ComfyUI model assets and composable generation capabilities for vLLM-Omni.
English · 简体中文
ComfyOmni is a vLLM-Omni plugin. Its first release targets reuse of existing ComfyUI/H3 model and component files, with loading, unloading and switching in RAM/VRAM. LoRA composition, tools, node workflows and other model families follow later and do not block this release. A new converted checkpoint or assembled model directory is not a normal startup or model-switch prerequisite.
This is an independent project, not an official project of ComfyUI, vLLM or MiniMax.
| Role | Name/path |
|---|---|
| Product | ComfyOmni |
| Repository and distribution | comfy-omni |
| Python package | comfy_omni |
| Source | src/comfy_omni/ |
| CLI | comfy-omni |
| Host entry point group | vllm_omni.general_plugins |
The migration is incomplete. Single plugin registration, component catalog/API and the audited legacy H3 curve-cache compatibility path exist. That existing path has real-host parity evidence; it remains the fixed working baseline. A beta4 DiT forward test does not establish complete original-quantized-H3 loading or generation.
Direct loading of the existing original H3 quantized assets and runtime model switching still need implementation and real-host acceptance. LoRA lifecycle, tools and nodes are deferred work. The user guide separates current capabilities from targets. Live progress and evidence belong in Issue #4.
- Reference existing ComfyUI component files and reuse one fixed model environment.
- Manage residency, request isolation, loading and recovery in RAM/VRAM.
- Support each format explicitly, using native quantization or bounded load-time memory adaptation.
- Preserve proven legacy behavior while delivering verified H3 slices.
In-memory LoRA composition, tools and node workflows are later extensions, not first-release gates.
Application services coordinate source bindings, component loaders and model sessions. The vLLM-Omni adapter connects those services to the pinned host. Existing exported packages remain usable as one source kind; runtime no longer targets mandatory package creation.
See the runtime architecture and H3-first refactoring plan. These describe the target; they do not claim the refactor is already implemented.
| ID | User-visible result | Acceptance |
|---|---|---|
| H1 | Load existing component sources | Fixed H3 loading regression, preserved legacy behavior, no new full model copy |
| H2 | Reuse resident components and switch models | One control instance and existing workers perform A→B→A; identities, cache invalidation, recovery and resource release are verified |
| H3 | Validate and deliver on the real host | Fixed-asset loading, generation and normal worker switching pass; affected software gates pass and the candidate reaches main |
This is a dependency sequence, not a live status table. The fixed model validation baseline defines assets and verification boundaries. Ordinary CI never downloads model payloads. Worker reconstruction is a reported recovery or fallback path, not successful normal hot loading. LoRA composition, tools and node workflows are outside this first-release sequence.
src/comfy_omni/ plugin, contracts, loaders, runtime and application code
tests/ unit, contract, integration, packaging and host checks
docs/ current design, usage, testing and source attribution
scripts/ container execution and repository checks
.worktrees/ ignored development isolation within this plugin
DOCKER_FIRST_POLICY: v1
Read AGENTS.md, CONTRIBUTING.md and the Docker execution policy. Project Python, tests and model execution run inside Docker; hosts edit files and operate Git, Docker and SSH.
./scripts/docker.sh docs 3.13
./scripts/docker.sh quality 3.10
./scripts/docker.sh quality 3.13
./scripts/docker.sh package 3.12PowerShell uses scripts/docker.ps1. Real runtime regressions reuse existing read-only model
mounts across branches. Use small fixtures for numerical and failure tests; reuse valid evidence
when the relevant code and inputs have not changed. Do not create new model copies for code tests.
Use short-lived branches and focused pull requests against protected main. Keep English and
Chinese READMEs synchronized. See CONTRIBUTING.md.
Apache License 2.0. Migrated code and assets retain their recorded attribution and applicable license terms.