Skip to content

unsloth

Recipe card from the charly-jupyter plugin (Images — the deployable catalog).

unsloth – LLM fine-tuning (Tier 1 post-install layer)

Section titled “unsloth – LLM fine-tuning (Tier 1 post-install layer)”
Property Value
Dependencies None (requires pixi env from parent)
Volumes models -> ~/.cache/huggingface
Aliases unsloth -> unsloth
Install files charly.yml, task:

This candy has no pixi.toml and no depends. It installs pip packages into the pixi environment established by a Tier 2 “environment-owner” parent candy. The user-phase tasks run AFTER the parent’s pixi COPY in the final image build.

Cannot be used standalone — must be composed into an environment-owner candy via the candy: field.

Variable Value
UNSLOTH_SKIP_LLAMA_CPP_INSTALL 1
HF_HOME ~/.cache/huggingface
  1. vLLM cu130 nightly wheel (pip install --no-deps) — installed here because vLLM wheel must run after pixi env exists
  2. unsloth + unsloth-zoo (pip install --no-deps) — incompatible with transformers 5.x in pixi solve
  3. opentelemetry runtime deps (pip install) — pixi resolver conflict prevents solving these via conda-forge
  4. vLLM torch.compile patch (patch_vllm_size_nodes.py) — fixes _decompose_size_nodes bug where getitem users and x.size(dim) patterns crash during model compilation (upstream: vllm-project/vllm#38360)

Use when the user asks about:

  • Unsloth fine-tuning setup or configuration
  • LLM fine-tuning environments
  • The unsloth Python library or unsloth-zoo
  • vLLM wheel installation inside pixi environments
  • HuggingFace model cache volume
  • The unsloth host alias
  • The two-tier layer architecture for ML layers
  • /charly-image:layer — candy authoring reference (charly.yml schema, task verbs, service declarations)
  • /charly-check:check — declarative testing (check: block, charly check box, charly check live)