coilyco-flight-deck/infrastructure

Watch

Converge Qwen Image and BiRefNet assets for ComfyUI #759

New issue

Closed

opened 2026-08-06 00:12:54 +00:00 by coilyco-ops · 2 comments

coilyco-ops commented

2026-08-06 00:12:54 +00:00

Owner

Outcome

The native ComfyUI runtime on kai-tower-3026 idempotently installs and verifies the approved Qwen Image structure-control and BiRefNet model bundle required by coilyco-bridge/agentic-os-xxx#21.

Approved source

Kai approved the bundle in AOSX commit 9aa7566:

.agents/skills/kai-comfyui-industrial/references/approved-capability-bundle.md

Required model files

qwen_image_fp8_e4m3fn.safetensors in the Qwen diffusion-model location.
qwen_2.5_vl_7b_fp8_scaled.safetensors in the text-encoder location.
qwen_image_vae.safetensors in the VAE location.
qwen_image_union_diffsynth_lora.safetensors in the LoRA location.
birefnet.safetensors in models/background_removal.

Expected payload is about 31.4 GB. Require at least 40 GB free before download. Do not add an upscaler or style LoRA in this change.

Convergence requirements

Extend the existing Windows ComfyUI convergence path rather than performing an untracked manual install.
Pin authoritative download URLs, expected byte sizes, and SHA-256 digests.
Download atomically and reject partial or mismatched files.
Keep the durable model store outside versioned application releases.
Preserve the existing Juggernaut XL checkpoint.
Make repeated convergence a no-op when every verified file is present.
Restart ComfyUI only when the owning runtime procedure requires it.
Keep private host and network identifiers out of tracked artifacts.

Repository verification

-WhatIf or the repository's equivalent dry path exposes the planned changes.
Script and documentation checks pass through Ward.
Update the ComfyUI runtime documentation and feature inventory when the public boundary changes.

Live acceptance

Ops performs the live convergence and returns read-only evidence:

Preflight reports sufficient free disk.
Every file passes size and digest verification.
ComfyUI returns healthy NVIDIA system stats after convergence.
Live object inventory lists the exact Qwen diffusion model, text encoder, VAE, control LoRA, and BiRefNet model.
Existing Juggernaut generation remains available.
AOSX issue #21 can execute its three-asset live acceptance after its workflow implementation lands.
Capture disk consumed, warm generation time, and peak VRAM or system-RAM offload during the first Qwen run.

Dependencies and authority

Infrastructure owns model installation and live convergence. AOSX issue coilyco-bridge/agentic-os-xxx#21 owns the workflow and CLI. Designer owns final asset review. Engineer and QA must not iterate against the live runtime.

## Outcome The native ComfyUI runtime on kai-tower-3026 idempotently installs and verifies the approved Qwen Image structure-control and BiRefNet model bundle required by coilyco-bridge/agentic-os-xxx#21. ## Approved source Kai approved the bundle in AOSX commit `9aa7566`: `.agents/skills/kai-comfyui-industrial/references/approved-capability-bundle.md` ## Required model files * `qwen_image_fp8_e4m3fn.safetensors` in the Qwen diffusion-model location. * `qwen_2.5_vl_7b_fp8_scaled.safetensors` in the text-encoder location. * `qwen_image_vae.safetensors` in the VAE location. * `qwen_image_union_diffsynth_lora.safetensors` in the LoRA location. * `birefnet.safetensors` in `models/background_removal`. Expected payload is about 31.4 GB. Require at least 40 GB free before download. Do not add an upscaler or style LoRA in this change. ## Convergence requirements * Extend the existing Windows ComfyUI convergence path rather than performing an untracked manual install. * Pin authoritative download URLs, expected byte sizes, and SHA-256 digests. * Download atomically and reject partial or mismatched files. * Keep the durable model store outside versioned application releases. * Preserve the existing Juggernaut XL checkpoint. * Make repeated convergence a no-op when every verified file is present. * Restart ComfyUI only when the owning runtime procedure requires it. * Keep private host and network identifiers out of tracked artifacts. ## Repository verification * `-WhatIf` or the repository's equivalent dry path exposes the planned changes. * Script and documentation checks pass through Ward. * Update the ComfyUI runtime documentation and feature inventory when the public boundary changes. ## Live acceptance Ops performs the live convergence and returns read-only evidence: 1. Preflight reports sufficient free disk. 2. Every file passes size and digest verification. 3. ComfyUI returns healthy NVIDIA system stats after convergence. 4. Live object inventory lists the exact Qwen diffusion model, text encoder, VAE, control LoRA, and BiRefNet model. 5. Existing Juggernaut generation remains available. 6. AOSX issue #21 can execute its three-asset live acceptance after its workflow implementation lands. 7. Capture disk consumed, warm generation time, and peak VRAM or system-RAM offload during the first Qwen run. ## Dependencies and authority Infrastructure owns model installation and live convergence. AOSX issue coilyco-bridge/agentic-os-xxx#21 owns the workflow and CLI. Designer owns final asset review. Engineer and QA must not iterate against the live runtime.

