[jamesbrink]

Summary

• ComfyUI v0.37.0 -> v0.44.0, with the vendored wheels pinned to exactly what v0.44.0's requirements.txt requires:
  • frontend 1.52.7 -> 1.57.6
  • workflow templates 0.11.66 -> 0.11.101 (core 0.3.392, JSON 0.1.127, media assets 01 0.1.48, media assets 02 0.1.27)
  • embedded docs 0.5.12 -> 0.5.16
  • comfy-kitchen 0.2.35 -> 0.2.40
  • comfy-aimdo 0.5.5 -> 0.6.0
• Flake inputs: nixpkgs (2026-09-07 -> 2026-11-07) and flake-parts.
• Template inputs: regenerated (816 files, up from 728).
• Build fixes the new nixpkgs needed:
  • valkey: deselect the timing-sensitive test_cwe_404.py::test_standalone_pipeline. It hit pytest-timeout under sandbox load and failed the whole build via narwhals' check inputs. The macOS port isolation is kept behind isDarwin.
  • timm: add openssl to the check inputs. Inductor's codecache shells out to openssl; nixpkgs' torch patches in a store path, but the CUDA/ROCm/XPU wheels look it up on PATH. This broke extra-python-packages-cuda.
• CHANGELOG updated.

Testing

• nix flake check passes locally (x86_64-linux).
• Smoke test of the CPU package on ./data:
  • ComfyUI 0.44.0, frontend 1.57.6 and torch 2.14.1 start cleanly.
  • All 5 model_downloader routes register.
  • A download through the API (POST /model-downloader/download, TAESD 9.8 MB) completed and shows up in /models/vae_approx.
  • In headless Chromium (Playwright), the frontend's detached <a download> click was intercepted and sent to the backend. The download completed and the Model Downloads widget shows "Complete".
• Peer reviewed by an Opus sub-agent. All 46 pythonhosted URLs and hashes were verified against PyPI JSON.

Not covered / follow-ups

• The CUDA, ROCm and XPU packages were not run on GPU; only the CPU package was smoke-tested. comfy-aimdo 0.6.0 (CUDA memory management) has not been exercised.
• The frontend console logs Unknown message type model_download_progress. This is not a regression; 1.57.6 just logs at a higher level. model_downloader.js registers through window.api, which no current frontend defines, so progress arrives only through the WebSocket fallback. A follow-up should switch it to import { api } from "../../scripts/api.js".
• Upstream dropped torchaudio from requirements.txt. It is still packaged here.
• The Impact-Pack SAM editor fails to import the legacy /extensions/core/clipspace.js. This already happened on frontend 1.52.7.
• The pre-built torch wheels' inductor still needs openssl on PATH at runtime for torch.compile. This predates the PR and should be checked in the launcher.

[jamesbrink]

CUDA verification on hal9000 (RTX 4090, driver 595.71.05, nix run .#cuda -- --base-directory ./data --use-pytorch-cross-attention --cuda-malloc --lowvram --enable-manager):

• Starts cleanly: ComfyUI 0.44.0, torch 2.10.0+cu130, xformers 0.0.35, cudaMallocAsync.
• comfy-aimdo 0.6.0 initializes for the 4090, with DynamicVRAM enabled. comfy-kitchen 0.2.40 CUDA backend is available.
• Downloaded SD 1.5 fp16 (2.1 GB) via /model-downloader/download: completed in 11.6 s.
• Ran a 512×512, 20-step txt2img: success, about 47 it/s, prompt executed in 1.90 s. All models loaded through dynamic VRAM. Output image looks correct.