How NCCL's wire protocols trade payload bytes against synchronization cost

Exploring Hopper Architecture on H100 and try to match PyTorch/cuBLAS performance for GEMMs using CUTLASS

Walkthrough of optimization techniques for GEMMs from a naive fp32 kernel to CUTLASS bf16 kernel

Worklog: Performance debugging Triton Kernel

Fused Softmax Triton kernel exploration

RMS Normalization Triton kernel implementation for LLMs

Visualizer for MXFP4 quantization

KV Caching: Training vs Inference in Multi-Head Attention

Load OpenAI's reference implementation

Testing out the huggingface version for the gpt-oss-20b locally on consumer hardware