●Change8

v0.3.68

Breaking Changes
📦 comfyuiView on GitHub →
⚠ 2 breaking✨ 5 features🐛 11 fixes🔧 15 symbols

Summary

This release introduces a new Mixed Precision Quantization System and significant performance improvements, especially around memory offloading and torch.compile optimizations for fp8 operations. Several API nodes were updated to use a new client structure, and various stability fixes were applied to caching and execution flow.

⚠️ Breaking Changes

  • The Comfy API key is removed from the queue API. If you were relying on this for queue management via the API, you will need to update your integration.
  • API Nodes for Luma, Minimax, Pixverse, Ideogram, StabilityAI, and Pika have been updated to use a new API client. This might require updates to how these nodes interact with their respective services.

Migration Steps

  1. If using the queue API, update your calls as the Comfy API key has been removed.
  2. Users experiencing issues with the portable version should update their NVIDIA drivers.
  3. Users are advised to update their PyTorch versions.

✨ New Features

  • Introduction of a Mixed Precision Quantization System.
  • Subgraph nodes can now execute multiple times.
  • Added RAM Pressure cache mode.
  • Added a ScaleROPE node, initially for WAN models, later expanded to Lumina models.
  • Clarified help text for the --fast argument.

🐛 Bug Fixes

  • Fixed an issue where weights were not being unpinned.
  • Fixed a race condition in --async-offload that could cause corruption.
  • Addressed issues related to pinned memory usage, including trying to fix slow load issues on low RAM hardware.
  • Fixed a small performance regression with fp8 fast and scaled fp8 operations.
  • Fixed a torch compile regression on fp8 operations, and subsequent fixes to restore fp8 torch compile performance.
  • Fixed img2img operation in the Dall2 API node.
  • Fixed rope scaling issues.
  • Fixed an issue where the Rodin3D API nodes returned absolute paths instead of relative paths for 3D models.
  • Fixed cache signature hashing to correctly treat bytes data as a primitive type.
  • Fixed handling of None outputs tuple case in caching.
  • Limited the amount of pinned memory on Windows to prevent issues.

🔧 Affected Symbols

queue apinodes_recraft.pyLuma API nodeMinimax API nodePixverse API nodeIdeogram API nodeDall2 API nodeScaleROPE nodeLTXV API nodesStabilityAI API nodeRodin3D nodesnodes_hypernetwork.pynodes_openai.pyPika API nodestorch.compile