFace cache, backed by the models volume), so the wheel install, the env, and the patch are each observable in the built image.

Acceptance plan

This candy’s plan: — the runnable spec charly check executes against a live deployment. check: steps are idempotent probes; run: steps change state.

Intent	Step
run	command=~/.pixi/envs/default/bin/python -m pip install –no-deps \ ‘https://wheels.vllm.ai/f53fa26e05c476a43f6db048a9e3b43bcb2b72fb/vllm-0.19.1rc1.dev39%2Bgf53fa26e0.cu130-cp38-abi3-manylinux_2_35_x86_64.whl’
run	command=~/.pixi/envs/default/bin/python /ctx/patch_vllm_size_nodes.py
check	the vLLM 0.19.x cu130 nightly wheel is installed in the pixi default environment
check	the candy exports UNSLOTH_SKIP_LLAMA_CPP_INSTALL=1 so the bundled llama.cpp build is skipped
check	the candy exports HF_HOME at the HuggingFace cache that backs the models volume
agent-check	vLLM’s compilation backend carries the _decompose_size_nodes fix (the x.size(dim) scalar guard) rather than the unpatched upstream code
check	file=${HOME}/.pixi/envs/default/bin/python