Add programmable CUDA boosting and verified objective comparisons - #25
Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7b18667daa
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| - uses: actions/checkout@v4 | ||
| with: | ||
| # Historical execution SHAs are part of the committed evidence contract. | ||
| fetch-depth: 0 |
There was a problem hiding this comment.
A fresh checkout of this commit cannot recover the execution commits merely by setting fetch-depth: 0: 9da718c has only parent 47108db, and its reachable history excludes the manifest revisions fe12beb and dd84247. The non-GPU CI command nevertheless runs test_glm_freeze.py and test_validation_freeze.py, which call git show on those revisions, so these evidence checks fail once the separate development branch is unavailable. Merge the execution history as parents, retain reachable refs, or make the verifiers use committed source archives instead.
AGENTS.md reference: AGENTS.md:L151-L155
Useful? React with 👍 / 👎.
1 participant
This change adds a programmable Python/CUDA boosting foundation: owned device
storage, named fields, histograms, split/feasibility/routing/leaf operations,
trees and resident transactions now support squared, Normal, binary and Poisson
recipes. Objective-owned loss-change bounds keep training acceptance, best-model
selection and patience separate even when reported losses round to the same value.
CPU execution gains verified incremental proposal evaluation, reusable encodings,
structural stopping results and bounded trace retention. The branch also preserves
the planning, failures and raw evidence accumulated since PR #24; most changed
files are evidence rather than production code.
Validation and evidence
ninety explicitly historical full-loss deselections. All ninety settings have
active objective-comparison replacements, verified by collection and execution.
Full production/test Ruff, strict MkDocs and wheel/sdist builds pass.
571/571 real T4 cases pass, including 153 new GLM cases and 418 regressions.
All 77 artifacts are retained; offline audit verifies 246 loss-change bounds and
replays all 32 final/best models from lossless input bytes.
bounded revised coverage is complete across two runs, with the original failed
verdict and historical disagreements preserved.
all 474 T4 checks and three frozen internal cost gates pass. Parallel field
validation lowers synthetic 100,000-row squared warm fit time from 13.513 to
8.947 seconds (33.79%), preserving model/prediction bytes. This is an internal
synthetic comparison, not external-library parity or formal E4.
hosted CPU runners and builds documentation strictly. The manual GPU workflow
remains non-executing; no new paid GPU/model invocation is part of this PR.
exposed nine platform-sensitive legacy reporting/decision assertions. Current
tests now exercise objective comparisons under native and controlled reporting;
original full-loss sources and hardware evidence remain unchanged. The matrix
retains all host/version outcomes with fail-fast disabled.
passes both Linux cells and identifies four one-ULP reporting differences on
macOS. Replay now pins original artifact bytes and requires exact decisions,
settings and source identities, with at most one ULP for recomputed reporting
scores. Tests reject larger errors and changes to coefficients or model prefixes.
Scope and remaining work
This remains an experimental foundation, not a drop-in replacement or completed
v1 release. Required multiclass, AFT and vector-topology CUDA cells, compatible
device train-many, real application quality and formal E4 remain open. Authoring
and adoption studies are deferred under Sprint 101; retained study infrastructure
does not constitute author/adoption evidence. All R/C/A requirements remain.
Merge strategy
Use Create a merge commit. Evidence tests resolve the original execution Git
SHAs; squash or rebase merging would break their availability in future clean
checkouts. Review the hosted Linux/macOS, Python 3.10/3.12 and documentation checks
before merging. This PR does not publish a package or dispatch hardware.