fix: move HuggingFace cache to user home, add gitignore#54
fix: move HuggingFace cache to user home, add gitignore#54vlordier wants to merge 1 commit intoLiquid4All:mainfrom
Conversation
There was a problem hiding this comment.
Pull request overview
This PR aims to prevent large HuggingFace cache directories from being committed into the examples/localcowork project by ignoring _models/.cache/ and documenting correct HuggingFace cache env var usage.
Changes:
- Add
_models/.cache/toexamples/localcowork/.gitignore. - Update
_models/config.yamlwith guidance aboutHF_HOME/HF_HUB_CACHE. - Update parts of
_models/config.yamlmodel registry content (model naming/paths and an LM Studio entry).
Reviewed changes
Copilot reviewed 2 out of 2 changed files in this pull request and generated 5 comments.
| File | Description |
|---|---|
| examples/localcowork/_models/config.yaml | Adds HF cache warning and modifies model config entries (including LFM2 naming/path and new LM Studio entry). |
| examples/localcowork/.gitignore | Ignores _models/.cache/ to prevent accidental cache commits. |
💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.
| lfm2-24b-a2b: | ||
| display_name: "LFM2-24B-A2B-Preview" | ||
| display_name: "LFM2-24B-A2B" | ||
| runtime: llama_cpp | ||
| model_path: "${LOCALCOWORK_MODELS_DIR:-~/Projects/_models}/LFM2-24B-A2B-Preview-Q4_K_M.gguf" | ||
| model_path: "${LOCALCOWORK_MODELS_DIR:-~/Projects/_models}/LFM2-24B-A2B-Q4_K_M.gguf" |
There was a problem hiding this comment.
Within the same model entry, the display_name/model_path were updated to drop -Preview, but other tooling in the repo still expects LFM2-24B-A2B-Preview-Q4_K_M.gguf (scripts + docs). If this filename change is intentional, the surrounding ecosystem needs to be updated together; otherwise, users will end up with a downloaded file that doesn’t match what the config suggests.
| # LM Studio headless server — any model loaded in LM Studio | ||
| # Run `lms server start` or enable "Run LLM server on login" in app settings. | ||
| # Default port is 1234. Uses OpenAI-compatible API. | ||
| # Note: The model_name here is informational - update to match your loaded model. | ||
| lmstudio-default: | ||
| display_name: "LM Studio (Default)" | ||
| runtime: lmstudio | ||
| model_name: "liquid/lfm2-24b-a2b" # Replace with your loaded model ID |
There was a problem hiding this comment.
The new lmstudio-default model and runtime: lmstudio addition is a functional/config expansion that isn’t mentioned in the PR description (which focuses on HuggingFace cache + gitignore). If this is intended, please call it out in the PR summary; otherwise consider moving it to a separate PR to keep scope tight and reduce review/rollback risk.
| active_model: lfm2-24b-a2b # Sparse MoE: 24B total, 2.3B active, 64 experts top-4 — 80% tool accuracy | ||
|
|
||
| # Default model directory for non-Ollama model files (GGUF, MLX, etc.) | ||
| models_dir: "${LOCALCOWORK_MODELS_DIR:-~/Projects/_models}" |
There was a problem hiding this comment.
This PR introduces a large number of whitespace-only edits (e.g., changing inline-comment spacing) throughout the config. This makes the diff noisier and increases the chance of merge conflicts without affecting behavior. Consider reverting purely-formatting changes and keeping this PR focused on the HuggingFace cache/gitignore fix.
| # LFM2-24B-A2B — Liquid AI's MoE hybrid model (private preview) | ||
| # Architecture: 24B total, 2.3B active per token, 64 experts top-4, 40 layers (1:3 attn:conv ratio) | ||
| # Download GGUF from: https://huggingface.co/LiquidAI/LFM2-24B-A2B-Preview (gated — request access) | ||
| # Download GGUF from: https://huggingface.co/LiquidAI/LFM2-24B-A2B (gated — request access) | ||
| # Benchmark plan: docs/model-analysis/lfm2-24b-a2b-benchmark.md | ||
| # Run: llama-server --model <path> --port 8080 --ctx-size 32768 --n-gpu-layers 99 --flash-attn |
There was a problem hiding this comment.
The LFM2-24B-A2B comments and download URL were changed from the ...-Preview repo to LiquidAI/LFM2-24B-A2B, but the rest of this repo still references the ...-Preview location/filename (e.g., scripts/start-model.sh, scripts/setup-dev.sh, and README.md). As-is, contributors following the scripts/docs will likely download a different filename/repo than this config implies. Please either update the related scripts/docs in the same PR, or keep the config pointing at the ...-Preview repo to stay consistent.
vlordier
force-pushed
the
fix/huggingface-cache-location
branch
2 times, most recently
from
1b4c55e to
c787d50
Compare
- Add _models/.cache/ to .gitignore to prevent accidental cache commits - Set HF_HOME and HF_HUB_CACHE in scripts to ensure cache goes to ~/.cache - Add warning in config about using local_dir outside project repo - Add note in .env.example about storing models outside project This prevents large HuggingFace model caches from being stored in the project directory (fixes 77GB cache issue).
vlordier
force-pushed
the
fix/huggingface-cache-location
branch
from
c787d50 to
a22f2af
Compare
|
@copilot open a new pull request to apply changes based on the comments in this thread |
Address Copilot AI review comments from PR Liquid4All#54: - Add _models/.cache/ to .gitignore to prevent accidental cache commits - Add comment about HuggingFace cache location (HF_HOME, HF_HUB_CACHE, XDG_CACHE_HOME)
Summary
This prevents large HuggingFace model caches from being stored in the project directory (fixes 77GB cache issue).