Repository navigation
Releases
Releases · goodguy1963/ComfyUI-ThinkingLLM
v3.1.4
Full Changelog: v3.1.3...v3.1.4
v3.1.3
v3.1.2
v3.1.1
v3.1.0
v3.0.2
v3.0.1
v3.0.0
v2.5.8
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_attnsetting 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.