dinkster-inference

This distribution packages upstream ComfyUI's inference library as dinkster_inference. The comfy import name aliases the same modules, not a second instance of model-management state. ComfyUI's application, server and node packages are not distributed here. Modules absent from inference can resolve from a ComfyUI checkout's comfy/ directory on sys.path; inference modules always take precedence over checkout modules.

python -m pip install .

Releases

Consumers pin an exact dinkster-inference==X.Y.Z release in package metadata and record its PyPI wheel hashes in their lockfile. Before the first PyPI publication of 0.1.0, build a local wheel with python -m build.

The upstream sync owner updates project.version when landed inference changes need a new release and validates that commit before requesting publication. An authorized maintainer pushes the matching vX.Y.Z tag from the landed commit. .github/workflows/release.yml builds the wheel and sdist, checks their metadata, and publishes those same artifacts using trusted publishing. Branches and pull requests build without publishing. After publication the sync owner updates Dinkster's exact requirement and lock together; setup scripts do not select a separate inference commit.

The repository owner must register the PyPI project dinkster-inference with a trusted publisher for owner Kosinkadink, repository dinkster-inference, workflow release.yml, environment pypi. Before publishing, protect that GitHub environment with required owner approval and restrict deployment to version tags. Publication requires the owner's authorization; no stored PyPI API token is used.

As in upstream ComfyUI, CPU callers select CPU execution before importing model management or loaders:

import dinkster_inference.cli_args as cli_args
cli_args.args.cpu = True
import comfy.sd
import dinkster_inference.sd
assert comfy.sd is dinkster_inference.sd

See upstream provenance for the fork point and omitted application tests. python tools/check_upstream_parity.py checks the library and retained test files against that upstream commit. New upstream files use tools/rewrite_upstream_imports.py for import renaming and ASCII source spelling; escaped string values and non-ASCII tokenizer resource data retain their meaning.

Workflow comparison

Install a built wheel and its CUDA dependencies in a separate environment, then run the harness with that environment's Python. The fork worker uses the installed package with isolated imports; the reference worker imports a clean, detached ComfyUI checkout at the recorded fork point. Both use the same interpreter, models, seed and deterministic PyTorch settings in separate processes.

CUDA_VISIBLE_DEVICES=1 /path/to/venv/bin/python -I tools/compare_workflows.py \
  --upstream /path/to/ComfyUI --output-dir /path/to/results \
  --checkpoint /path/to/v1-5-pruned-emaonly-fp16.safetensors \
  --lora /path/to/lcm-lora-sdv1-5.safetensors \
  --controlnet /path/to/control_v11p_sd15_canny.safetensors \
  --h3-model /path/to/minimax_h3_fl2va_int8_convrot.safetensors \
  --h3-clip /path/to/qwen3vl_32b_minimax_h3_int8_convrot.safetensors \
  --h3-vae /path/to/minimax_h3_video_vae_fp16.safetensors

The five SD1.5 cases use 512x512, seed 459, Euler/normal, 20 steps and CFG 7: baseline; LoRA strength 0.8; five linearly interpolated hook LoRA keyframes from strength 0 to 1 over the first half of sampling (first guarantee 1, subsequent guarantees 0, matching upstream's interpolated-keyframe node); ControlNet strength 1 over timestep range 0.0-0.6 using a fixed asymmetric edge map; and overlapping 16-frame context windows over a 32-frame batch. H3 uses 1344x768, 124 frames, seed 1064, Euler/normal, 20 steps and CFG 6, with the 32B text encoder on CPU on both sides to avoid GPU dequantization memory exhaustion. If H3 weights are absent, supply a bundled Flux dev checkpoint via --flux-checkpoint instead (512x512, four steps, seed 459, CFG 1, guidance 3.5). --cases selects individual cases for diagnosis.

comparison.json records input hashes, installed-package provenance, environment, per-case raw latent and PNG hashes, and maximum absolute differences. The result directory also contains tensors, PNGs and worker logs. Exit status is zero only when every selected case is identical; missing model files, worker errors and non-finite values never count as parity. No model downloads happen in the harness.