Skip to content

My kernel learning journal

Copy this file to results/local/my-learning-journal.md and fill it yourself. results/local/ is ignored by Git, so your notes stay private until you choose to move them elsewhere. Reference-solution validation in this repo does not mark any of these as completed.

Session record (repeat this section)

  • Date and lesson:
  • File I changed:
  • What I expected before running:
  • Command I ran:
  • First failing case and what it taught me:
  • What I changed and why:
  • Final checker result:
  • One explanation in my own words:
  • One question I still cannot answer:

Experiment record

  • Operation, formula and supported input assumptions:
  • GPU, software versions, shape and dtype:
  • Baseline and candidate:
  • Hypothesis (written before the run):
  • Timing mode and what it excludes:
  • Correctness comparison and tolerance:
  • Raw JSON path:
  • Baseline median / candidate median / computed ratio:
  • Did the samples vary enough to question the result?
  • What I observed (facts):
  • Why I think it happened (hypothesis, unless independently verified):
  • One losing case or limitation:
  • Next experiment that could disprove my explanation:

Readiness demonstration

  • [ ] I can derive masks and addresses without looking at a solution.
  • [ ] I can calculate stable softmax for a three-element row.
  • [ ] I can explain storage dtype versus accumulation dtype.
  • [ ] I can write add and row sum from a blank file.
  • [ ] I can explain the softmax load padding identity.
  • [ ] I can draw GEMM output ownership and its K loop.
  • [ ] I can explain why the saved graph and dispatch timings differ.
  • [ ] I have one experiment with my own prediction, result and explanation.
  • [ ] I can identify which parts of this portfolio I implemented or studied myself.