PR Walkthroughs

Staging Large FMA Dots via SLM

Fix the PTSS overflow cliff on non-DPAS Intel iGPUs by staging tt.dot operands in SLM and tiling K.

6 Modules 1-2 Hours

TD Q-Load for Non-Power-of-2 GQA

When a GPU block-I/O fast path demands power-of-2 extents, pad the awkward axis, read wider, and mask the slack — taught through attention's Q load and prefill tile sizing.

7 Modules 1-2 Hours

GPU Optimization: 10 vLLM PRs

Learn optimization techniques from actual merged vLLM PRs with verified performance numbers.

10 Modules 2-3 Hours

PyTorch Optimization: 10 PRs

Learn optimization techniques from actual merged PyTorch PRs with verified performance numbers.

10 Modules 2-3 Hours

Want a specific PR broken down?

Suggest a merged PR in GPU kernels, ML compilers, or systems — I turn the good ones into walkthroughs

Get in Touch