chore: prepare v2.5.8 release

__init__.py

@@ -3,7 +3,7 @@
     "name": "ComfyUI-ThinkingLLM",
     "description": "A multimodal ComfyUI AI node with Qwen3.5, Qwen3-VL, Qwen2.5-VL, Qwen3, and Gemma 4 integrations. Features live thinking in the terminal to see what the LLM is doing in real time.",
     "author": "goodguy1963",
-    "version": "2.5.7",
+    "version": "2.5.8",
     "url": "https://github.com/goodguy1963/ComfyUI-ThinkingLLM",
     "category": "image"
 }

custom_nodes.json

@@ -9,5 +9,5 @@
   "description": "A multimodal ComfyUI AI node with Qwen3.5, Qwen3-VL, Qwen2.5-VL, Qwen3, and Gemma 4 integrations. Features live thinking in the terminal to see what the LLM is doing in real time.",
   "category": "image",
   "tags": ["vision", "language", "multimodal", "thinkingllm", "qwen", "gguf", "flash_attention", "performance", "wan2.2", "video", "i2v", "t2v", "storyboard", "caching", "fork", "gpl-3.0"],
-  "version": "2.5.7"
+  "version": "2.5.8"
 }

pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "ComfyUI-ThinkingLLM"
 description = "A multimodal ComfyUI AI node with Qwen3.5, Qwen3-VL, Qwen2.5-VL, Qwen3, and Gemma 4 integrations. Features live thinking in the terminal to see what the LLM is doing in real time."
-version = "2.5.7"
+version = "2.5.8"
 license = {file = "LICENSE"}
 dependencies = ["transformers>=5.2.0", "torch", "huggingface-hub", "hf_xet", "psutil", "numpy", "Pillow", "opencv-python", "bitsandbytes", "accelerate", "sentencepiece", "tiktoken", "faster-whisper", "llama-cpp-python"]
 

update.md

@@ -1,5 +1,18 @@
 # ComfyUI-QwenVL Update Log
 
+## Version 2.5.8 (2026/09/27) - ROCm SageAttention 1 Support
+
+### 2.5.8 Changes
+
+- Added a SageAttention 1.x Triton path for Hugging Face Qwen nodes on ROCm, including SageAttention 1.0.6. Select `attention_mode=sage`, or use `auto` to try Sage before Flash Attention 2 and SDPA.
+- Retained NVIDIA CUDA kernel selection. Unsupported attention masks, head sizes, and model layers use the original SDPA forward.
+- Protected the KV cache from SageAttention 1's in-place key smoothing and documented that the GGUF `flash_attn` setting is separate from Hugging Face attention selection.
+
+### 2.5.8 Validation
+
+- Seven focused SageAttention tests and all 153 feature and compatibility tests pass on the local Windows CPU test path.
+- End-to-end inference on ROCm hardware remains unverified.
+
 ## Version 2.5.7 (2026/08/26) - Safe Gemma 4 GGUF Audio
 
 ### 2.5.7 Changes