Vijay Janapa Reddi
|
3750ee12e9
|
Enforce Computed Arithmetic Rule across all chapters (1,064 inline refs, 0 unresolved)
Replace every hand-typed derived number with Python-computed inline
references. Add just-in-time compute cells before prose so that changing
any input constant automatically propagates to all derived values.
Vol 1 chapters fixed: dl_primer, dnn_architectures, serving,
model_compression, hw_acceleration, benchmarking, ops, appendix_machine,
appendix_data, frameworks, data_engineering, training, ml_systems,
responsible_engr, data_selection, workflow, introduction.
Vol 2 chapters fixed: distributed_training, inference, infrastructure,
storage, sustainable_ai, fault_tolerance, ops_scale, edge_intelligence,
ai_for_good, privacy_security.
Key corrections caught by forcing computation:
- training.qmd carbon footprint: 64 GPUs → 1024 GPUs (original was
mathematically impossible for 7B params × 1T tokens)
- hw_acceleration.qmd systolic energy: 10 pJ/250× → 11 pJ/233× (exact)
- hw_acceleration.qmd GPT-2 utilization: 0.6% → 0.7% (exact)
- serving.qmd tokens/hour: ~190M → ~192M (exact)
Also adds calc/validate_inline_refs.py pre-render guardrail and
extends calc/viz.py with Harvard Crimson plotting palette.
|
2026-02-01 11:13:42 -05:00 |
|
Vijay Janapa Reddi
|
7ad6d51f96
|
Update two-volume textbook content, config, and tooling
- Edit all Vol 1 and Vol 2 chapters for print readiness and pedagogical clarity
- Update Quarto config files for both volumes (PDF, HTML, EPUB)
- Add frontmatter updates (about, acknowledgements, socratiq)
- Remove unused _brand assets (scss, favicon, scripts, manifest)
- Add new utility scripts (audit_figure_placement, format_div_spacing, audit_refs)
- Update format_python_in_qmd script
- Add references.bib entries and seminal papers corpus
|
2026-01-30 02:42:59 -05:00 |
|