Repository navigation
Revert grouped GPU decoding (api#1234) - #1241
Conversation
In production, grouped decoding doubled full_report instead of cutting it: on the same song, wall time went from 181 s to 367 s and billed seconds from 179 to 365. generate_batch goes back to answering prompts one at a time, as verified after api#1233. This reverts merge commit 026071c. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
|
Important Review skippedReview was skipped due to path filters ⛔ Files ignored due to path filters: 3
CodeRabbit blocks several paths by default. You can override this behavior by explicitly including those paths in the path filters. For example, including
|
Codex Review Summary: ✅ CompletedThis 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. |
Why
#1234 decoded a batch's prompts together on the GPU to make
full_reportfaster. It was merged without a GPU run, with a stated fallback: verify on production and revert if it didn't help. In production it doubled the time and the billed seconds.Same song: a public-domain 1915 Edison recording, 3:59. Each run followed a fresh Modal deploy.
full_reportMoment descriptions barely moved (80 s → 75 s). The greedy presets' answers drifted by a few tokens, as #1234 expected, so the problem is speed and cost, not answer quality.
Callers can't fall back to one-at-a-time decoding:
group_sizeis a Modal field the API doesn't send. So the revert is the only fallback.What changes
This reverts the merge commit 026071c.
generate_batchgoes back to answering prompts one at a time: the code verified in production after #1233. Onlymodal/changes, and its three files matchmainbefore #1234.Follow-up
Find out where the time goes before trying grouped decoding again. Candidates:
lyrics).#1234's
benchentrypoint can measure this, but needs access to the shared Modal workspace.🤖 Generated with Claude Code