coilyco-ops referenced this issue from a commit

2026-08-06 00:41:28 +00:00

feat: converge Qwen Image ComfyUI models

coilyco-ops referenced this issue from a commit

2026-08-06 00:44:04 +00:00

feat: broker ComfyUI convergence through Ward

coilyco-ops closed this issue

2026-08-06 01:00:15 +00:00

coilyco-ops commented

2026-08-06 01:04:07 +00:00

Author

Owner

Live convergence completed from landed commit 3bd3f8b.

Repository verification: Ward PowerShell parse and binding checks passed, the full pre-commit suite passed, and Forgejo lint.yml plus trufflehog.yml passed.
Preflight: 179.2 GiB free against the 40 GiB minimum.
Convergence: exact pinned byte-size and SHA-256 verification passed for all five approved files. The bundle consumed about 29.3 GiB.
Runtime: ComfyUI 0.29.0 remained healthy on the NVIDIA CUDA device. No restart was required because the launcher and scheduled-task state were already converged.
Inventory: /object_info exposed Juggernaut XL plus the Qwen diffusion model, text encoder, VAE, control LoRA, and BiRefNet.
Idempotence: the second ward exec windows-comfyui-converge run required no downloads and returned RESULT=OK.
Rollback: only the five new model files need removal. Juggernaut XL was unchanged.
Follow-up owner: first Qwen generation timing, peak allocation or system-RAM offload, and final asset acceptance remain tracked by coilyco-bridge/agentic-os-xxx#21, which owns the workflow and CLI.

Live convergence completed from landed commit `3bd3f8b`. * Repository verification: Ward PowerShell parse and binding checks passed, the full pre-commit suite passed, and Forgejo `lint.yml` plus `trufflehog.yml` passed. * Preflight: 179.2 GiB free against the 40 GiB minimum. * Convergence: exact pinned byte-size and SHA-256 verification passed for all five approved files. The bundle consumed about 29.3 GiB. * Runtime: ComfyUI 0.29.0 remained healthy on the NVIDIA CUDA device. No restart was required because the launcher and scheduled-task state were already converged. * Inventory: `/object_info` exposed Juggernaut XL plus the Qwen diffusion model, text encoder, VAE, control LoRA, and BiRefNet. * Idempotence: the second `ward exec windows-comfyui-converge` run required no downloads and returned `RESULT=OK`. * Rollback: only the five new model files need removal. Juggernaut XL was unchanged. * Follow-up owner: first Qwen generation timing, peak allocation or system-RAM offload, and final asset acceptance remain tracked by `coilyco-bridge/agentic-os-xxx#21`, which owns the workflow and CLI.

coilyco-ops commented

2026-08-06 01:14:11 +00:00

Author

Owner

Follow-up correction: the three-asset generation and design acceptance is tracked by coilyco-bridge/agentic-os-xxx#22. Its first live run measured 116.3 seconds and about 23.37 GiB maximum observed device allocation, then exposed a missing RGBA branch. The product defect is recorded in coilyco-bridge/agentic-os-xxx#23. Infrastructure convergence remains complete and idempotent.

Follow-up correction: the three-asset generation and design acceptance is tracked by `coilyco-bridge/agentic-os-xxx#22`. Its first live run measured 116.3 seconds and about 23.37 GiB maximum observed device allocation, then exposed a missing RGBA branch. The product defect is recorded in `coilyco-bridge/agentic-os-xxx#23`. Infrastructure convergence remains complete and idempotent.

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference

coilyco-flight-deck/infrastructure#759

No description provided.

Rows
Columns