Unverified50% confidenceReleaseExact time
此次更新新增了convrot的int4量化支持，并针对NVIDIA Turing架构显卡进行了int8/int4专项优化
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
7/16/2026
First Seen
Valid until: 10/14/2026
Sources
Related Claims
UnverifiedRTX 30系列（Ampere架构）的Tensor Core对INT4的原生支持有限，INT4运算通常需要拆包模拟67% similarUnverified有开发者将多个主流图像模型转换为INT4 ConvRot W4A4格式并适配最新版本的ComfyUI67% similarUnverifiedApple M4 Pro的Neural Engine达38TOPS，高通骁龙X Elite集成45TOPS NPU，英特尔Core Ultra也已集成NPU支持本地AI推理66% similarUnverifiedComfyUI-INT4-Fast支持原生混合精度检查点，按逐层粒度决定计算路径：主干模块路由至INT4 Tensor Core，敏感层（patch projection）以INT8格式导向Triton的INT8执行路径66% similarUnverifiedComfyUI-INT4-Fast在BobJohnson24的ComfyUI-INT8-Fast基础上独立改进而来，底层依托comfy-kitchen调用GPU的Tensor Core66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/535366API
curl https://kongchang.com/api/v1/knowledge/claims/535366MCP
get_claim(id=535366)