coilyco-flight-deck/infrastructure

Watch

Install and live-verify Qwen Image Edit 2511 for AOSX #763

New issue

Open

opened 2026-08-06 11:24:17 +00:00 by coilyco-ops · 3 comments

coilyco-ops commented

2026-08-06 11:24:17 +00:00

Owner

Outcome

Converge the official Qwen Image Edit 2511 model selected by coilyco-bridge/agentic-os-xxx#24, then prove the new AOSX multi-reference workflow on the target 24 GiB runtime.

Selected model contract

Diffusion model: qwen_image_edit_2511_fp8mixed.safetensors from the official Comfy-Org Qwen Image Edit repository.
Reused official files: qwen_2.5_vl_7b_fp8_scaled.safetensors and qwen_image_vae.safetensors.
The diffusion file is 20.5 GB and Apache-2.0 licensed.
A configured AOSX filename is not installation or live-proof evidence.

Required operator work

Author and apply the infrastructure-owned model convergence for the selected diffusion file, including integrity verification and rollback-safe storage handling.
Verify live inventory exposes the configured edit model and the native TextEncodeQwenImageEditPlus, FluxKontextImageScale, FluxKontextMultiReferenceLatentMethod, ModelSamplingAuraFlow, and CFGNorm nodes.
Run the fixed seed 67401 pylon edit with the operator-controlled pylon target, warehouse baseline, and radar baseline.
Capture measured process peak, offload behavior, system-memory use, runtime, and failure evidence. Do not infer 24 GiB fit from file size or idle VRAM.
Preserve separate source, mask, and normalized 100 by 100 RGBA outputs.
Compare the old and edited pylon on transparent, light, dark, and Factory Game ground backgrounds. Designer verifies that topology and electrical identity remain intact while material, palette, edge treatment, lighting, and detail density converge toward the accepted family.

Acceptance

Infrastructure convergence is tracked and repeatable.
Live inventory distinguishes installed files from configured names.
The fixed-seed run completes without incompatible-model or missing-node errors.
Measured evidence establishes whether the selected model fits through supported offload behavior.
Source, mask, and RGBA artifacts remain separate.
Designer records the topology-preservation and visual-family verdict.

This issue is the live-operations handoff from coilyco-bridge/agentic-os-xxx#24. The Engineer-owned offline implementation and verification remain in AOSX.

## Outcome Converge the official Qwen Image Edit 2511 model selected by `coilyco-bridge/agentic-os-xxx#24`, then prove the new AOSX multi-reference workflow on the target 24 GiB runtime. ## Selected model contract * Diffusion model: `qwen_image_edit_2511_fp8mixed.safetensors` from the official Comfy-Org Qwen Image Edit repository. * Reused official files: `qwen_2.5_vl_7b_fp8_scaled.safetensors` and `qwen_image_vae.safetensors`. * The diffusion file is 20.5 GB and Apache-2.0 licensed. * A configured AOSX filename is not installation or live-proof evidence. ## Required operator work 1. Author and apply the infrastructure-owned model convergence for the selected diffusion file, including integrity verification and rollback-safe storage handling. 2. Verify live inventory exposes the configured edit model and the native `TextEncodeQwenImageEditPlus`, `FluxKontextImageScale`, `FluxKontextMultiReferenceLatentMethod`, `ModelSamplingAuraFlow`, and `CFGNorm` nodes. 3. Run the fixed seed `67401` pylon edit with the operator-controlled pylon target, warehouse baseline, and radar baseline. 4. Capture measured process peak, offload behavior, system-memory use, runtime, and failure evidence. Do not infer 24 GiB fit from file size or idle VRAM. 5. Preserve separate source, mask, and normalized 100 by 100 RGBA outputs. 6. Compare the old and edited pylon on transparent, light, dark, and Factory Game ground backgrounds. Designer verifies that topology and electrical identity remain intact while material, palette, edge treatment, lighting, and detail density converge toward the accepted family. ## Acceptance * Infrastructure convergence is tracked and repeatable. * Live inventory distinguishes installed files from configured names. * The fixed-seed run completes without incompatible-model or missing-node errors. * Measured evidence establishes whether the selected model fits through supported offload behavior. * Source, mask, and RGBA artifacts remain separate. * Designer records the topology-preservation and visual-family verdict. This issue is the live-operations handoff from `coilyco-bridge/agentic-os-xxx#24`. The Engineer-owned offline implementation and verification remain in AOSX.

coilyco-ops added the

autonomy

live-collab

label

2026-08-06 11:25:17 +00:00

coilyco-ops referenced this issue from a commit

2026-08-06 11:26:29 +00:00

feat: add Qwen multi-reference edit workflow

coilyco-ops commented

2026-08-06 11:45:06 +00:00

Author

Owner

Ops checkpoint:

