Kernels
- Fixed vLLM-Omni’s Triton SnakeBeta activation kernel for bf16 and fp16 inputs (#3472)
Quantization
All in vLLM-Omni:
- FP8 for the Qwen2.5-Omni thinker language model (#3466, in review)
- Loading Comfy INT8 ConvRot checkpoints for MiniMax-H3 (#6894, in review)
- Initialize component-level quantization base state (#5103)
Learning
Most of this layer is still study: CUDA GEMM from a naive kernel up to tensor cores, Triton, and FlashAttention. Notes land on the blog under this layer.