Skip to content

v2.5.8

Choose a tag to compare

@github-actions github-actions released this 27 Sep 13:01
· 16 commits to main since this release

ComfyUI-ThinkingLLM v2.5.8

SageAttention 1.x, including 1.0.6, can now be selected for Hugging Face Qwen nodes on ROCm through its Triton backend. Set attention_mode=sage; auto tries Sage before Flash Attention 2 and PyTorch SDPA.

  • NVIDIA CUDA kernel selection remains available.
  • Unsupported masks, head sizes, and model layers use the original SDPA path.
  • The SageAttention 1 path copies keys before the kernel to protect the KV cache from in-place smoothing.
  • The GGUF flash_attn setting belongs to llama.cpp and does not select SageAttention.

Validation: 7 focused tests, 153 feature and compatibility tests, and catalog validation passed on the local Windows CPU test path. End-to-end ROCm inference has not been verified on AMD hardware.