Authored model convergence for qwen_image_edit_2511_fp8mixed.safetensors at immutable Comfy-Org revision 4c7c4ea236326cbae56d403d22a03c6cd86ad9a0, exact size 20533762817, and SHA-256 c9fdc158e46d3b61ef75f21ae866ca2fe808bf4a53643120d1c1e87c19280a4e.
Added missing-model-aware disk preflight with 8 GiB rollback headroom and live inventory checks for the five AOSX Qwen Edit nodes.
Updated the runtime, Ansible, and feature documentation.
ward exec windows-comfyui-convergence-test passed. ward exec pre-commit passed.
The validated change remains staged because the required signed commit cannot complete. The long native-session GPG home defect is recorded at coilyco-flight-deck/agentic-os#954. Using the short physical GPG home fixes key import, but signing then fails with Bad passphrase, so the SSM-backed signing key and passphrase require attended reconciliation. Signing was not bypassed.
No live tower convergence or fixed-seed generation has run. The next checkpoint remains commit and canonical push, followed by the attended model download, inventory verification, measured run, artifact preservation, and Designer verdict.

Ops checkpoint: * Authored model convergence for `qwen_image_edit_2511_fp8mixed.safetensors` at immutable Comfy-Org revision `4c7c4ea236326cbae56d403d22a03c6cd86ad9a0`, exact size `20533762817`, and SHA-256 `c9fdc158e46d3b61ef75f21ae866ca2fe808bf4a53643120d1c1e87c19280a4e`. * Added missing-model-aware disk preflight with 8 GiB rollback headroom and live inventory checks for the five AOSX Qwen Edit nodes. * Updated the runtime, Ansible, and feature documentation. * `ward exec windows-comfyui-convergence-test` passed. `ward exec pre-commit` passed. * The validated change remains staged because the required signed commit cannot complete. The long native-session GPG home defect is recorded at https://forgejo.coilysiren.me/coilyco-flight-deck/agentic-os/issues/954. Using the short physical GPG home fixes key import, but signing then fails with `Bad passphrase`, so the SSM-backed signing key and passphrase require attended reconciliation. Signing was not bypassed. * No live tower convergence or fixed-seed generation has run. The next checkpoint remains commit and canonical push, followed by the attended model download, inventory verification, measured run, artifact preservation, and Designer verdict.

coilyco-ops referenced this issue

2026-08-06 16:08:51 +00:00

Land Qwen Image Edit 2511 convergence source changes #764

coilyco-ops referenced this issue from a commit

2026-08-06 16:14:43 +00:00

feat: add Qwen Image Edit 2511 convergence

coilyco-ops commented

2026-08-06 16:39:24 +00:00

Author

Owner

2026-08-06 live convergence checkpoint:

Before state: native ComfyUI was healthy on the RTX 3090 Ti, the scheduled task was running, the queue had zero running and zero pending jobs, all five required Qwen Edit nodes were already exposed, and the Qwen Image Edit 2511 model was absent. Free space was 153946902528 bytes.
Updated the clean tower checkout to signed commit a2d34cf on canonical main and ran the tracked converger in WhatIf mode before applying.
Applied the converger once. It downloaded qwen_image_edit_2511_fp8mixed.safetensors, verified the exact 20533762817-byte size and SHA-256, verified every retained model asset, and completed with RESULT=CHANGED.
After state: native ComfyUI is healthy on CUDA, the scheduled task is running, the queue remains empty, the new model is present and exposed through live object inventory, and none of the five required native nodes are missing. Free space is 133398855680 bytes.
The ser8 gateway returns ready, proxies system stats successfully, and exposes the new model in object inventory.
The tower checkout is clean and synchronized with origin/main. No rollback was needed.

The model installation and live inventory acceptance are complete. The fixed-seed 67401 generation, process and offload measurements, artifact preservation, and Designer verdict remain open on this issue.

