docs(reports): add technical reports for the #1053/#1054/#1045/#1040 chain - #1077
Merged
Conversation
…chain Bilingual reports for the four PRs merged in that chain: #1058 and #1059 (the MLX CUDA JIT cache key omitting input dtypes, and the sweep plus CI guard for the rest of that class), #1061 (the fused decode-MoE parity determination), and #1062 (promoting chunked GLA prefill to the bailing_moe_linear default). `TECHNICAL_REPORTS/` is gitignored and no report has been tracked before, so these are force-added by explicit path. The directory carries no `.keep-reports` marker, so this does not change the default for future reports; it tracks these eight files only. Each report records what the measurement actually settled rather than what the issue expected, since three of the four landed against their issue's premise: #1045's kernel turned out more accurate than the reference it disagreed with, #1040's opt-in path turned out better rather than merely different, and #1053/#1054 turned out to share one root cause in a dependency's backend asymmetry. The two CUDA reports also state plainly which acceptance criterion could not be met on an Apple Silicon host and what a GB10 session should run.
inureyes
added a commit
that referenced
this pull request
Aug 7, 2026
## Summary Adds `TECHNICAL_REPORTS/.keep-reports`, the marker the report workflow already defines as the opt-in for tracking reports in git. ## Why now `TECHNICAL_REPORTS/` is in `.gitignore`, so reports stayed local and each had to be force-added by hand. PR #1077 did that for eight reports and left the directory in a half-and-half state: eight files tracked out of seventy-seven, inside a directory the ignore rules still say to skip. Either the reports are tracked here or they are not, and the marker is how the workflow expects that to be declared. ## What changes With the marker present, the report workflow: - may generate a report for a still-open PR rather than only a merged one, - dates it today instead of by `mergedAt`, - force-adds, commits and pushes it, so the report is inside the squash merge rather than trailing behind it. The marker force-adds itself, so the opt-in survives a fresh clone. ## What does not change The sixty-nine reports predating #1077 stay untracked. Adding them is a separate decision and several belong to work whose reports were deliberately left local, so this PR does not sweep them in. ## Validation One new file, no code, no build or test surface touched. The file is force-added by explicit path.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Bilingual technical reports for the four PRs merged in the
#1053, #1054, #1045, #1040chain:1058-paged-v2-dtype-kernel-key1059-cuda-jit-dtype-key-sweep1061-fused-moe-parity-checkpoint-dependent1062-chunked-gla-prefill-defaultbailing_moe_lineardefaultWhy these read the way they do
Three of the four landed against their issue's premise, so each report leads with what the measurement settled rather than with what the issue expected:
gather_qmmreference it disagreed with, in 96 of 96 calls on each of two families. The disagreement is mostly the reference's own distance from truth.Tracking note
TECHNICAL_REPORTS/is gitignored and no report had been tracked before, so these eight files are force-added by explicit path. The directory carries no.keep-reportsmarker, so this changes nothing about future reports; it tracks these eight only. Adding the marker would be a separate decision, since it also makes the report workflow commit and push on its own before a merge.Verification gaps carried into the reports
The two CUDA reports state plainly which acceptance criterion could not be met on an Apple Silicon host, and what a GB10 session should run:
cargo test --release --features cuda -p mlxcel-core --lib -- --test-threads=1has not been run against either fix. There is nonvccon the implementing host and no reachable CUDA node (re-verified 2026-08-07).1062also records thatinclusionAI/Ring-flash-linear-2.0, the larger sibling the issue lists, was not measured.Validation
scripts/ci/check_cross_repo_refs.pyis advisory and lists the bare#NNNreferences it found; all thirteen distinct numbers cited (268, 886, 1039, 1040, 1044, 1045, 1048, 1053, 1054, 1058, 1059, 1061, 1062) were confirmed viaghto be genuinelablup/mlxcelissues or PRs, so no internal number is leaked. No code, build, or test surface is touched.