This commit standardizes percentage formatting across the entire codebase to prevent critical rendering bugs (like the `19250000000000%` effective utilization bug in Vol 2).
Root Cause:
When dividing two Pint Quantities (e.g., `flop/second` by `TFLOPs/second`), Pint creates a mixed unit (`flop/TFLOPs`). The raw `.magnitude` of this fraction is $10^{12}$. When passed to `fmt(x * 100)`, it multiplied that massive magnitude by 100, resulting in an incorrect display.
Fix:
1. Fortified `fmt_percent` and `display_percent` in `mlsys/formatting.py` to defensively strip units using `.m_as('')`. This forces Pint to cancel out the units (e.g., `flop/TFLOPs` becomes `1.0`) *before* extracting the number.
2. Replaced all instances of `fmt(X * 100)` with the fortified `fmt_percent(X)` across Vol 1 and Vol 2.
3. Fixed inline f-strings in `appendix_assumptions.qmd` by moving formatting logic into the Python setup cell as `_str` variables, adhering to the book's standard practice.
Validation:
- Audited all `.magnitude` extractions in the codebase to ensure they are safe (e.g., explicitly converting to dimensionless units first).
- Ran `validate_inline_refs.py` and confirmed no Python variables are trapped inside LaTeX math mode.
- Successfully built full PDFs for both Volume 1 and Volume 2.
- Add missing attributes to FleetFoundations in appendix_fleet.qmd
- Fix regression_testing.png image path in fault_tolerance.qmd
- Add pgfplots package to header-includes.tex for TikZ compatibility
- Fortify fmt_percent in formatting.py to handle Pint Quantities properly, fixing the 19250000000000% display bug
Introduces `--only-from-report` and `--only-keys` arguments to the `references` validation command.
These allow re-validating only specific citation keys, either from a previous validation report or a custom list.
This significantly improves the workflow for correcting references by enabling focused re-runs and reducing validation time.
Removes the standalone `README_REFERENCE_CHECK.md` documentation, as its content is now implicitly handled by the integrated CLI help and broader documentation.
Flagged by the sustainable_ai editor agent as newly discovered during fixing:
- line 635: "If your cluster consumes...how much...actually went...how much was wasted?"
→ impersonal declarative; removes "your", two embedded rhetorical questions, two "actually"
- line 2261: "You want to fine-tune a small language model" in .callout-notebook
→ "Consider fine-tuning a small language model" (impersonal)
Post-commit verification found 6 additional violations not caught by the
initial audit agents:
distributed_training (4 fixes):
- line 108: second person "If you could purchase a single GPU" → impersonal
- line 280: rhetorical Q "How exactly do 1,024 GPUs...agree" → declarative
- line 784: second person "Your AllReduce...Where do you look?" in
.callout-perspective → impersonal problem statement
- line 1347: rhetorical Q "where did the missing 25%...go?" → declarative
sustainable_ai (2 fixes):
- line 2047: embedded rhetorical Q "where does the dominant share of energy go?" → declarative
- line 2414: closing rhetorical Q "what happens to these clusters...?" → declarative noun phrase
