ethanfel's picture
fix: define explicit config schema with candidates — re-enable dataset viewer
a47f2ff verified
metadata
license: gpl-3.0
tags:
  - lora
  - merge
  - comfyui
  - stable-diffusion
configs:
  - config_name: config
    data_files: config/*.json
    features:
      - name: algo_version
        dtype: string
      - name: arch_preset
        dtype: string
      - name: lora_content_hashes
        sequence: string
      - name: score
        dtype: float64
      - name: config
        dtype:
          struct:
            - name: merge_mode
              dtype: string
            - name: sparsification
              dtype: string
            - name: sparsification_density
              dtype: float64
            - name: dare_dampening
              dtype: float64
            - name: merge_refinement
              dtype: string
            - name: auto_strength
              dtype: string
            - name: optimization_mode
              dtype: string
            - name: strategy_set
              dtype: string
      - name: candidates
        sequence:
          struct:
            - name: rank
              dtype: int64
            - name: config
              dtype:
                struct:
                  - name: merge_mode
                    dtype: string
                  - name: sparsification
                    dtype: string
                  - name: sparsification_density
                    dtype: float64
                  - name: dare_dampening
                    dtype: float64
                  - name: merge_refinement
                    dtype: string
                  - name: auto_strength
                    dtype: string
                  - name: optimization_mode
                    dtype: string
                  - name: strategy_set
                    dtype: string
            - name: score_heuristic
              dtype: float64
            - name: score_measured
              dtype: float64
            - name: score_final
              dtype: float64

LoRA Optimizer — Community Cache

Shared analysis results for the LoRA Optimizer ComfyUI node.

LoRA merge analysis is hardware-agnostic — the same LoRA files always produce the same conflict metrics and optimal merge config regardless of GPU tier. This dataset lets users share and reuse those results so nobody has to run the AutoTuner from scratch.


How It Works

The AutoTuner computes pairwise conflict metrics (cosine similarity, sign conflicts, subspace overlap) and tests merge parameter combinations to find the best config for a set of LoRAs. These results are keyed by content hash (SHA256[:16] of file contents) — not by filename — so they're portable across systems and private by design.

When community_cache=upload_and_download is set in the AutoTuner node:

  • Download: Before running analysis, the node checks this dataset for existing results. A config hit skips the entire sweep (~30–120s saved). Lora/pair cache hits speed up the analysis phase even without a full config hit.
  • Upload: After a successful sweep (or when replaying from local memory), results are uploaded if the local score beats the current community score for that LoRA set.

Privacy

LoRA filenames are never stored here. Only SHA256[:16] content hashes are used as keys. The uploaded data contains:

  • Per-prefix conflict metrics (cosine similarity, sign conflict ratios, subspace overlap)
  • Winning merge configuration (sparsification method, merge strategy, refinement level, etc.)
  • A composite quality score

No file paths, no usernames, no LoRA names.


File Structure

lora/
  {content_hash}.lora.json       # Per-LoRA per-prefix conflict stats
pair/
  {hash_a}_{hash_b}.pair.json   # Pairwise conflict metrics (hashes sorted)
config/
  {hash_a}_{hash_b}_..._{arch}.config.json  # Best merge config + score for a LoRA set

All files include an algo_version field. Results from incompatible algorithm versions are ignored automatically.


Usage

In the LoRA AutoTuner node, set community_cache to upload_and_download. That's the only option — there's no passive download-only mode. If you benefit from the cache, you contribute back.

Value Behavior
disabled (default) No network interaction
upload_and_download Download precomputed results and contribute yours back

Network errors are silently ignored — the node always falls back to local computation.


Setup

One time:

pip install huggingface_hub
huggingface-cli login

The node picks up your stored token automatically. No environment variables needed for most users.

Headless/server alternative: set HF_TOKEN as an environment variable.

Then: set community_cache=upload_and_download in the AutoTuner node and run as normal. Everything else is automatic.


Score-Based Replacement

Configs are only uploaded when your local score beats the community score. Users with more thorough sweeps (top_n=10) or better hardware naturally contribute higher-quality results over time.