2026-08-06 live convergence checkpoint: * Before state: native ComfyUI was healthy on the RTX 3090 Ti, the scheduled task was running, the queue had zero running and zero pending jobs, all five required Qwen Edit nodes were already exposed, and the Qwen Image Edit 2511 model was absent. Free space was 153946902528 bytes. * Updated the clean tower checkout to signed commit a2d34cf on canonical main and ran the tracked converger in WhatIf mode before applying. * Applied the converger once. It downloaded qwen_image_edit_2511_fp8mixed.safetensors, verified the exact 20533762817-byte size and SHA-256, verified every retained model asset, and completed with RESULT=CHANGED. * After state: native ComfyUI is healthy on CUDA, the scheduled task is running, the queue remains empty, the new model is present and exposed through live object inventory, and none of the five required native nodes are missing. Free space is 133398855680 bytes. * The ser8 gateway returns ready, proxies system stats successfully, and exposes the new model in object inventory. * The tower checkout is clean and synchronized with origin/main. No rollback was needed. The model installation and live inventory acceptance are complete. The fixed-seed 67401 generation, process and offload measurements, artifact preservation, and Designer verdict remain open on this issue.

coilyco-ops commented

2026-08-20 05:17:27 +00:00

Author

Owner

Evidence drop from a live pass on kai-tower-3026 on 2026-08-19, so whoever picks this up knows which steps remain.

Required operator work item 1 is done

The infrastructure-owned convergence exists and the selected diffusion file is installed and integrity-verified. Checked against the pins in scripts/converge-comfyui-windows.ps1, computing digests directly rather than trusting the converger's own report:

MATCH  diffusion_models\qwen_image_edit_2511_fp8mixed.safetensors
MATCH  text_encoders\qwen_2.5_vl_7b_fp8_scaled.safetensors
MATCH  vae\qwen_image_vae.safetensors

Sizes and SHA-256 both match, including the 20533762817 byte edit model. The two reused official files named in the model contract are present and verified as well. The other three bundle members (qwen_image_fp8_e4m3fn, the union control LoRA, and BiRefNet) also verify, and the Juggernaut XL checkpoint is preserved.

This satisfies the "installed, not merely configured" distinction the issue calls out. A configured AOSX filename is not evidence, so the above is measured from disk.

Everything else here still stands

Items 2 through 6 are untouched. In particular item 2, live inventory of the configured model plus the native TextEncodeQwenImageEditPlus, FluxKontextImageScale, FluxKontextMultiReferenceLatentMethod, ModelSamplingAuraFlow, and CFGNorm nodes, cannot be checked right now because the runtime is down. Filed as #883: the process died at 20:42 PDT and the scheduled task did not restart it, while Tailscale Serve keeps forwarding tailnet 8188 to the dead loopback listener. The fixed-seed 67401 run and the Designer verdict are blocked behind that.

One caveat about how you verify

Do not use just windows-comfyui-converge -WhatIf to confirm installation state on a branch that predates #882. The preview reported this fully converged host as completely empty, planning all six models as missing and demanding 56.42 GiB. Fix is in #885. Until that lands, verify from disk the way this comment does, not from the preview.

Evidence drop from a live pass on `kai-tower-3026` on 2026-08-19, so whoever picks this up knows which steps remain. ## Required operator work item 1 is done The infrastructure-owned convergence exists and the selected diffusion file is installed and integrity-verified. Checked against the pins in `scripts/converge-comfyui-windows.ps1`, computing digests directly rather than trusting the converger's own report: ``` MATCH diffusion_models\qwen_image_edit_2511_fp8mixed.safetensors MATCH text_encoders\qwen_2.5_vl_7b_fp8_scaled.safetensors MATCH vae\qwen_image_vae.safetensors ``` Sizes and SHA-256 both match, including the 20533762817 byte edit model. The two reused official files named in the model contract are present and verified as well. The other three bundle members (`qwen_image_fp8_e4m3fn`, the union control LoRA, and BiRefNet) also verify, and the Juggernaut XL checkpoint is preserved. This satisfies the "installed, not merely configured" distinction the issue calls out. A configured AOSX filename is not evidence, so the above is measured from disk. ## Everything else here still stands Items 2 through 6 are untouched. In particular item 2, live inventory of the configured model plus the native `TextEncodeQwenImageEditPlus`, `FluxKontextImageScale`, `FluxKontextMultiReferenceLatentMethod`, `ModelSamplingAuraFlow`, and `CFGNorm` nodes, cannot be checked right now because **the runtime is down**. Filed as #883: the process died at 20:42 PDT and the scheduled task did not restart it, while Tailscale Serve keeps forwarding tailnet 8188 to the dead loopback listener. The fixed-seed `67401` run and the Designer verdict are blocked behind that. ## One caveat about how you verify Do not use `just windows-comfyui-converge -WhatIf` to confirm installation state on a branch that predates #882. The preview reported this fully converged host as completely empty, planning all six models as missing and demanding 56.42 GiB. Fix is in #885. Until that lands, verify from disk the way this comment does, not from the preview.

coilyco-ops added the

role/sysadmin

priority

labels

2026-08-25 08:02:16 +00:00

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#763

No description provided.

Rows
Columns